Probably worth mentioning that David Baker’s lab released a similar model (predicts protein structure along with bound DNA and ligands), just a couple of months ago, and it is open source [1].
It’s also worth remembering that it was David Baker who originally came up with the idea of extending AlphaFold from predicting just proteins to predicting ligands as well [2].
Unlike AlphaFold 3, which predicts only a small, preselected subset of ligands, RosettaFold All Atom predicts a much wider range of small molecules. While I am certain that neither network is up to the task of designing an enzyme, these are exciting steps.
One of the more exciting aspects of the RosettaFold paper is that they train the model for predicting structures, but then also use the structure predicting model as the denoising model in a diffusion process, enabling them to actually design new functional proteins. Presumably, DeepMind is working on this problem as well.
I appreciated this, but it's probably worth mentioning: when you say AlphaFold 3, you're talking about AlphaFold 2.
TFA announces AlphaFold 3.
Post: "Unlike AlphaFold 3, which predicts only a small, preselected subset of ligands, RosettaFold All Atom predicts a much wider range of small molecules"
TFA: "AlphaFold 3...*models large biomolecules such as proteins, DNA and RNA*, as well as small molecules, also known as ligands"
Post: "they also use the structure predicting model as the denoising model in a diffusion process...Presumably, DeepMind is working on this problem as well."
TFA: "AlphaFold 3 assembles its predictions using a diffusion network, akin to those found in AI image generators."
Coming up with ideas is cheaper than executing the ideas. Predicting a wide range of molecules okay-ish is cheaper than predicting a small range of molecules very well.
Important caveat: it's only about 70% accurate. Why doesn't the press release say this explicitly? It seems intentionally misleading to only report accuracy relative to existing methods, which apparently are just not so good (30%, 50% in various settings).
https://www.fastcompany.com/91120456/deepmind-alphafold-3-dn...
They also had a headline for Alphazero that convinced everyone that they crushed Stockfish and that classical chess engines were stuff of the past, when in fact it was about 50 elo better than the Stockfish version they were testing against, or roughly the same as how much Stockfish improves each year.
I think Alphazero is a lot more interesting than Stockfish though. Most notably it lead me to reevaluate positional play. Iirc A0 at around 2-3 ply is still above SuperGM Level which is pretty mind-blowing. Based on this I have increased my strategy to tactics ratio quite a bit. FWIW Stockfish is always evolving and adapting and has incorporated ideas from A0.
And then what happened is AlphaZero changed the professional game in various interesting ways, and all its ideas were absorbed into Stockfish. A little bombast is forgivable for technology that goes on to have a big impact, and I don’t doubt it’s the same story here.
Just "Alpha-- models" in general?? That's not a remotely reasonable way to reason about it. Even if it were, why should it stop DeepMind from clearly communicating accuracy?
Which specific AlphaX model evolved like that? Most of the ones that were in the press had essentially a single showing, typically very good, but didn't really improve after that.
Very sad to see they did not make it open source. When you have a technology that has the potential to be a gateway for drug development, to the cures of new diseases, and instead you choose to make it closed, it is a very huge disservice to the community at large. Sure, release your own product alongside it, but making it closed source does not help the scientific community upon which all these innovations were built. Especially if you have lost a loved one to a disease which this technology will one day be able to create cures for, it is very disappointing.
Sometimes the lines get blurred - but I don't think you can expect Google to release as much of their work for free as people who are paid via central taxes.
Worth noting that they did release the AlphaFold 2 weights after a while. Milking an expensive discovery for a limited period should be considered laudable, unless you think tax funding of all research would be awesome and it's just a weird anomaly how the org producing these results was a tiny heterodox startup very recently.
The closer it gets to enabling full drug discovery, the closer it also gets to enabling bioterrorism. Taking it to the extreme, if they had the theory of everything, I don't think I'd want it to be made available to the whole world as it is today.
On a related note, I highly recommend The Talos Principle 2, which really made me think about these questions.
Oh please, like a terrorist cant fork over a couple bucks to do the bioterrorism. This excuse is utter bs, whether its applied to LLMs or to alphafold. The motivator is profit, not safety.
> What is different about the new AlphaFold3 model compared to AlphaFold2?
> AlphaFold3 can predict many biomolecules in addition to proteins. AlphaFold2 predicts structures of proteins and protein-protein complexes. AlphaFold3 can generate predictions containing proteins, DNA, RNA, ions,ligands, and chemical modifications. The new model also improves the protein complex modelling accuracy. Please refer to our paper for more information on performance improvements.
AlphaFold 2 generally produces looping “ribbon-like” predictions for disordered regions. AlphaFold3 also does this, but will occasionally output segments with secondary structure within disordered regions instead, mostly spurious alpha helices with very low confidence (pLDDT) and inconsistent position across predictions.
So the criticism towards AlphaFold 2 will likely still apply? For example, it’s more accurate for predicting structures similar to existing ones, and fails at novel patterns?
I am not aware of anybody currently criticiszing AF2's abilities outside of its training set. In fact the most recent papers (written by crystallographers) they are mostly arguing about atomic-level details of side chains at this point.
>So the criticism towards AlphaFold 2 will likely still apply? For example, it’s more accurate for predicting structures similar to existing ones, and fails at novel patterns?
Yes, and there is simply no way to bridge that gap with this technique. We can make it better and better at pattern matching, but it is not going to predict novel folds.
As a software engineer, I kind of feel uncomfortable about this new model. It outperforms Alphafold 2 at ligand binding, but Alphafold 2 also had some more hardcoded and interpretable structural reasoning baked into the model architecture.
There's so many things you can incorporate into a protein folding model such as structural constraints, rotational equivariance, etc, etc
This new model simple does away with some of that, achieving greater results. And the authors simply use distillation from data outputted from Alphafold2 and Alphafold2-multimer to get those better results for those cases where you wind up with implausible results.
You have to run all those previous models, and output their predictions to do the distillation to achieve a real end-to-end training from scratch for this new model! Makes me feel a bit uncomfortable.
Why? Do compilers which can't bootstrap themselves also make you uncomfortable due to dependencies on pre-built artifacts? I'm not saying you're unjustified to feel that way, but sometimes more abstracted systems are quicker to build and may have better performance than those built from the ground up. Selecting which one is better depends on your constraints and taste
Compilers are deterministic (for the most part, and it's incredibly rare to introduce a compiler bug that self-replicates in future compilers (unless you're Ken Thompson and are reflecting upon trust itself)).
Alternatively, AlphaFold 2's output is noisy, and using that to train AlphaFold 3, which presumably may be used to train what becomes AlphaFold 4, results in a cascade of errors.
Ahh, but the new young master is able to explain their work and processes to the satisfaction of the old masters. In the 'Science' of our modern times it's a requirement to show your work (yes, yes, I know about the replication crisis and all that terrible jazz).
Not being able to ascertain how and why the ML/AI is achieving results is not quite the same and more akin to the alchemists and sorcerers with their cyphers and hidden laboratories.
I am trying to understand how accurate the docking predictions are.
Looking at the PoseBusters paper [1] they mention, they say they are 50% more accurate than traditional methods.
DiffDock, which is the best DL based systems gets 30-70% depending on the dataset, and traditional gets 50-70%. The paper highlighted some issues with the DL-based methods and given that DeepMind would have had time to incorporate this into their work and develop with the PoseBusters paper in mind, I'd hope it's significantly better than 50-70%. They say 50% better than traditional so I expected something like 70-85% across all datasets.
I hope a paper will appear soon to illuminate these and other details.
So much of the talk about their "free server" seems to be trying to distract from the fact that they are not releasing the model.
I feel like it's an important threshold moment if this gets accepted into scientific use without the model being available - reproducibility of results becomes dependent on the good graces of a single commercial entity. I kind of hope that like OpenAI it just spurs creation of equivalent open models that then actually get used.
This tool reminds me that the human body functions much like a black box. While physics can be modeled with equations and constraints, biology is inherently probabilistic and unpredictable. We verify the efficacy of a medicine by observing its outcomes: the medicine is the input, and the changes in symptoms are the output. However, we cannot model what happens in between, as we cannot definitively prove that the medicine affects only its intended targets. In many ways, much of what we understand about medicine is based on observing these black-box processes, and this tool helps to model that complexity.
>However, if the radio has tunable components, such as those
found in my old radio (indicated by yellow arrows in Figure 2,
inset) and in all live cells and organisms, the outcome will not be
so promising. Indeed, the radio may not work because several
components are not tuned properly, which is not reflected in their
appearance or their connections. What is the probability that this
radio will be fixed by our biologists? I might be overly pessimistic,
but a textbook example of the monkey that can, in principle, type
a Burns poem comes to mind. In other words, the radio will not
play music unless that lucky chance meets a prepared mind.
I’d say it’s always been the case for medicine, when people first used medicines, the intention was never to fully understand what happens, just save a life, eliminate or reduce symptoms.
Now we’ve built explainable systems like computers and software, we try to overlay that onto everything and it might not work.
To quote Alan Watts, humans like to try square out wiggly systems because we’re not great and understanding wiggles.
> Thrilled to announce AlphaFold 3 which can predict the structures and interactions of nearly all of life’s molecules with state-of-the-art accuracy including proteins, DNA and RNA. [1]
There's a slight mismatch between the blog's title and Demis Hassabis' tweet, where he uses "nearly all".
The blog's title suggests that it's a 100% solved problem.
This is a basic question, but how is the accuracy of the predicted biomolecular interactions measured? Are the predicted interactions compared to known interactions? How would the accuracy of predicting unknown interactions be assessed?
Accuracy can be assessed two main ways: computationally and experimentally. Computationally, they would compare the predicted structures and interactions with known data from databases like PDB (Protein Database). Experimentally, they can use tools like x-ray crystallography and NMR (nuclear magnetic resonance) to obtain the actual molecule structure and compare it to the predicted result. The outcomes of each approach would be fed back into the model for refining future predictions.
AlphaFold very explicitly (unless something has changed) removes NMR structures as references because they are not accurate enough. I have a PhD in NMR biomolecular structure and I wouldn't trust. the structures for anything.
Can someone tell me what are the direct implication of this? I often see "helps with a drug design" but I'm too far from this industry and have never seen an example of such drugs
In transformer based architectures, where one typically uses variation of attention mechanism to model interactions, even if one does not consider the autoregressive assumption of the domain's "nodes"(amino acids, words, image patches), if the number of final states that nodes take eventually can be permuted only in a finite way(i.e. they have sparse interactions between them), then these architectures are efficient way of modeling such domains.
In plain english the final state of words in a sentence and amino acids in a protein have only so many ways they can be arranged and transformers do a good job of modeling it.
Also can one assume this won't do well for domains where there is, say, sensitivity to initial conditions, like chaotic systems like wheather where the # final states just explodes?
For a couple of years I've been expecting that ML models would be able to 'accelerate' bio-molecular simulations, using physics-based simulations as ground truth. But this seems to be a step beyond that.
When I competed in CASP 20 years ago (and lost terribly) I predicted that the next step to improve predictions would be to develop empirically fitted force fields to make MD produce accurate structure predictions (MD already uses empirically fitted force fields, but they are not great). This area was explored, there are now better force fields, but that didn't really push protein structure prediction forward.
Another approach is fully differentiable force fields- the idea that the force field function itself is a trainable structure (rather than just the parameters/weights/constants) that can be optimized directly towards a goal. Also explored, produced some interesting results, but nothing that woudl be considered transformative.
The field still generally believes that if you had a perfect force field and infinite computing time, you could directly recapitulate the trajectories of proteins folding (from fully unfolded to final state along with all the intermediates), but that doesn't address any practical problems, and is massively wasteful of resources compared to using ML models that exploit evolutionary information encoded in sequence and structures.
In retrospect I'm pretty relieved I was wrong, as the new methods are more effective with far fewer resources.
This version has Isomorphic Labs far more in the focus of the press release, which seems to be now the commercial arm more or less licensing access out.
The new AlphaFold server does not do everything the paper says AlphaFold 3 says it does. You cannot predict docking with the server! That is the main interest of pharma companies, 'does our medication bind to the target protein?'. From the FAQ: 'AlphaFold Server is a web-service that offers customized biomolecular structure prediction. It makes several newer AlphaFold3 capabilities available, including support for a wider range of molecule type' - that's not ALL AlphaFold3 capabilities. Isomorphic prints the money with those additional capabilities.
It's hilarious that Google says they don't allow this for safety reasons, pure OpenAI fluff. It's just money.
Here's something that bugs me about ML: all we have is prediction and no explanation how we come to that prediction, ie: no deeper understanding on the underlying principles.
So despite that we got a good match this time, how can we be sure that the match will be equally good next time? And how to use ML to predict structure that we have no baseline to start with or experimental result to benchmark ? In the absence of physics-like principles, How can we ever be sure that ML results next time is correct ?
There is a biannual structural prediction contest called CASP [1], in which a set of newly determined structures is used to benchmark the prediction methods. Some of these structures will be "novel", and so can be used to estimate the performance of current methods on predicting "structure that we have no baseline to start with".
CASP-style assessments are something that should done for more research fields, but it's really hard to persuade funders and researchers to put up the money and embargo the data as required.
Speaking of physics, we should borrow the quote "Shut up and calculate" to describe the situation: it works so use it now and worry about the explanations later.
I'm interested in how they measure accuracy of binding site identification and binding pose prediction. This was missing for the hitherto widely-used binding pose prediction tool Autodock Vina (and in silico binding pose tools in general). Despite the time I invested in learning & exercising that tool, I avoided using it for published research because I could not credibly cite its general-use accuracy. Is / will Alphafold 3 be citeable in the sense of "I have run Alphafold on this particular target of interest and this array of ligands, and have found these poses of X kJ/mol binding energy, and this is known to an accuracy of Y% because of Alphafold 3's training set results cited below'
I've never trusted those predicted binding energies. If you have predicted a ligand/protein complex and have high confidence in it and want to study the binding energy I really think you should do a full MD simulation, you can pull the ligand-protein complex apart and measure the change in free energy explicitly.
Also, and this is an unfounded guess only, the problem of protein / ligand docking is quite a bit more complex than protein folding - there seems to be a finite set of overall folds used in nature, while docking a small ligand to a big protein with flexible sidechains and even flexible large-scale structures can have induced fits that are really important to know and estimate, and I'm just very sceptical that it's going to be possible to in a general fashion ever predict these accurately by the AI model with the limited training data.
Though you just need some hints, then you can run MD sims on them to see what happens for real.
Would anyone more familiar with the field be able to provide some cursory resources on the protein folding problem? I have a background in computer science and a half a background in biology (took two semesters of OChem, biology, anatomy; didn’t go much further).
Thanks for this informative video summary. As a layperson, with a BS in Chemistry, it was quite helpful in understanding main bulletpoints of this accomplishment.
I wonder in the not too distant future if these AI predictions could be explained back into “humanized” understanding. Much like ChatGPT can simplify complex topics … cold the model in the future provide feedback to researchers why it is making this prediction?
The improvement on predicting protein/RNA/ligand interactions might facilitate many commercially relevant use cases. I assume pharma and biotech will eagerly get in line to use this.
>Unlike RoseTTAFold and AlphaFold2, scientists will not be able to run their own version of AlphaFold3, nor will the code underlying AlphaFold3 or other information obtained after training the model be made public. Instead, researchers will have access to an ‘AlphaFold3 server’, on which they can input their protein sequence of choice, alongside a selection of accessory molecules. [. . .] Scientists are currently restricted to 10 predictions per day, and it is not possible to obtain structures of proteins bound to possible drugs.
This is unfortunate. I wonder how long until David Baker's lab upgrades RoseTTAFold to catch up.
That sucks a bit. I was just wondering why they are touting that 3rd party company in their own blog post, who commercialise research tools, as well. Maybe there are some corporate agreements with them that prevents them from opening the system...
Imagine the goodwill for humanity for releasing these pure research systems for free. I just have a hard time understanding how you can motivate to keep it closed. Let's hope it will be replicated by someone who doesn't have to hide behind the "responsible AI" curtain as it seems they are now.
Are they really thinking that someone who needs to predict 11 structures per day are more likely to be a nefarious evil protein guy than someone who predicts 10 structures a day? Was AlphaFold-2 (that was open-sourced) used by evil researchers?
- "Imagine the goodwill for humanity for releasing these pure research systems for free."
The entire point[0] is that they want to sell an API to drug-developer labs, at exclusive-monopoly pricing. Those labs in turn discover life-saving drugs, and recoup their costs from e.g. parents of otherwise-terminally-ill children—again, priced as an exclusive monopoly.
[0] As signaled by "it is not possible to obtain structures of proteins bound to possible drugs"
It's a massive windfall for Alphabet, and it'd be a profound breach of their fiduciary duties as a public company to do anything other than lock-down and hoard this API, and squeeze it for every last billion.
This is a deeply, deeply, deeply broken situation.
Isomorphic Labs? That's an Alphabet owned startup run by Denis Hassabis that they created to commercialise the Alphafold work, so it's not really a 3rd party at all.
> AlphaFold 3 will be available as a non-commercial usage only server at https://www.alphafoldserver.com, with restrictions on allowed ligands and covalent modifications. Pseudocode describing the algorithms is available in the Supplementary Information. Code is not provided.
How easy/hard would be for the scientific community to come up with an "OpenFold" model which is pretty much AF3 but fully open source and without restrictions in it?
I can image training will be expensive, but I don't think it will be at a GPT-4 level of expensive.
If you need to submit to their server, I don't know who would use it for commercial reasons anyway. Most biotech startups and pharma companies are very careful about entering sequences into online tools like this.
The AI call is rolling fast, I see similarities with cryptography in the 90s.
I have a history to tell for the record, back in the 90s we developed a home banking for Palm (with a modem), it was impossible to perform RSA because of the speed so I contacted the CEO of Certicom which was the unique elliptic curve cryptography implementation at that time. Fast forward and ECC is everywhere.
Not just unfortunate, but doesn't this make it completely untrustable? How can you be sure the data was not modified in any way? How can you verify any results?
You determine a crystal structure of a known protein which does not previously have a known structure, and compare the prediction to the experimentally determined structure.
There is a biennial (biannual?) competition known as CASP where some new structures, not yet published, are used for testing predictions from a wide range of protein structure prediction (so, basically blind predictions which are then compared when the competition wraps up). AlphaFold beat all the competitors by a very wide margin (much larger than the regular rate of improvement in the competition), and within a couple years, the leading academic groups adopted the same techniques and caught up.
It was one of the most important and satisfying moments in structure prediction in the past two+ decades. The community was a bit skeptical but as it's been repeatedly tested, validated, and reproduced, people are generally of the opinion that DeepMind "solved" protein structure prediction (with some notable exceptions), and did so without having the solve the full "protein folding problem" (which is actually great news while also being somewhat depressing).
No. It just means that scientific purposes will have an additional tax paid to google. This will likely reduce use in academia but won't deter pharmaceutical companies.
The second amendment prevents the government's overreaching perversion to restrict me from having the ability to print biological weapons from the comfort of my couch.
I know this is tongue in cheek, but you absolutely can be restricted from having a biological weapons factory in your basement (similar to not being able to pick "nuclear bombs" as your arms to bear).
The logical consequence is to put all scientific publications under a license that restricts the right to train commercial ai models on them.
Science advances because of an open exchange of ideas, the original idea of patents was to grant the inventor exclusive use in exchange for disclosure of knowledge.
Those who did not patent, had to accept that their inventions would be studied and reverse engineered.
Well, it's because you can design deadly viruses using this technology. Viruses gain entry to living cells via cell-surface receptor proteins whose normal job is to bind signalling molecules, alter their conformation and translate that external signal into the cellular interior where it triggers various responses from genomic transcription to release of other signal molecules. Viruses hijack such mechanisms to gain entry to cells.
Thus if you can design a viral coat protein to bind to a human cell-surface receptor, such that it gets translocated into the cell, then it doesn't matter so much where that virus came from. The cell's firewall against viruses is the cell membrane, and once inside, the biomolecular replication machinery is very similar from species to species, particularly within restricted domains, such as all mammals.
Thus viruses from rats, mice, bats... aren't going to have major problems replicating in their new host - a host they only gained access to because some nation-state actors working in collaboration on such gain-of-function research in at least two labs on opposite sides of the world with funds and material provided by the two largest economic powers for reasons that are still rather opaque, though suspiciously banal...
Now while you don't need something like AlphaFold3 to do recklessly stupid things (you could use directed evolution, making millions of mutatad proteins, throwing them at a wall of human cell receptors and collecting what stuck), it makes it far easier. Thus Google doesn't want to be seen as enabling, though given their prediliction for classified military-industrial contracting to a variety of nation-states, particularly with AI, with revenue now far more important than silly "don't be evil" statements, they might bear watching.
On the positive side, AlphaFold3 will be great for fields like small molecular biocatalysis, i.e. industrial applications in which protein enzymes (or more robust heterogenous catalysts designed based on protein structures) convert N2 to ammonia, methane to methanol, or selectively bind CO2 for carbon capture, modification of simple sugars and amino acids, etc.
Stepping back, the high-order bit here is an ML method is beating physically-based methods for accurately predicting the world.
What happens when the best methods for computational fluid dynamics, molecular dynamics, nuclear physics are all uninterpretable ML models? Does this decouple progress from our current understanding of the scientific process - moving to better and better models of the world without human-interpretable theories and mathematical models / explanations? Is that even iteratively sustainable in the way that scientific progress has proven to be?
If you're a scientist who works in protein folding (or one of those other areas) and strongly believe that science's goal is to produce falsifiable hypotheses, these new approaches will be extremely depressing, especially if you aren't proficient enough with ML to reproduce this work in your own hands.
If you're a scientist who accepts that probabilist models beat interpretable ones (articulated well here: https://norvig.com/chomsky.html), then you'll be quite happy because this is yet another validation of the value of statistical approaches in moving our ability to predict the universe forward.
If you're the sort of person who believes that human brains are capable of understanding the "why" of how things work in all its true detail, you'll find this an interesting challenge- can we actually interpret these models, or are human brains too feeble to understand complex systems without sophisticated models?
If you're the sort of person who likes simple models with as few parameters as possible, you're probably excited because developing more comprehensible or interpretable models that have equivalent predictive ability is a very attractive research subject.
(FWIW, I'm in the camp of "we should simultaneously seek simpler, more interpretable models, while also seeking to improve native human intelligence using computational augmentation")
The goal of science has always been to discover underlying principles and not merely to predict the outcome of experiments. I don't see any way to classify an opaque ML model as a scientific artifact since by definition it can't reveal the underlying principles. Maybe one could claim the ML model itself is the scientist and everyone else is just feeding it data. I doubt human scientists would be comfortable with that, but if they aren't trying to explain anything, what are they even doing?
> If you're the sort of person who believes that human brains are capable of understanding the "why" of how things work in all its true detail, you'll find this an interesting challenge- can we actually interpret these models, or are human brains too feeble to understand complex systems without sophisticated models?
I think chess engines, weirdly enough, have disabused me of this notion.
There are lots of factors a human considers when looking at a board. Piece activity. Bishop and knight imbalances. King safety. Open and semi-open file control. Tempo. And on and on.
But all of them are just convenient shortcuts that allow us to substitute reasonable guesses for what really matters: exhaustively calculating a winning line through to the end. “Positional play” is a model that only matters when you can’t calculate trillions of lines thirty moves deep, and it’s infinitely more important that a move survives your opponent’s best possible responses than it is to satisfy some cohesive higher level principle.
I'm in the following camp:
It is wrong to think about the world or the models as "complex systems" that may or may not be understood by human intelligence. There is no meaning beyond that which is created by humans. There is no 'truth' that we can grasp in parts but not entirely.
Being unable to understand these complex systems means that we have framed them in such a way (f.e. millions of matrix operations) that does not allow for our symbol-based, causal reasoning mode. That is on us, not our capabilities or the universe.
All our theories are built on observation, so these empirical models yielding such useful results is a great thing - it satisfies the need for observing and acting. Missing explainability of the models merely means we have less ability to act more precisely - but it does not devalue our ability to act coarsely.
> ... and strongly believe that science's goal is to produce falsifiable hypotheses, these new approaches will be extremely depressing
I don't quite understand this point — could you elaborate?
My understanding is that the ML model produces a hypothesis, which can then be tested via normal scientific method (perform experiment, observe results).
If we have a magic oracle that says "try this, it will work", and then we try it, and it works, we still got something falsifiable out of it.
Or is your point that we won't necessarily have a coherent/elegant explanation for why it works?
What if our understanding of the laws of the natural sciences are subtly flawed and AI just corrects perfectly for our flawed understanding without telling us what the error in our theory was?
Forget trying to understand dark matter. Just use this model to correct for how the universe works. What is actually wrong with our current model and if dark matter exists or not or something else is causing things doesn't matter. "Shut up and calculate" becomes "Shut up and do inference."
There have been times in the past when usable technology surpassed our scientific understanding, and instead of being depressing it provided a map for scientific exploration. For example, the steam engine was developed by engineers in the 1600s/1700s (Savery, Newcomen, and others) but thermodynamics wasn’t developed by scientists until the 1800s (Carnot, Rankine, and others).
What always struck me about Chomskyists is that they chose a notion of interpretable model that required unrealistic amounts of working interpretation. So Chomsky grammars have significant polynomial memory and computational costs for grammars as they approach something resembling human grammar. And you say, ok, the human brain can handle much more computation than that, and that's fine. But (for example) context-free grammars aren't just O(n^3) in computational cost; for a realistic description of human language they're O(n^3) in human-interpretable rules.
Other Chomsky-like models of human grammars have different asymptotic behavior and different choices of n, but the same fundamental problem; the big-O constant factor isn't neurons firing but rather human connections between the n inputs. How can you conceive of human minds being able to track O(n^3) (or whatever) cost where that n is everything being communicated -- words, concepts, symbols, representations, all that jazz and the polynomial relationships between them?
But I feel an apology is in order: I've had quite a few beers before coming home, and it's probably a mistake to try to express academically charged and difficult views on the Internet while in an inebriated state. Probably the alcohol has substantially decreased my mental computational power. However, it has only mildly impaired my ability to string together words and sentences in a grammatically complex fashion. In fact, I often feel that the more sober and clear-minded I am, the simpler my language is. Maybe human grammar is actually sub-polynomial. I have observed the same in ChatGPT; the more flowery and wordy it has become over time, the dumber its output.
> If you're the sort of person who believes that human brains are capable of understanding the "why" of how things work in all its true detail
This seems to me an empirical question about the world. It’s clear our minds are limited, and we understand complex phenomena through abstraction. So either we discover we can continue converting advanced models to simpler abstractions we can understand, or that’s impossible. Either way, it’s something we’ll find out and will have to live with in the coming decades. If it turns out further abstractions aren’t possible, well, enlightenment thought had lasted long enough. It’s exciting to live at a time in humanity’s history when we enter a totally uncharted new paradigm.
> can we actually interpret these models, or are human brains too feeble to understand complex systems without sophisticated models?
I think we will have to develop a methodology and supporting toolset to be able to derive the underlying patterns driving such ML models. It's just too much for a human to comb through by themselves and make sense of.
So the work to simplify ML models, reduce dimensions, etc. becomes the numeric way to seek simple actual scientific models. Scientific computing and science become one.
The frontier in model space is kind of fluid. It's all about solving differential equations.
In theoretical physics, you know the equations, you solve equations analytically, but you can only do that when the model is simple.
In numerical physics, you know the equations, you discretize the problem on a grid, and you solve the constraint defined by the equations with various numerical integration schemes like RK4, but you can only do that when the model is small and you know the equations, and you find a single solution.
Then you want the result faster, so you use mesh-free methods and adaptive grids. It works on bigger models but you have to know the equations, finding a single solution to the differential equations.
Then you compress this adaptive grid with a neural network, while still knowing the governing equations, and you have things like Physics Informed Neural Networks ( https://arxiv.org/pdf/1711.10561 and following papers) where you can bound the approximation error. This method allows solve all solutions to the differential equations simultaneously, sharing the computations.
Then when knowing explicitly your governing equations is too complex, so you assume that there are some governing stochastic equations implicitly, which you learn the end-result of the dynamic with a diffusion model, that's what this alpha-fold is doing.
ML is kind of a memoization technique, analog to hashlife in the game of life, that allows you reuse your past computational efforts. You are free to choose on this ladder which memory-compute trade-off you want to use to model the world.
As a steelman, wouldn't the abundance of infinitely generate-able situations make it _easier_ for us to develop strong theories and models? The bottleneck has always been data. You have to do expensive work in the real world and accurately measure it before you can start fitting lines to it. If we were to birth an e.g. atomically accurate ML model of quantum physics, I bet it wouldn't take long until we have mathematical theories that explain why it works. Our current problem is that this stuff is super hard to manipulate and measure.
This is an important aspect that's being ignored IMO.
For a lot of problems, currently you either don't have an an analytical solution and the alternative is a brute force-ish numerical approach. As a result the computational cost of simulating things enough times to be able to detect behavior that can inform theories/models (potentially yielding a good analytical result) is not viable.
It depends whether the value of science is human understanding or pure prediction. In some realms (for drug discovery, and other situations where we just need an answer and know what works and what doesn’t), pure prediction is all we really need. But if we could build an uninterpretable machine learning model that beats any hand-built traditional ‘physics’ model, would it really be physics?
Maybe there’ll be an intermediate era for a while where ML models outperform traditional analytical science, but then eventually we’ll still be able to find the (hopefully limited in number) principles from which it can all be derived. I don’t think we’ll ever find that Occam’s razor is no use to us.
> But if we could build an uninterpretable machine learning model that beats any hand-built traditional ‘physics’ model, would it really be physics?
At that point I wonder if it would be possible to feed that uninterpretable model back into another model that makes sense of it all and outputs sets of equations that humans could understand.
The success of these ML models has me wondering if this is what Quantum Mechanics is. QM is notoriously difficult to interpret yet makes amazing predictions. Maybe wave functions are just really good at predicting system behavior but don't reflect the underlying way things work.
OTOH, Newtonian mechanics is great at predicting things under certain circumstances yet, in the same way, doesn't necessarily reflect the underlying mechanism of the system.
So maybe philosophers will eventually tell us the distinction we are trying to draw, although intuitive, isn't real
Pure prediction is only all we need if the total end-to-end process is predicted correctly - otherwise there could be pretty nasty traps (e.g., drug works perfectly for the target disease but does something unexpected elsewhere etc.).
In case it's not clear, this does not "beat" experimental structure determination. The matches to experiment are pretty close, but they will be closer in some cases than others and may or may not be close enough to answer a given question about the biochemistry. It certainly doesn't give much information about the dynamics or chemical perturbations that might be relevant in biological context. That's not to pooh-pooh alphafold's utility, just that it's a long way from making experimental structure determination unnecessary, and much much further away from replacing a carefully chosen scientific question and careful experimental design.
It means we now have an accurate surrogate model or "digital twin" that can be experimented on almost instantaneously. So we can massively accelerate the traditional process of developing mechanistic understanding through experiment, while also immediately be able to benefit from the ability to make accurate predictions, even without needing understanding.
In reality, science has already pretty much gone this way long ago, even if people don't like to admit it. Simple, reductionist explanations for complex phenomena in living systems don't really exist. Virtually all of medicine nowadays is empirical: try something, and if you can prove its safe and effective, you keep doing it. We almost never have a meaningful explanation for how it really works, and when we think we do, it gets proven wrong repeatedly, while the treatment keeps working as always.
Medicine can be explained fairly simply, and the why of how it works as it does is also explained by this:
Imagine a very large room that has every surface covered by on-off switches.
We cannot see inside of this room. We cannot see the switches. We cannot fit inside of this room, but a toddler fits through the tiny opening leading into the room. The toddler cannot reach the switches, so we equip the toddler with a pole that can flip the switches. We train the toddler, as much as possible, to flip a switch using the pole.
Then, we send the toddler into the room and ask the toddler to flip the switch or switches we desire to be flipped, and then do tests on the wires coming out of the room to see if the switches were flipped correctly. We also devise some tests for other wires to see if that naughty toddler flipped other switches on or off.
We cannot see inside the room. We cannot monitor the toddler. We can't know what _exactly_ the toddler did inside the room.
That room is the human body. The toddler with a pole is a medication.
We can't see or know enough to determine what was activated or deactivated. We can invent tests to narrow the scope of what was done, but the tests can never be 100% accurate because we can't test for every effect possible.
We introduce chemicals then we hope-&-pray that the chemicals only turned on or off the things we wanted turned on or off. Craft some qualifications testing for proofs, and do a 'long-term' study to determine if there were other things turned on or off, or a short circuit occurred, or we broke something.
I sincerely hope that even without human understanding, our AI models can determine what switches are present, which ones are on and off, and how best to go about selecting for the correct result.
Right now, modern medicine is almost a complete crap-shoot. Hopefully modern AI utilities can remedy the gambling aspect of medicine discovery and use.
It makes me think about how Einstein was famous for making falsifiable real-world predictions to accompany his theoretical work. And, sometimes it took years for proper experiments to be run (such as measuring a solar eclipse during the breakout of a world war).
Perhaps the opportunity here is to provide a quicker feedback loop for theory about predictions in the real world. Almost like unit tests.
> Perhaps the opportunity here is to provide a quicker feedback loop for theory about predictions in the real world. Almost like unit tests.
Or jumping the gap entirely to move towards more self-driven reinforcement learning.
Could one structure the training setup to be able to design its own experiments, make predictions, collect data, compare results, and adjust weights...? If that loop could be closed, then it feels like that would be a very powerful jump indeed.
In the area of LLMs, the SPAG paper from last week was very interesting on this topic, and I'm very interested in seeing how this can be expanded to other areas:
Many of our existing physical models can be decomposed into "high-confidence, well tested bit" plus "hand-wavy empirically fitted bit". I'd like to see progress via ML replacing the empirical part - the real scientific advancement then becomes steadily reducing that contribution to the whole by improving the robust physical model incrementally. Computational performance is another big influence though. Replacing the whole of a simulation with an ML model might still make sense if the model training is transferrable and we can take advantage of the GPU speed-ups, which might not be so easy to apply to the foundational physical model solution. Whether your model needs to be verified against real physical models depends on the seriousness of your use-case; for nuclear weapons and aerospace weather forecasts I imagine it will remain essential, while for a lot of consumer-facing things the ML will be good enough.
Physics-informed machine learning is a whole (nascent) subfield that is very much in line with this thinking. Steve Brunton has some good stuff about this on YouTube.
"Best methods" is doing a lot of heavy lifting here. "Best" is a very multidimensional thing, with different priorities leading to different "bests." Someone will inevitably prioritize reliability/accuracy/fidelity/interpretability, and that's probably going to be a significant segment of the sciences. Maybe it's like how engineers just need an approximation that's predictive enough to build with, but scientists still want to understand the underlying phenomena. There will be an analogy to how some people just want an opaque model that works on a restricted domain for their purposes, but others will be interested in clearer models or unrestricted/less restricted domain models.
It could lead to a very interesting ecosystem of roles.
Even if you just limit the discussion to using the best model of X to design a better Y, limited to the model's domain of validity, that might translate the usage problem to finding argmax_X of valueFunction of modelPrediction of design of X. In some sense a good predictive model is enough to solve this with brute force, but this still leaves room for tons of fascinating foundational work. Maybe you start to find that the (wow so small) errors in modelPrediction are correlated with valueFunction, so the most accurate predictions don't make it the best for argmax (aka optimization might exploit model errors rather than optimizing the real thing). Or maybe brute force just isn't computationally feasible, so you need to understand something deeper about the problem to simplify the optimization to make it cheap.
Physicists like to retroactively believe that our understanding of physical phenomena preceded the implementation of uses of those phenomena, when the reality is that physics has always come in to clean up after the engineers. There are some rare exceptions, but usually the reason that scientific progress can be made in an area is that the equipment to perform experiments has been commoditized sufficiently by engineering demand for it.
We had semiconductors and superconductors before we understood how they worked -- on both cases arguably we still don't completely understand the phenomena. Things like the dynamo and the electric motor were invented by practice and later explained by scientists, not derived from first principles. Steam engines and pumps were invented before we had the physics to describe how they worked.
The most moneyed and well-coordinated organizations have honed a large hammer, and they are going to use it for everything, and so almost certainly future big findings in the areas you mention, probabilistically inclined models coming from ML will be the new gold standard.
But yet the only thing that can save us from ML will be ML itself because it is ML that has the best chance to be able to extrapolate patterns from these blackbox models to develop human interpretable models. I hope we do dedicate explicit effort to this endeavor, and so continue the human advances and expanse of human knowledge in tandem with human ingenuity with computers at our assistance.
Spoiler: "Interpretable ML" will optimize for output that either looks plausible to humans, reinforces our preconceptions, or appeals to our aesthetic instincts. It will not converge with reality.
It's interesting to compare this situation to earlier eras in science. Newton, for example, gave us equations that were very accurate but left us with no understanding at all of why they were accurate.
It seems like we're repeating that here, albeit with wildly different methods. We're getting better models but by giving up on the possibility of actually understanding things from first principles.
Not comparable. Our current knowledge of the physics involved in these systems is complete. It is just impossibly difficult to calculate from first principles.
A new-ish field of "mechanistic interpretability" is trying to poke at weights and activations and find human-interpretable ideas w/in them. Making lots of progress lately, and there are some folks trying to apply ideas from the field to Alphafold 2. There are hopes of learning the ideas about biology/molecular interactions that the model has "discovered".
Perhaps we're in an early stage of Ted Chiang's story "The Evolution of Human Science", where AIs have largely taken over scientific research and a field of "meta-science" developed where humans translate AI research into more human-interpretable artifacts.
1. Research can then focus on where things go wrong
2. ML models, despite being "black boxes," can still have brute-force assessment performed of the parameter space over covered and uncovered areas by input information
3. We tend to assume parsimony (i.e Occam's razor) to give preference to simpler models when all else is equal. More complex black-box models exceeding in prediction let us know the actual causal pathway may be more complex than simple models allow. This is okay too. We'll get it figured out. Not everything is closed-form, especially considering quantum effects may cause statistical/expected outcomes instead of deterministic outcomes.
Interesting times indeed. I think the early history of medicines takes away from your observation though. In the 19th and early 20th century people didn't know why medicines worked, they just did. The whole "try a bunch of things on mice, pick the best ones and try them on pigs, and then the best of those and try a few on people" kind of thing. In many ways the mice were a stand in for these models, at the time scientists didn't understand nearly as much about how mice worked (early mice models were pretty crude by today's standards) but they knew they were a close enough analog to the "real thing" that the information provided by mouse studies was usefully translated into things that might help/harm humans.
So when you're tools can produce outputs that you find useful, you can then use those tools to develop your understanding and insights. As a tool, this is quite good.
I asked a friend of mine who is chemistry professor at a large research university something along these lines a while ago. He said that so far these models don't work well in regions where either theory or data is scarce, which is where most progress happens. So he felt that until they can start making progress in those areas it won't change things much.
This is the topic of epistemology of the sciences in books such as "New Direction in the Philosophy of Mathematics" [1] and happened before with problems such as the four color theorem [2] where AI was not involved.
Going back to the uninterpretable ML models in the context of AlphaFold 3, I think one method for trying to explain the findings is similar to the experimental methods of physics with reality: you perform experiments with the reality (in this case AlphaFold 3) to came up with sound conclusions. AI/ML is an interesting black-box system.
There are other open discussions on this topic. For example, can our human brain absorbe that knowledge or it is limited somehow with the scientific language that we have now?
In physics, we already deal with the fact that many of the core equations cannot be analytically solved for more than the most basic scenarios. We've had to adapt to using approximation methods and numerical methods. This will have to be another place where we adapt to a practical way of getting results.
Reminds me of the novel Blindsight - in it there's special individuals who work as synthesists, whos job it is to observe and understand and then somehow translate back to "lay person" the seemingly undecipherable actions/decisions of advanced computers and augmented humans.
I'd say it's not new. Take fluid dynamics as an example, the navier stokes equations predict the motion of fluids very well but you need to approximately solve them on a computer in order to get useful predictions for most setups. I guess the difference is the equation is compact and the derivation from continuum mechanics is easy enough to follow. People still rely on heuristics to answer "how does a wing produce lift?". These heuristic models are completely useless at "how much lift will this particular wing produce under these conditions?". Seems like the same kind of situation.
Maybe progress forward will look like producing compact models or tooling to reason about why a particular thing happened.
I think it likely that instead of replacing existing methods, we will see a fusion. Or rather, many different kinds of fusions - depending on the exact needs of the problems at hand (or in science, the current boundary of knowledge). If nothing else then to provide appropriate/desirable level of explainability, correctness etc. Hypothetically the combination will also have better predictive performance and be more data efficient - but it remains to be seen how well this plays out in practice. The field of "physics informed machine learning" is all about this.
Is alphafold doing model generation or is it just reducing a massive state space?
The current computational and systems biochemistry approaches struggle to model large biomolecules and their interactions due to the large degrees of freedom of the models.
I think it is reasonable to rely on statistical methods to lead researchers down paths that have a high likelihood of being correct versus brute forcing the chemical kinetics.
Our metaphors and intuitions were crumbling already and stagnating. See quantum physics: sometimes a particle, sometimes a wave, and what constitute a measurement anyway?
I’ll take prediction over understanding if that’s the best our brains can do. We’ve evolved to deal with a few orders of magnitude around a meter and a second. Maybe dealing with light-years and femtometer/seconds is too much to ask.
> Does this decouple progress from our current understanding of the scientific process - moving to better and better models of the world without human-interpretable theories and mathematical models / explanations?
Replace "human-interpretable theories" with "every man interpretable theories", and you'll have a pretty good idea of how > 90% of the world feels about modern science. It is indistinguishable from magic, by the common measure.
Obtuse example: My parents were alive when the first nuclear weapon was detonated. They didn't know that they didn't know this weapon was being built, let alone that it might have ignited the atmosphere.
With sophisticated enough ML, that 90% will become 99.9% - save the few who have access to (and can trust) ML tools that can decipher the "logic" from the original ML tools.
"better and better models of the world" does not always mean "more accurate" and never has.
We already know how to model the vast majority of things, just not at a speed and cost which makes it worthwhile. There are dimensions of value - one is accuracy, another speed, another cost, and in different domains additional dimensions. There are all kinds of models used in different disciplines which are empirical and not completely understood. Reducing things to the lowest level of physics and building up models from there has never been the only approach. Biology, geology, weather, materials all have models which have hacks in them, known simplifications, statistical approximations, so the result can be calculated. It's just about choosing the best hacks to get the best trade off of time/money/accuracy.
This is a key but secondary concern to many of us working in molecular geneticist who will use AlphaFold 3 to evaluate pair-wise interactions. We often have genetic support for an interaction between proteins A and B. For example, in a study of genetic variation in responses of mice to morphine I currently have two candidate proteins that interact epistatically, suggesting a possible “lock and key” model—-the mu opiate receptor (MOR) and FGF12. I can now evaluate the likelihood of a direct molecular interaction between these proteins and possible amino acids substitutions that account for individuals difference.
In other words I bring a hypothesis to AF3 and ask for it to refute or affirm.
You are conflating the whole scientific endeavor to a very specific problem to which this specific approach is effective at producing results that fit with the observable world.
This has nothing to do with science as a whole.
I think it is fine & better for society to have applications and models for things we don't fully understand... We can model lots of small aspects of weather, and we have a lot of factors nailed down, but not necessarily all the interactions.. and not all of the factors. (Additional example for the same reason: Gravity)
Used responsibly. Of course. I wouldn't think an AI model designing an airplane that no engineers understand how it works is a good idea :-)
And presumably all of this is followed by people trying to understand the results (expanding potential research areas)
To paraphrase Kahan, it's not interesting to me whether a method is accurate enough or not, but whether you can predict how accurate you can be. So, if ML methods can predict that they're right 98% of times then we can build this in our systems, even if we don't understand how they work.
Deterministic methods can predict result with a single run, ML methods will need ensemble of results to show the same confidence. It is possible at the end of day that the difference in cost might not he that high over time.
Science has always given us better, but error prone tooling to see further and make better guesses. There is still a scientific test. In a clinical trial, is this new drug safe and effective.
Perhaps an ai can be made to produce the work as well as a final answer, even if it has to reconstruct or invent the work backwards rather than explain it's own internal inscrutable process.
"produce a process that arrives at this result" should be just another answer it can spit out. We don't necessarily care if the answer it produces is actually the same as what originally happened inside itself. All we need is that the answer checks out when we try it.
No, science doesn't work that way. You can just calculate your way to scientific discoveries, you got to test them in the real world. Learning, both in humans and AI, is based on the signals provided by the environment. There are plenty of things not written anywhere, so the models can't simply train on human text to discover new things. They learn directly from the environment to do that, like AlphaZero did when it beat humans at Go.
I think at some point, we will be able to produce models that are able to pass data into a target model and observe its activations and outputs and put together some interpretable pattern or loose set of rules that govern the input-output relationship in the target model. Using this on a model like AlphaFold might enable us to translate inferred chemical laws into natural language.
Even if we don’t understand the models themselves, you can still use them as a basis for understanding
For example, I have no idea how a computer works in every minute detail (ie, exactly the physics and chemistry of every process that happens in real time), but I have enough of an understanding of what to do with it, that I can use it as an incredibly useful tool for many things
Not the same. There is a difference between "I cannot understand the deeper details of certain model but some others can and there's the possibility of explaining it in detail" and "Nobody can understand it and there's not a clear cause-effect that we know" .
Except for weird cases, computers (or cars, or cameras, or lots of other man made devices) are clearly known and you (or another specialist) can clearly show why a device does X when you input Y on it.
> Does this decouple progress from our current understanding of the scientific process?
Thank God! As a person who uses my brain, I think I can say, pretty definitively, that people are bad at understanding things.
If this actually pans out, it means we will have harnessed knowledge/truth as a fundamental force, like fire or electricity. The "black box" as a building block.
I can only assume that existing methods would still be used for verification. At least we understand the logic used behind these methods. The ML models might become more accurate on average but they could still throw out results that are way off occasionally, so their error rate would have to become equal to the existing methods.
I wonder if ML can someday be employed in deciphering such black box problems; a second model that can look under the hood at all the number crunching performed by the predictive model, identify the pattern that resulted in a prediction, and present it in a way we can understand.
That said, I don’t even know if ML is good at finding patterns in data.
The models are learning an encoding based on evolutionary related and known structures. We should be able to derive fundamental properties from those encodings eventually. Or at least our biophysical programmed models should map into that encoding. That might be a reasonable approach to look at the folding energy landscape.
In terms of docking, you can call the conventional approaches "physically-based", however, they are rather poor physical models. Namely, they lack proper electrostatics, and, most importantly, basically ignore entropic contributions. There is no reason for concern.
Might be easier to come up with new models with analytic solutions if you have a probabilistic model at hand. A lot easier to evaluate against data and iterate. Also, I wouldn't be surprised if we develop better tools for introspecting these models over time.
Perhaps for understanding the structure itself, but having the structure available allows us to focus on a coarser level. We also don't want to use quantum mechanics to understand the everyday world, and that's why we have classic mechanics etc.
These processes are both beyond human comprehension because they contain vast layers of tiny interactions and also not practical to simulate. This tech will allow for exploration for accurate simulations to better understand new ideas if needed.
I'm not a scientist by any means, but I imagine even accurate opaque models can be useful in moving the knowledge forward. For example, they can allow you to accurately simulate reality, making experiments faster and cheaper to execute.
There will be an iterative process built around curated training datasets - continually improved, top tier models, teams reverse engineering the model's understanding and reasoning, and applying that to improve datasets and training.
This is a neat observation. Slightly terrifying, but still interesting. Seems like there will also be cases where we discover new theories through the uninterpretable models—much easier and faster to experiment endlessly with a computer.
I think it creates new studies, such as diagnosing these models behaviors without the doctor having an intricate understanding of all of the model's processes/states just like with natural organisms
I would assume that given enough hints from AI and if it is deemed important enough humans will come in to figure out the “first principles” required to arrive at the conclusion.
I believe this is the case also. With a well enough performing AI/ML/probabilistic model where you can change the model's input parameters and get a highly accurate prediction basically instantly, we can test theories approximately and extremely fast rather than running completely new experiments, which will always come with it's own set of errors and problems.
every time the two systems disagree, it's an opportunity to learn something. both kinds of models can be improved with new information, done through real-world experiments
Hook the protein model up to an LLM model, have the LLM interpret the results. Problem solved :-) Then we just have to trust the LLM is giving us correct interpretations.
We will get better with understanding black boxes, if a model can be compressed into simple math formula then it's both easier to understand and to compute.
Is it capable of predictions though? Ie can it accurately predict the folding of new molecules? Otherwise how do you distinguish accuracy from overfitting.
This is exactly how the physicists felt at the dawn of quantum physics - the loss of meaningful human inquiry to blindly effective statistics. Sobering stuff…
Personally, I’m convinced that human reason is less pure than we think it to be, and that the move to large mathematical models might just be formalizing a lack-of-control that was always there. But that’s less of a philosophy of science discussion and more of a cognitive science one
We already have the absolute best method for accurately predicting the world, and it is by experimentation. In the protein folding case, it works by actually making the protein and analyzing it. For designing airplanes, computer models are no match for building the thing, or even using physical models and wind tunnels.
And despite having these "best method", it didn't prevent progress in theoretical physics, theory and experimentation complement each other.
ML models are just another kind of model that can help both engineering and fundamental research. Their working is close to the old guy in the shop who knows intuitively what is good design, because he has seen it all. That old guys in shops are sometimes better than modeling using physics equations help scientific progress, as scientists can work together with the old guy, combining the strength of intuition and experience with that of scientific reasoning.
It’s much easier to reverse engineer a solution that you don’t understand (and discover important underlying theories on that journey), than it is to arrive at that same solution and the underlying theories without knowing in advance where you are going.
For this reason, discoveries made by AI will be immensely useful for accelerating scientific progress, even if those discoveries are opaque at first.
What would a universe look like that lacked simple things, and somehow only complex things existed?
It makes me think of how Gaussian integers have irreducibles but not prime numbers, where some large things cannot be uniquely expressed as combination of smaller things.
what happens is an opportunity has entered the chat.
There is a wave coming—I won't try to predict if it's the next one—where the hot thing in AI/ML is going to be profoundly powerful tools for analyze other such tools and render them intelligible to us,
which will I imagine mean providing something like a zoomable explainer. At every level there are footnotes; if you want to understand why the simplified model is a simplification, you look at the fine print. Which has fine print. Which has...
Which doesn't mean there is not a stable level at which some formal notion of "accurate" cannot be said to exist, which is the minimum viable level of simplification.
Etc.
This sort of thing will of course will the input to many other things.
So it’s okay now to publish a computational paper with no code? I guess Nature’s reporting standards don’t apply to everyone.
> A condition of publication in a Nature Portfolio journal is that authors are required to make materials, data, code, and associated protocols promptly available to readers without undue qualifications.
> Authors must make available upon request, to editors and reviewers, any previously unreported custom computer code or algorithm used to generate results that are reported in the paper and central to its main claims.
If you read the standards it applies broadly beyond reviewers or editors.
> A condition of publication in a Nature Portfolio journal is that authors are required to make materials, data, code, and associated protocols promptly available to readers without undue qualifications.
Also makes me wonder -- where's the line? Is it reasonable to have "layperson" reviewers? Is it reasonable to think that regular citizens could review such content?
This reminds me of Google’s claim that another “AI” discovered millions of new materials. The results turned out to be a lot of useless noise but that was only apparent after actual expert spent hundreds of hours reviewed the results[0]
The alphafold work has been used across the industry (successfully, in the sense of blind prediction), and has been replicated independently. The work on alphafold will likely net Demis and John a Nobel prize in the next few years.
(that said, one should always inspect Google publications with a fine-toothed comb and lots of skepticism, as they have a tendency to juice the results)
"In many cases, AlphaFold predictions matched experimental maps remarkably closely. In other cases, even very high-confidence predictions differed from experimental maps on a global scale through distortion and domain orientation, and on a local scale in backbone and side-chain conformation. We suggest considering AlphaFold predictions as exceptionally useful hypotheses."
> We have yet to find any strikingly novel compounds in the GNoME and Stable Structure listings, although we anticipate that there must be some among the 384,870 compositions. We also note that, while many of the new compositions are trivial adaptations of known materials, the computational approach delivers credible overall compositions, which gives us confidence that the underlying approach is sound.
AlphaFold has been widely validated- it's now appreciated that its predictions are pretty damn good, with a few important exceptions, instances of which are addressed with the newer implementation.
So... what percentage of the time? If you made an AI to pilot an airplane, how would you verify its edge conditions, you know, like plummeting out of the sky because it thought it had to nosedive?
Because these AIs are black box neural networks, how do you know they are predicting things correctly for things that aren't in the training dataset?
So after 6 years of this "revolutionary technology", what we have to show for all the hype and breathless press releases is: ....another press release saying how "revolutionary" it is. Fantastic. Thanks DeepMind.
I don't blame them for hyping their products - if only to fight the sentiment that Google is far behind OpenAI because they were not first to release a LLM.
I’m inclined to ignore such pr fluff until they actually demonstrate a _practical_ result. Eg. cure some form of cancer or some autoimmune disease. All this “prediction of structure” has been in the news for years, and it seems to have resulted in nothing practically usable IRL as far as I can tell. I could be wrong of course, I do not work in this field
the R&D of all major pharma is currently using AlphaFold predictions when they don't have experimentally determined structures. I cannot share further details but the results suggest that we will see future pharmaceuticals based on AF predictions.
The important thing to recognize is that protein structures are primarily hypothesis-generation machines and tools to stimulate ideas, rather that direct targets of computational docking. Currently structures rarely capture the salient details required to identify a molecule that has precisely the biological outcome desired, because the biological outcome is an extremely complex function that incorporates a wide array of other details, such as other proteins, metabolism, and more.
Sure. If/when we see anything practical, that’ll be the right moment to pay attention. This is much like “quantum computing” where everyone who doesn’t know what it is is excited for some reason, and those that do know can’t even articulate any practical applications
Probably worth mentioning that David Baker’s lab released a similar model (predicts protein structure along with bound DNA and ligands), just a couple of months ago, and it is open source [1].
It’s also worth remembering that it was David Baker who originally came up with the idea of extending AlphaFold from predicting just proteins to predicting ligands as well [2].
1. https://github.com/baker-laboratory/RoseTTAFold-All-Atom
2. https://alexcarlin.bearblog.dev/generalized/
Unlike AlphaFold 3, which predicts only a small, preselected subset of ligands, RosettaFold All Atom predicts a much wider range of small molecules. While I am certain that neither network is up to the task of designing an enzyme, these are exciting steps.
One of the more exciting aspects of the RosettaFold paper is that they train the model for predicting structures, but then also use the structure predicting model as the denoising model in a diffusion process, enabling them to actually design new functional proteins. Presumably, DeepMind is working on this problem as well.
I appreciated this, but it's probably worth mentioning: when you say AlphaFold 3, you're talking about AlphaFold 2.
TFA announces AlphaFold 3.
Post: "Unlike AlphaFold 3, which predicts only a small, preselected subset of ligands, RosettaFold All Atom predicts a much wider range of small molecules"
TFA: "AlphaFold 3...*models large biomolecules such as proteins, DNA and RNA*, as well as small molecules, also known as ligands"
Post: "they also use the structure predicting model as the denoising model in a diffusion process...Presumably, DeepMind is working on this problem as well."
TFA: "AlphaFold 3 assembles its predictions using a diffusion network, akin to those found in AI image generators."
And that tech just got $1b in funding.
can you expand? Who got 1bn of funding?
Coming up with ideas is cheaper than executing the ideas. Predicting a wide range of molecules okay-ish is cheaper than predicting a small range of molecules very well.
Important caveat: it's only about 70% accurate. Why doesn't the press release say this explicitly? It seems intentionally misleading to only report accuracy relative to existing methods, which apparently are just not so good (30%, 50% in various settings). https://www.fastcompany.com/91120456/deepmind-alphafold-3-dn...
They also had a headline for Alphazero that convinced everyone that they crushed Stockfish and that classical chess engines were stuff of the past, when in fact it was about 50 elo better than the Stockfish version they were testing against, or roughly the same as how much Stockfish improves each year.
I think Alphazero is a lot more interesting than Stockfish though. Most notably it lead me to reevaluate positional play. Iirc A0 at around 2-3 ply is still above SuperGM Level which is pretty mind-blowing. Based on this I have increased my strategy to tactics ratio quite a bit. FWIW Stockfish is always evolving and adapting and has incorporated ideas from A0.
7 replies →
Alphazero indeed crushed stockfish with a novel technique, I think it deserved all the praise.
1 reply →
And then what happened is AlphaZero changed the professional game in various interesting ways, and all its ideas were absorbed into Stockfish. A little bombast is forgivable for technology that goes on to have a big impact, and I don’t doubt it’s the same story here.
10 replies →
That's what I thought. They go from "predicting all of life's molecules" to "it's a 50% improvement...and we HOPE to...transform drug discovery..."
Seems unfortunately typical of Google these days: "Gemini will destroy GPT-4..."
IIRC the next best models all have all been using AlphaFold 2's methodology, so that's still a massive improvement.
Edit: I see now that you're probably objecting to the headline that got edited on HN.
Not just the headline, the whole press release. And not questioning that it's a big improvement.
That's pretty good. Based on the previous performance improvements of Alpha-- models, it'll be nearing 100% in the next couple of years.
> it'll be nearing 100% in the next couple of years.
What are you basing this on? There is no established "moores law" for computational models.
3 replies →
Just "Alpha-- models" in general?? That's not a remotely reasonable way to reason about it. Even if it were, why should it stop DeepMind from clearly communicating accuracy?
7 replies →
Which specific AlphaX model evolved like that? Most of the ones that were in the press had essentially a single showing, typically very good, but didn't really improve after that.
Very sad to see they did not make it open source. When you have a technology that has the potential to be a gateway for drug development, to the cures of new diseases, and instead you choose to make it closed, it is a very huge disservice to the community at large. Sure, release your own product alongside it, but making it closed source does not help the scientific community upon which all these innovations were built. Especially if you have lost a loved one to a disease which this technology will one day be able to create cures for, it is very disappointing.
I suspect https://www.isomorphiclabs.com/ is the reason.
There are 3 basic ways to fund research.
- Taxes - most academic research
- begging - research charities
- profits - companies like Google.
Sometimes the lines get blurred - but I don't think you can expect Google to release as much of their work for free as people who are paid via central taxes.
Worth noting that they did release the AlphaFold 2 weights after a while. Milking an expensive discovery for a limited period should be considered laudable, unless you think tax funding of all research would be awesome and it's just a weird anomaly how the org producing these results was a tiny heterodox startup very recently.
4 replies →
I'd be willing to bet that OpenFold has an implementation inside of a few months.
https://openfold.io/
Openness and collaboration could have far-reaching implications for public health and well-being but there are lots of aspects of being open
> the community at large
Which community? Not any I'm part of.
The closer it gets to enabling full drug discovery, the closer it also gets to enabling bioterrorism. Taking it to the extreme, if they had the theory of everything, I don't think I'd want it to be made available to the whole world as it is today.
On a related note, I highly recommend The Talos Principle 2, which really made me think about these questions.
Any organization/country that has the ability to use a tool like this to create a bio weapon is already sophisticated enough to do bioterrorism today.
26 replies →
Why do you need AI for bioterrorism? There are plenty of well known biological organisms that can kill us today...
Oh please, like a terrorist cant fork over a couple bucks to do the bioterrorism. This excuse is utter bs, whether its applied to LLMs or to alphafold. The motivator is profit, not safety.
Why on Earth would any terrorist want to invest into this when you can just purchase guns and explosives much easier?
1 reply →
> What is different about the new AlphaFold3 model compared to AlphaFold2?
> AlphaFold3 can predict many biomolecules in addition to proteins. AlphaFold2 predicts structures of proteins and protein-protein complexes. AlphaFold3 can generate predictions containing proteins, DNA, RNA, ions,ligands, and chemical modifications. The new model also improves the protein complex modelling accuracy. Please refer to our paper for more information on performance improvements.
AlphaFold 2 generally produces looping “ribbon-like” predictions for disordered regions. AlphaFold3 also does this, but will occasionally output segments with secondary structure within disordered regions instead, mostly spurious alpha helices with very low confidence (pLDDT) and inconsistent position across predictions.
So the criticism towards AlphaFold 2 will likely still apply? For example, it’s more accurate for predicting structures similar to existing ones, and fails at novel patterns?
I am not aware of anybody currently criticiszing AF2's abilities outside of its training set. In fact the most recent papers (written by crystallographers) they are mostly arguing about atomic-level details of side chains at this point.
problem is biomolecules, are "chaperoned" to fold properly, only specific regions such as, alpha helix, or beta pleatedsheet will fold de novo.
Chaperone (protein)
https://en.wikipedia.org/wiki/Chaperone_(protein)
Chaperones exist - however many proteins will quite happily fold in isolation without any external help.
2 replies →
In principle couldn’t we just incorporate knowledge about chaperones into the model?
6 replies →
>So the criticism towards AlphaFold 2 will likely still apply? For example, it’s more accurate for predicting structures similar to existing ones, and fails at novel patterns?
Yes, and there is simply no way to bridge that gap with this technique. We can make it better and better at pattern matching, but it is not going to predict novel folds.
alphafold has been shown to accurately predict some novel folds. The technique doesn't entirely depend on whole-domain homology.
> but it is not going to predict novel folds
https://www.nature.com/articles/s42003-022-03357-1
As a software engineer, I kind of feel uncomfortable about this new model. It outperforms Alphafold 2 at ligand binding, but Alphafold 2 also had some more hardcoded and interpretable structural reasoning baked into the model architecture.
There's so many things you can incorporate into a protein folding model such as structural constraints, rotational equivariance, etc, etc
This new model simple does away with some of that, achieving greater results. And the authors simply use distillation from data outputted from Alphafold2 and Alphafold2-multimer to get those better results for those cases where you wind up with implausible results.
You have to run all those previous models, and output their predictions to do the distillation to achieve a real end-to-end training from scratch for this new model! Makes me feel a bit uncomfortable.
> Makes me feel a bit uncomfortable.
Why? Do compilers which can't bootstrap themselves also make you uncomfortable due to dependencies on pre-built artifacts? I'm not saying you're unjustified to feel that way, but sometimes more abstracted systems are quicker to build and may have better performance than those built from the ground up. Selecting which one is better depends on your constraints and taste
Compilers are deterministic (for the most part, and it's incredibly rare to introduce a compiler bug that self-replicates in future compilers (unless you're Ken Thompson and are reflecting upon trust itself)).
Alternatively, AlphaFold 2's output is noisy, and using that to train AlphaFold 3, which presumably may be used to train what becomes AlphaFold 4, results in a cascade of errors.
Consider that humans also learn from other humans, and sometimes surpass their teachers.
A bit more comfortable?
Ahh, but the new young master is able to explain their work and processes to the satisfaction of the old masters. In the 'Science' of our modern times it's a requirement to show your work (yes, yes, I know about the replication crisis and all that terrible jazz).
Not being able to ascertain how and why the ML/AI is achieving results is not quite the same and more akin to the alchemists and sorcerers with their cyphers and hidden laboratories.
3 replies →
I am trying to understand how accurate the docking predictions are.
Looking at the PoseBusters paper [1] they mention, they say they are 50% more accurate than traditional methods.
DiffDock, which is the best DL based systems gets 30-70% depending on the dataset, and traditional gets 50-70%. The paper highlighted some issues with the DL-based methods and given that DeepMind would have had time to incorporate this into their work and develop with the PoseBusters paper in mind, I'd hope it's significantly better than 50-70%. They say 50% better than traditional so I expected something like 70-85% across all datasets.
I hope a paper will appear soon to illuminate these and other details.
[1] https://pubs.rsc.org/en/content/articlehtml/2024/sc/d3sc0418...
So much of the talk about their "free server" seems to be trying to distract from the fact that they are not releasing the model.
I feel like it's an important threshold moment if this gets accepted into scientific use without the model being available - reproducibility of results becomes dependent on the good graces of a single commercial entity. I kind of hope that like OpenAI it just spurs creation of equivalent open models that then actually get used.
This tool reminds me that the human body functions much like a black box. While physics can be modeled with equations and constraints, biology is inherently probabilistic and unpredictable. We verify the efficacy of a medicine by observing its outcomes: the medicine is the input, and the changes in symptoms are the output. However, we cannot model what happens in between, as we cannot definitively prove that the medicine affects only its intended targets. In many ways, much of what we understand about medicine is based on observing these black-box processes, and this tool helps to model that complexity.
Classic essay in this vein:
>Can a biologist fix a radio? — Or, what I learned while studying apoptosis
https://www.cell.com/cancer-cell/pdf/S1535-6108(02)00133-2.p...
>However, if the radio has tunable components, such as those found in my old radio (indicated by yellow arrows in Figure 2, inset) and in all live cells and organisms, the outcome will not be so promising. Indeed, the radio may not work because several components are not tuned properly, which is not reflected in their appearance or their connections. What is the probability that this radio will be fixed by our biologists? I might be overly pessimistic, but a textbook example of the monkey that can, in principle, type a Burns poem comes to mind. In other words, the radio will not play music unless that lucky chance meets a prepared mind.
I’d say it’s always been the case for medicine, when people first used medicines, the intention was never to fully understand what happens, just save a life, eliminate or reduce symptoms.
Now we’ve built explainable systems like computers and software, we try to overlay that onto everything and it might not work.
To quote Alan Watts, humans like to try square out wiggly systems because we’re not great and understanding wiggles.
> Thrilled to announce AlphaFold 3 which can predict the structures and interactions of nearly all of life’s molecules with state-of-the-art accuracy including proteins, DNA and RNA. [1]
There's a slight mismatch between the blog's title and Demis Hassabis' tweet, where he uses "nearly all".
The blog's title suggests that it's a 100% solved problem.
[1] https://twitter.com/demishassabis/status/1788229162563420560
First time reading a Deep Mind PR? This is literally their modus operandi.
How to make the share price go up…surprised?
Marketing vs. Reality :)
This is a basic question, but how is the accuracy of the predicted biomolecular interactions measured? Are the predicted interactions compared to known interactions? How would the accuracy of predicting unknown interactions be assessed?
Accuracy can be assessed two main ways: computationally and experimentally. Computationally, they would compare the predicted structures and interactions with known data from databases like PDB (Protein Database). Experimentally, they can use tools like x-ray crystallography and NMR (nuclear magnetic resonance) to obtain the actual molecule structure and compare it to the predicted result. The outcomes of each approach would be fed back into the model for refining future predictions.
https://www.rcsb.org/
AlphaFold very explicitly (unless something has changed) removes NMR structures as references because they are not accurate enough. I have a PhD in NMR biomolecular structure and I wouldn't trust. the structures for anything.
14 replies →
Can someone tell me what are the direct implication of this? I often see "helps with a drug design" but I'm too far from this industry and have never seen an example of such drugs
We're this much closer to being hacked?
Noob here. Can one make the following deduction:
In transformer based architectures, where one typically uses variation of attention mechanism to model interactions, even if one does not consider the autoregressive assumption of the domain's "nodes"(amino acids, words, image patches), if the number of final states that nodes take eventually can be permuted only in a finite way(i.e. they have sparse interactions between them), then these architectures are efficient way of modeling such domains.
In plain english the final state of words in a sentence and amino acids in a protein have only so many ways they can be arranged and transformers do a good job of modeling it.
Also can one assume this won't do well for domains where there is, say, sensitivity to initial conditions, like chaotic systems like wheather where the # final states just explodes?
If you work in this space would be interested to know what material impact has alphafold caused in your workflow since its release 4 years ago?
For a couple of years I've been expecting that ML models would be able to 'accelerate' bio-molecular simulations, using physics-based simulations as ground truth. But this seems to be a step beyond that.
When I competed in CASP 20 years ago (and lost terribly) I predicted that the next step to improve predictions would be to develop empirically fitted force fields to make MD produce accurate structure predictions (MD already uses empirically fitted force fields, but they are not great). This area was explored, there are now better force fields, but that didn't really push protein structure prediction forward.
Another approach is fully differentiable force fields- the idea that the force field function itself is a trainable structure (rather than just the parameters/weights/constants) that can be optimized directly towards a goal. Also explored, produced some interesting results, but nothing that woudl be considered transformative.
The field still generally believes that if you had a perfect force field and infinite computing time, you could directly recapitulate the trajectories of proteins folding (from fully unfolded to final state along with all the intermediates), but that doesn't address any practical problems, and is massively wasteful of resources compared to using ML models that exploit evolutionary information encoded in sequence and structures.
In retrospect I'm pretty relieved I was wrong, as the new methods are more effective with far fewer resources.
The article was heavy on the free research aspect, but light on the commercial application.
I'm curious about the business strategy. Does Google intend to license out tools, partner, or consult for commercial partners?
This version has Isomorphic Labs far more in the focus of the press release, which seems to be now the commercial arm more or less licensing access out.
The new AlphaFold server does not do everything the paper says AlphaFold 3 says it does. You cannot predict docking with the server! That is the main interest of pharma companies, 'does our medication bind to the target protein?'. From the FAQ: 'AlphaFold Server is a web-service that offers customized biomolecular structure prediction. It makes several newer AlphaFold3 capabilities available, including support for a wider range of molecule type' - that's not ALL AlphaFold3 capabilities. Isomorphic prints the money with those additional capabilities.
It's hilarious that Google says they don't allow this for safety reasons, pure OpenAI fluff. It's just money.
I wonder what the license for RoseTTAFold is. On github you have:
https://github.com/RosettaCommons/RoseTTAFold/blob/main/LICE...
But there's also:
https://files.ipd.uw.edu/pub/RoseTTAFold/Rosetta-DL_LICENSE....
Which is it?
[flagged]
as soon as google tries to think commercially this will shut down so the longer it stays pure research the better. google is bad with productization.
I don't think it was ever pure research. The article talks about infinity labs, which is the co. Mercial branch for drug discovery.
I do agree that Google seems bad at commercialization, which is why I'm curious on what the strategy is.
It is hard to see them being paid consultants or effective partners for pharma companies, let alone developing drugs themselves.
Here's something that bugs me about ML: all we have is prediction and no explanation how we come to that prediction, ie: no deeper understanding on the underlying principles.
So despite that we got a good match this time, how can we be sure that the match will be equally good next time? And how to use ML to predict structure that we have no baseline to start with or experimental result to benchmark ? In the absence of physics-like principles, How can we ever be sure that ML results next time is correct ?
There is a biannual structural prediction contest called CASP [1], in which a set of newly determined structures is used to benchmark the prediction methods. Some of these structures will be "novel", and so can be used to estimate the performance of current methods on predicting "structure that we have no baseline to start with".
CASP-style assessments are something that should done for more research fields, but it's really hard to persuade funders and researchers to put up the money and embargo the data as required.
[1] https://en.wikipedia.org/wiki/CASP
Speaking of physics, we should borrow the quote "Shut up and calculate" to describe the situation: it works so use it now and worry about the explanations later.
Except the model is not open-source. You can't calculate anything.
I'm interested in how they measure accuracy of binding site identification and binding pose prediction. This was missing for the hitherto widely-used binding pose prediction tool Autodock Vina (and in silico binding pose tools in general). Despite the time I invested in learning & exercising that tool, I avoided using it for published research because I could not credibly cite its general-use accuracy. Is / will Alphafold 3 be citeable in the sense of "I have run Alphafold on this particular target of interest and this array of ligands, and have found these poses of X kJ/mol binding energy, and this is known to an accuracy of Y% because of Alphafold 3's training set results cited below'
I've never trusted those predicted binding energies. If you have predicted a ligand/protein complex and have high confidence in it and want to study the binding energy I really think you should do a full MD simulation, you can pull the ligand-protein complex apart and measure the change in free energy explicitly.
Also, and this is an unfounded guess only, the problem of protein / ligand docking is quite a bit more complex than protein folding - there seems to be a finite set of overall folds used in nature, while docking a small ligand to a big protein with flexible sidechains and even flexible large-scale structures can have induced fits that are really important to know and estimate, and I'm just very sceptical that it's going to be possible to in a general fashion ever predict these accurately by the AI model with the limited training data.
Though you just need some hints, then you can run MD sims on them to see what happens for real.
Would anyone more familiar with the field be able to provide some cursory resources on the protein folding problem? I have a background in computer science and a half a background in biology (took two semesters of OChem, biology, anatomy; didn’t go much further).
A video summary of why this research is important: https://youtu.be/Mz7Qp73lj9o?si=29vjdQtTtIOk_0CV
Thanks for this informative video summary. As a layperson, with a BS in Chemistry, it was quite helpful in understanding main bulletpoints of this accomplishment.
Very cool that anyone can login to https://golgi.sandbox.google.com/ and check it out
I wonder in the not too distant future if these AI predictions could be explained back into “humanized” understanding. Much like ChatGPT can simplify complex topics … cold the model in the future provide feedback to researchers why it is making this prediction?
Would be very useful if one they used it to predict the structure and interaction of the known variants to.
Would be very helpful when predicting if a mutation on a protein would lead to loss of function for the protein.
Does it also simulate prion (=misfolded structures) based diseases?
The improvement on predicting protein/RNA/ligand interactions might facilitate many commercially relevant use cases. I assume pharma and biotech will eagerly get in line to use this.
A lot of accelerated article previews as of recently. Seems like humanity is making a lot of breakthroughs.
This is nothing short of amazing for all those suffering from disease.
Excited but also it's been a fair bit now and I have yet to see something truly remarkable come out of this
Zero-shot nearly beating trained catboost is pretty amazing.
How can I get my outputs in PDB format? Can not?
And google is giving the service for free. Pretty good.
Google's Game of Life 3D: Spiral edition
From: https://www.nature.com/articles/d41586-024-01383-z
>Unlike RoseTTAFold and AlphaFold2, scientists will not be able to run their own version of AlphaFold3, nor will the code underlying AlphaFold3 or other information obtained after training the model be made public. Instead, researchers will have access to an ‘AlphaFold3 server’, on which they can input their protein sequence of choice, alongside a selection of accessory molecules. [. . .] Scientists are currently restricted to 10 predictions per day, and it is not possible to obtain structures of proteins bound to possible drugs.
This is unfortunate. I wonder how long until David Baker's lab upgrades RoseTTAFold to catch up.
That sucks a bit. I was just wondering why they are touting that 3rd party company in their own blog post, who commercialise research tools, as well. Maybe there are some corporate agreements with them that prevents them from opening the system...
Imagine the goodwill for humanity for releasing these pure research systems for free. I just have a hard time understanding how you can motivate to keep it closed. Let's hope it will be replicated by someone who doesn't have to hide behind the "responsible AI" curtain as it seems they are now.
Are they really thinking that someone who needs to predict 11 structures per day are more likely to be a nefarious evil protein guy than someone who predicts 10 structures a day? Was AlphaFold-2 (that was open-sourced) used by evil researchers?
- "Imagine the goodwill for humanity for releasing these pure research systems for free."
The entire point[0] is that they want to sell an API to drug-developer labs, at exclusive-monopoly pricing. Those labs in turn discover life-saving drugs, and recoup their costs from e.g. parents of otherwise-terminally-ill children—again, priced as an exclusive monopoly.
[0] As signaled by "it is not possible to obtain structures of proteins bound to possible drugs"
It's a massive windfall for Alphabet, and it'd be a profound breach of their fiduciary duties as a public company to do anything other than lock-down and hoard this API, and squeeze it for every last billion.
This is a deeply, deeply, deeply broken situation.
10 replies →
Isomorphic Labs? That's an Alphabet owned startup run by Denis Hassabis that they created to commercialise the Alphafold work, so it's not really a 3rd party at all.
There is at least some difference between a monitored server and a privately ran one, if negative consequences are possible
Also no commercial use, from the paper:
> AlphaFold 3 will be available as a non-commercial usage only server at https://www.alphafoldserver.com, with restrictions on allowed ligands and covalent modifications. Pseudocode describing the algorithms is available in the Supplementary Information. Code is not provided.
How easy/hard would be for the scientific community to come up with an "OpenFold" model which is pretty much AF3 but fully open source and without restrictions in it?
I can image training will be expensive, but I don't think it will be at a GPT-4 level of expensive.
2 replies →
If you need to submit to their server, I don't know who would use it for commercial reasons anyway. Most biotech startups and pharma companies are very careful about entering sequences into online tools like this.
What's the point in that that - I mean who does non-commercial drug research?
2 replies →
Yes, because that's going to stop competitors.. it's why they didn't release code I guess.
This is yet another large part of a biotech related Gutenberg moment.
3 replies →
Or OpenFold, which is the more literal reproduction of AlphaFold 2: https://github.com/aqlaboratory/openfold
Time for an OpenFold3? Or would it be an OpenFold2?
The AI call is rolling fast, I see similarities with cryptography in the 90s.
I have a history to tell for the record, back in the 90s we developed a home banking for Palm (with a modem), it was impossible to perform RSA because of the speed so I contacted the CEO of Certicom which was the unique elliptic curve cryptography implementation at that time. Fast forward and ECC is everywhere.
Not just unfortunate, but doesn't this make it completely untrustable? How can you be sure the data was not modified in any way? How can you verify any results?
You determine a crystal structure of a known protein which does not previously have a known structure, and compare the prediction to the experimentally determined structure.
There is a biennial (biannual?) competition known as CASP where some new structures, not yet published, are used for testing predictions from a wide range of protein structure prediction (so, basically blind predictions which are then compared when the competition wraps up). AlphaFold beat all the competitors by a very wide margin (much larger than the regular rate of improvement in the competition), and within a couple years, the leading academic groups adopted the same techniques and caught up.
It was one of the most important and satisfying moments in structure prediction in the past two+ decades. The community was a bit skeptical but as it's been repeatedly tested, validated, and reproduced, people are generally of the opinion that DeepMind "solved" protein structure prediction (with some notable exceptions), and did so without having the solve the full "protein folding problem" (which is actually great news while also being somewhat depressing).
1 reply →
in other words, this has been converted to a novelty, and has no use for scientific purposes.
No. It just means that scientific purposes will have an additional tax paid to google. This will likely reduce use in academia but won't deter pharmaceutical companies.
The second amendment prevents the government's overreaching perversion to restrict me from having the ability to print biological weapons from the comfort of my couch.
Google has no such restriction.
I know this is tongue in cheek, but you absolutely can be restricted from having a biological weapons factory in your basement (similar to not being able to pick "nuclear bombs" as your arms to bear).
2 replies →
Sergey once said "We don't have an army per-se" (he was referring the size of Google's physical security group) at TGIF.
There was a nervous chuckle from the audience.
/s is strong with this one
The logical consequence is to put all scientific publications under a license that restricts the right to train commercial ai models on them.
Science advances because of an open exchange of ideas, the original idea of patents was to grant the inventor exclusive use in exchange for disclosure of knowledge.
Those who did not patent, had to accept that their inventions would be studied and reverse engineered.
The „as a service“ model, breaks that approach.
This turns it into a tool that deserves to be dethroned by another group, frankly. What a strange choice.
Well, it's because you can design deadly viruses using this technology. Viruses gain entry to living cells via cell-surface receptor proteins whose normal job is to bind signalling molecules, alter their conformation and translate that external signal into the cellular interior where it triggers various responses from genomic transcription to release of other signal molecules. Viruses hijack such mechanisms to gain entry to cells.
Thus if you can design a viral coat protein to bind to a human cell-surface receptor, such that it gets translocated into the cell, then it doesn't matter so much where that virus came from. The cell's firewall against viruses is the cell membrane, and once inside, the biomolecular replication machinery is very similar from species to species, particularly within restricted domains, such as all mammals.
Thus viruses from rats, mice, bats... aren't going to have major problems replicating in their new host - a host they only gained access to because some nation-state actors working in collaboration on such gain-of-function research in at least two labs on opposite sides of the world with funds and material provided by the two largest economic powers for reasons that are still rather opaque, though suspiciously banal...
Now while you don't need something like AlphaFold3 to do recklessly stupid things (you could use directed evolution, making millions of mutatad proteins, throwing them at a wall of human cell receptors and collecting what stuck), it makes it far easier. Thus Google doesn't want to be seen as enabling, though given their prediliction for classified military-industrial contracting to a variety of nation-states, particularly with AI, with revenue now far more important than silly "don't be evil" statements, they might bear watching.
On the positive side, AlphaFold3 will be great for fields like small molecular biocatalysis, i.e. industrial applications in which protein enzymes (or more robust heterogenous catalysts designed based on protein structures) convert N2 to ammonia, methane to methanol, or selectively bind CO2 for carbon capture, modification of simple sugars and amino acids, etc.
Stepping back, the high-order bit here is an ML method is beating physically-based methods for accurately predicting the world.
What happens when the best methods for computational fluid dynamics, molecular dynamics, nuclear physics are all uninterpretable ML models? Does this decouple progress from our current understanding of the scientific process - moving to better and better models of the world without human-interpretable theories and mathematical models / explanations? Is that even iteratively sustainable in the way that scientific progress has proven to be?
Interesting times ahead.
If you're a scientist who works in protein folding (or one of those other areas) and strongly believe that science's goal is to produce falsifiable hypotheses, these new approaches will be extremely depressing, especially if you aren't proficient enough with ML to reproduce this work in your own hands.
If you're a scientist who accepts that probabilist models beat interpretable ones (articulated well here: https://norvig.com/chomsky.html), then you'll be quite happy because this is yet another validation of the value of statistical approaches in moving our ability to predict the universe forward.
If you're the sort of person who believes that human brains are capable of understanding the "why" of how things work in all its true detail, you'll find this an interesting challenge- can we actually interpret these models, or are human brains too feeble to understand complex systems without sophisticated models?
If you're the sort of person who likes simple models with as few parameters as possible, you're probably excited because developing more comprehensible or interpretable models that have equivalent predictive ability is a very attractive research subject.
(FWIW, I'm in the camp of "we should simultaneously seek simpler, more interpretable models, while also seeking to improve native human intelligence using computational augmentation")
The goal of science has always been to discover underlying principles and not merely to predict the outcome of experiments. I don't see any way to classify an opaque ML model as a scientific artifact since by definition it can't reveal the underlying principles. Maybe one could claim the ML model itself is the scientist and everyone else is just feeding it data. I doubt human scientists would be comfortable with that, but if they aren't trying to explain anything, what are they even doing?
55 replies →
> If you're the sort of person who believes that human brains are capable of understanding the "why" of how things work in all its true detail, you'll find this an interesting challenge- can we actually interpret these models, or are human brains too feeble to understand complex systems without sophisticated models?
I think chess engines, weirdly enough, have disabused me of this notion.
There are lots of factors a human considers when looking at a board. Piece activity. Bishop and knight imbalances. King safety. Open and semi-open file control. Tempo. And on and on.
But all of them are just convenient shortcuts that allow us to substitute reasonable guesses for what really matters: exhaustively calculating a winning line through to the end. “Positional play” is a model that only matters when you can’t calculate trillions of lines thirty moves deep, and it’s infinitely more important that a move survives your opponent’s best possible responses than it is to satisfy some cohesive higher level principle.
10 replies →
I'm in the following camp: It is wrong to think about the world or the models as "complex systems" that may or may not be understood by human intelligence. There is no meaning beyond that which is created by humans. There is no 'truth' that we can grasp in parts but not entirely. Being unable to understand these complex systems means that we have framed them in such a way (f.e. millions of matrix operations) that does not allow for our symbol-based, causal reasoning mode. That is on us, not our capabilities or the universe.
All our theories are built on observation, so these empirical models yielding such useful results is a great thing - it satisfies the need for observing and acting. Missing explainability of the models merely means we have less ability to act more precisely - but it does not devalue our ability to act coarsely.
21 replies →
> ... and strongly believe that science's goal is to produce falsifiable hypotheses, these new approaches will be extremely depressing
I don't quite understand this point — could you elaborate?
My understanding is that the ML model produces a hypothesis, which can then be tested via normal scientific method (perform experiment, observe results).
If we have a magic oracle that says "try this, it will work", and then we try it, and it works, we still got something falsifiable out of it.
Or is your point that we won't necessarily have a coherent/elegant explanation for why it works?
7 replies →
What if our understanding of the laws of the natural sciences are subtly flawed and AI just corrects perfectly for our flawed understanding without telling us what the error in our theory was?
Forget trying to understand dark matter. Just use this model to correct for how the universe works. What is actually wrong with our current model and if dark matter exists or not or something else is causing things doesn't matter. "Shut up and calculate" becomes "Shut up and do inference."
6 replies →
There have been times in the past when usable technology surpassed our scientific understanding, and instead of being depressing it provided a map for scientific exploration. For example, the steam engine was developed by engineers in the 1600s/1700s (Savery, Newcomen, and others) but thermodynamics wasn’t developed by scientists until the 1800s (Carnot, Rankine, and others).
2 replies →
What always struck me about Chomskyists is that they chose a notion of interpretable model that required unrealistic amounts of working interpretation. So Chomsky grammars have significant polynomial memory and computational costs for grammars as they approach something resembling human grammar. And you say, ok, the human brain can handle much more computation than that, and that's fine. But (for example) context-free grammars aren't just O(n^3) in computational cost; for a realistic description of human language they're O(n^3) in human-interpretable rules.
Other Chomsky-like models of human grammars have different asymptotic behavior and different choices of n, but the same fundamental problem; the big-O constant factor isn't neurons firing but rather human connections between the n inputs. How can you conceive of human minds being able to track O(n^3) (or whatever) cost where that n is everything being communicated -- words, concepts, symbols, representations, all that jazz and the polynomial relationships between them?
But I feel an apology is in order: I've had quite a few beers before coming home, and it's probably a mistake to try to express academically charged and difficult views on the Internet while in an inebriated state. Probably the alcohol has substantially decreased my mental computational power. However, it has only mildly impaired my ability to string together words and sentences in a grammatically complex fashion. In fact, I often feel that the more sober and clear-minded I am, the simpler my language is. Maybe human grammar is actually sub-polynomial. I have observed the same in ChatGPT; the more flowery and wordy it has become over time, the dumber its output.
1 reply →
For some, this conversation started when the machine derived four colour map proof was announced which is almost 5 decades ago in 1976
> If you're the sort of person who believes that human brains are capable of understanding the "why" of how things work in all its true detail
This seems to me an empirical question about the world. It’s clear our minds are limited, and we understand complex phenomena through abstraction. So either we discover we can continue converting advanced models to simpler abstractions we can understand, or that’s impossible. Either way, it’s something we’ll find out and will have to live with in the coming decades. If it turns out further abstractions aren’t possible, well, enlightenment thought had lasted long enough. It’s exciting to live at a time in humanity’s history when we enter a totally uncharted new paradigm.
> can we actually interpret these models, or are human brains too feeble to understand complex systems without sophisticated models?
I think we will have to develop a methodology and supporting toolset to be able to derive the underlying patterns driving such ML models. It's just too much for a human to comb through by themselves and make sense of.
So the work to simplify ML models, reduce dimensions, etc. becomes the numeric way to seek simple actual scientific models. Scientific computing and science become one.
Do you think a model will also be able to truly comprehend everything too ?
The goal of science should always be to seek good explanations hard to vary.
The frontier in model space is kind of fluid. It's all about solving differential equations.
In theoretical physics, you know the equations, you solve equations analytically, but you can only do that when the model is simple.
In numerical physics, you know the equations, you discretize the problem on a grid, and you solve the constraint defined by the equations with various numerical integration schemes like RK4, but you can only do that when the model is small and you know the equations, and you find a single solution.
Then you want the result faster, so you use mesh-free methods and adaptive grids. It works on bigger models but you have to know the equations, finding a single solution to the differential equations.
Then you compress this adaptive grid with a neural network, while still knowing the governing equations, and you have things like Physics Informed Neural Networks ( https://arxiv.org/pdf/1711.10561 and following papers) where you can bound the approximation error. This method allows solve all solutions to the differential equations simultaneously, sharing the computations.
Then when knowing explicitly your governing equations is too complex, so you assume that there are some governing stochastic equations implicitly, which you learn the end-result of the dynamic with a diffusion model, that's what this alpha-fold is doing.
ML is kind of a memoization technique, analog to hashlife in the game of life, that allows you reuse your past computational efforts. You are free to choose on this ladder which memory-compute trade-off you want to use to model the world.
As a steelman, wouldn't the abundance of infinitely generate-able situations make it _easier_ for us to develop strong theories and models? The bottleneck has always been data. You have to do expensive work in the real world and accurately measure it before you can start fitting lines to it. If we were to birth an e.g. atomically accurate ML model of quantum physics, I bet it wouldn't take long until we have mathematical theories that explain why it works. Our current problem is that this stuff is super hard to manipulate and measure.
Maybe; AI chess engines have improved human understanding of the game very rapidly, even though humans cannot beat engines.
1 reply →
This is an important aspect that's being ignored IMO.
For a lot of problems, currently you either don't have an an analytical solution and the alternative is a brute force-ish numerical approach. As a result the computational cost of simulating things enough times to be able to detect behavior that can inform theories/models (potentially yielding a good analytical result) is not viable.
In this regard, ML models are promising.
It depends whether the value of science is human understanding or pure prediction. In some realms (for drug discovery, and other situations where we just need an answer and know what works and what doesn’t), pure prediction is all we really need. But if we could build an uninterpretable machine learning model that beats any hand-built traditional ‘physics’ model, would it really be physics?
Maybe there’ll be an intermediate era for a while where ML models outperform traditional analytical science, but then eventually we’ll still be able to find the (hopefully limited in number) principles from which it can all be derived. I don’t think we’ll ever find that Occam’s razor is no use to us.
> But if we could build an uninterpretable machine learning model that beats any hand-built traditional ‘physics’ model, would it really be physics?
At that point I wonder if it would be possible to feed that uninterpretable model back into another model that makes sense of it all and outputs sets of equations that humans could understand.
The success of these ML models has me wondering if this is what Quantum Mechanics is. QM is notoriously difficult to interpret yet makes amazing predictions. Maybe wave functions are just really good at predicting system behavior but don't reflect the underlying way things work.
OTOH, Newtonian mechanics is great at predicting things under certain circumstances yet, in the same way, doesn't necessarily reflect the underlying mechanism of the system.
So maybe philosophers will eventually tell us the distinction we are trying to draw, although intuitive, isn't real
1 reply →
Pure prediction is only all we need if the total end-to-end process is predicted correctly - otherwise there could be pretty nasty traps (e.g., drug works perfectly for the target disease but does something unexpected elsewhere etc.).
1 reply →
In case it's not clear, this does not "beat" experimental structure determination. The matches to experiment are pretty close, but they will be closer in some cases than others and may or may not be close enough to answer a given question about the biochemistry. It certainly doesn't give much information about the dynamics or chemical perturbations that might be relevant in biological context. That's not to pooh-pooh alphafold's utility, just that it's a long way from making experimental structure determination unnecessary, and much much further away from replacing a carefully chosen scientific question and careful experimental design.
It means we now have an accurate surrogate model or "digital twin" that can be experimented on almost instantaneously. So we can massively accelerate the traditional process of developing mechanistic understanding through experiment, while also immediately be able to benefit from the ability to make accurate predictions, even without needing understanding.
In reality, science has already pretty much gone this way long ago, even if people don't like to admit it. Simple, reductionist explanations for complex phenomena in living systems don't really exist. Virtually all of medicine nowadays is empirical: try something, and if you can prove its safe and effective, you keep doing it. We almost never have a meaningful explanation for how it really works, and when we think we do, it gets proven wrong repeatedly, while the treatment keeps working as always.
Medicine can be explained fairly simply, and the why of how it works as it does is also explained by this:
Imagine a very large room that has every surface covered by on-off switches.
We cannot see inside of this room. We cannot see the switches. We cannot fit inside of this room, but a toddler fits through the tiny opening leading into the room. The toddler cannot reach the switches, so we equip the toddler with a pole that can flip the switches. We train the toddler, as much as possible, to flip a switch using the pole.
Then, we send the toddler into the room and ask the toddler to flip the switch or switches we desire to be flipped, and then do tests on the wires coming out of the room to see if the switches were flipped correctly. We also devise some tests for other wires to see if that naughty toddler flipped other switches on or off.
We cannot see inside the room. We cannot monitor the toddler. We can't know what _exactly_ the toddler did inside the room.
That room is the human body. The toddler with a pole is a medication.
We can't see or know enough to determine what was activated or deactivated. We can invent tests to narrow the scope of what was done, but the tests can never be 100% accurate because we can't test for every effect possible.
We introduce chemicals then we hope-&-pray that the chemicals only turned on or off the things we wanted turned on or off. Craft some qualifications testing for proofs, and do a 'long-term' study to determine if there were other things turned on or off, or a short circuit occurred, or we broke something.
I sincerely hope that even without human understanding, our AI models can determine what switches are present, which ones are on and off, and how best to go about selecting for the correct result.
Right now, modern medicine is almost a complete crap-shoot. Hopefully modern AI utilities can remedy the gambling aspect of medicine discovery and use.
2 replies →
instead of "in mice", we'll be able to say "in the cloud"
4 replies →
It makes me think about how Einstein was famous for making falsifiable real-world predictions to accompany his theoretical work. And, sometimes it took years for proper experiments to be run (such as measuring a solar eclipse during the breakout of a world war).
Perhaps the opportunity here is to provide a quicker feedback loop for theory about predictions in the real world. Almost like unit tests.
> Perhaps the opportunity here is to provide a quicker feedback loop for theory about predictions in the real world. Almost like unit tests.
Or jumping the gap entirely to move towards more self-driven reinforcement learning.
Could one structure the training setup to be able to design its own experiments, make predictions, collect data, compare results, and adjust weights...? If that loop could be closed, then it feels like that would be a very powerful jump indeed.
In the area of LLMs, the SPAG paper from last week was very interesting on this topic, and I'm very interested in seeing how this can be expanded to other areas:
https://github.com/Linear95/SPAG
Agreed. At the very least, models of this nature let us iterate/filter our theories a little bit more quickly.
1 reply →
Many of our existing physical models can be decomposed into "high-confidence, well tested bit" plus "hand-wavy empirically fitted bit". I'd like to see progress via ML replacing the empirical part - the real scientific advancement then becomes steadily reducing that contribution to the whole by improving the robust physical model incrementally. Computational performance is another big influence though. Replacing the whole of a simulation with an ML model might still make sense if the model training is transferrable and we can take advantage of the GPU speed-ups, which might not be so easy to apply to the foundational physical model solution. Whether your model needs to be verified against real physical models depends on the seriousness of your use-case; for nuclear weapons and aerospace weather forecasts I imagine it will remain essential, while for a lot of consumer-facing things the ML will be good enough.
Physics-informed machine learning is a whole (nascent) subfield that is very much in line with this thinking. Steve Brunton has some good stuff about this on YouTube.
"Best methods" is doing a lot of heavy lifting here. "Best" is a very multidimensional thing, with different priorities leading to different "bests." Someone will inevitably prioritize reliability/accuracy/fidelity/interpretability, and that's probably going to be a significant segment of the sciences. Maybe it's like how engineers just need an approximation that's predictive enough to build with, but scientists still want to understand the underlying phenomena. There will be an analogy to how some people just want an opaque model that works on a restricted domain for their purposes, but others will be interested in clearer models or unrestricted/less restricted domain models.
It could lead to a very interesting ecosystem of roles.
Even if you just limit the discussion to using the best model of X to design a better Y, limited to the model's domain of validity, that might translate the usage problem to finding argmax_X of valueFunction of modelPrediction of design of X. In some sense a good predictive model is enough to solve this with brute force, but this still leaves room for tons of fascinating foundational work. Maybe you start to find that the (wow so small) errors in modelPrediction are correlated with valueFunction, so the most accurate predictions don't make it the best for argmax (aka optimization might exploit model errors rather than optimizing the real thing). Or maybe brute force just isn't computationally feasible, so you need to understand something deeper about the problem to simplify the optimization to make it cheap.
Physicists like to retroactively believe that our understanding of physical phenomena preceded the implementation of uses of those phenomena, when the reality is that physics has always come in to clean up after the engineers. There are some rare exceptions, but usually the reason that scientific progress can be made in an area is that the equipment to perform experiments has been commoditized sufficiently by engineering demand for it.
We had semiconductors and superconductors before we understood how they worked -- on both cases arguably we still don't completely understand the phenomena. Things like the dynamo and the electric motor were invented by practice and later explained by scientists, not derived from first principles. Steam engines and pumps were invented before we had the physics to describe how they worked.
The most moneyed and well-coordinated organizations have honed a large hammer, and they are going to use it for everything, and so almost certainly future big findings in the areas you mention, probabilistically inclined models coming from ML will be the new gold standard.
But yet the only thing that can save us from ML will be ML itself because it is ML that has the best chance to be able to extrapolate patterns from these blackbox models to develop human interpretable models. I hope we do dedicate explicit effort to this endeavor, and so continue the human advances and expanse of human knowledge in tandem with human ingenuity with computers at our assistance.
Spoiler: "Interpretable ML" will optimize for output that either looks plausible to humans, reinforces our preconceptions, or appeals to our aesthetic instincts. It will not converge with reality.
3 replies →
Interpretable AI is the same scam as alignment.
It's interesting to compare this situation to earlier eras in science. Newton, for example, gave us equations that were very accurate but left us with no understanding at all of why they were accurate.
It seems like we're repeating that here, albeit with wildly different methods. We're getting better models but by giving up on the possibility of actually understanding things from first principles.
Not comparable. Our current knowledge of the physics involved in these systems is complete. It is just impossibly difficult to calculate from first principles.
A new-ish field of "mechanistic interpretability" is trying to poke at weights and activations and find human-interpretable ideas w/in them. Making lots of progress lately, and there are some folks trying to apply ideas from the field to Alphafold 2. There are hopes of learning the ideas about biology/molecular interactions that the model has "discovered".
Perhaps we're in an early stage of Ted Chiang's story "The Evolution of Human Science", where AIs have largely taken over scientific research and a field of "meta-science" developed where humans translate AI research into more human-interpretable artifacts.
A few things:
1. Research can then focus on where things go wrong
2. ML models, despite being "black boxes," can still have brute-force assessment performed of the parameter space over covered and uncovered areas by input information
3. We tend to assume parsimony (i.e Occam's razor) to give preference to simpler models when all else is equal. More complex black-box models exceeding in prediction let us know the actual causal pathway may be more complex than simple models allow. This is okay too. We'll get it figured out. Not everything is closed-form, especially considering quantum effects may cause statistical/expected outcomes instead of deterministic outcomes.
Interesting times indeed. I think the early history of medicines takes away from your observation though. In the 19th and early 20th century people didn't know why medicines worked, they just did. The whole "try a bunch of things on mice, pick the best ones and try them on pigs, and then the best of those and try a few on people" kind of thing. In many ways the mice were a stand in for these models, at the time scientists didn't understand nearly as much about how mice worked (early mice models were pretty crude by today's standards) but they knew they were a close enough analog to the "real thing" that the information provided by mouse studies was usefully translated into things that might help/harm humans.
So when you're tools can produce outputs that you find useful, you can then use those tools to develop your understanding and insights. As a tool, this is quite good.
I asked a friend of mine who is chemistry professor at a large research university something along these lines a while ago. He said that so far these models don't work well in regions where either theory or data is scarce, which is where most progress happens. So he felt that until they can start making progress in those areas it won't change things much.
Major breakthroughs happen when clear connections can be made and engineered between the many bits of solved but obscured solutions.
> What happens when the best methods for computational fluid dynamics, molecular dynamics, nuclear physics are all uninterpretable ML models?
A better analogy is "weather forecasting".
interesting choice considering the role chaos theory plays in forever rendering long term weather predictions impossible, by humans or LLMs.
This is the topic of epistemology of the sciences in books such as "New Direction in the Philosophy of Mathematics" [1] and happened before with problems such as the four color theorem [2] where AI was not involved.
Going back to the uninterpretable ML models in the context of AlphaFold 3, I think one method for trying to explain the findings is similar to the experimental methods of physics with reality: you perform experiments with the reality (in this case AlphaFold 3) to came up with sound conclusions. AI/ML is an interesting black-box system.
There are other open discussions on this topic. For example, can our human brain absorbe that knowledge or it is limited somehow with the scientific language that we have now?
[1] https://www.google.com.ar/books/edition/New_Directions_in_th...
[2] https://en.wikipedia.org/wiki/Four_color_theorem
In physics, we already deal with the fact that many of the core equations cannot be analytically solved for more than the most basic scenarios. We've had to adapt to using approximation methods and numerical methods. This will have to be another place where we adapt to a practical way of getting results.
Reminds me of the novel Blindsight - in it there's special individuals who work as synthesists, whos job it is to observe and understand and then somehow translate back to "lay person" the seemingly undecipherable actions/decisions of advanced computers and augmented humans.
I'd say it's not new. Take fluid dynamics as an example, the navier stokes equations predict the motion of fluids very well but you need to approximately solve them on a computer in order to get useful predictions for most setups. I guess the difference is the equation is compact and the derivation from continuum mechanics is easy enough to follow. People still rely on heuristics to answer "how does a wing produce lift?". These heuristic models are completely useless at "how much lift will this particular wing produce under these conditions?". Seems like the same kind of situation. Maybe progress forward will look like producing compact models or tooling to reason about why a particular thing happened.
I think it likely that instead of replacing existing methods, we will see a fusion. Or rather, many different kinds of fusions - depending on the exact needs of the problems at hand (or in science, the current boundary of knowledge). If nothing else then to provide appropriate/desirable level of explainability, correctness etc. Hypothetically the combination will also have better predictive performance and be more data efficient - but it remains to be seen how well this plays out in practice. The field of "physics informed machine learning" is all about this.
Is alphafold doing model generation or is it just reducing a massive state space?
The current computational and systems biochemistry approaches struggle to model large biomolecules and their interactions due to the large degrees of freedom of the models.
I think it is reasonable to rely on statistical methods to lead researchers down paths that have a high likelihood of being correct versus brute forcing the chemical kinetics.
After all chemistry is inherently stochastic…
Our metaphors and intuitions were crumbling already and stagnating. See quantum physics: sometimes a particle, sometimes a wave, and what constitute a measurement anyway?
I’ll take prediction over understanding if that’s the best our brains can do. We’ve evolved to deal with a few orders of magnitude around a meter and a second. Maybe dealing with light-years and femtometer/seconds is too much to ask.
> Does this decouple progress from our current understanding of the scientific process - moving to better and better models of the world without human-interpretable theories and mathematical models / explanations?
Replace "human-interpretable theories" with "every man interpretable theories", and you'll have a pretty good idea of how > 90% of the world feels about modern science. It is indistinguishable from magic, by the common measure.
Obtuse example: My parents were alive when the first nuclear weapon was detonated. They didn't know that they didn't know this weapon was being built, let alone that it might have ignited the atmosphere.
With sophisticated enough ML, that 90% will become 99.9% - save the few who have access to (and can trust) ML tools that can decipher the "logic" from the original ML tools.
Yes, interesting times ahead... indeed.
"better and better models of the world" does not always mean "more accurate" and never has.
We already know how to model the vast majority of things, just not at a speed and cost which makes it worthwhile. There are dimensions of value - one is accuracy, another speed, another cost, and in different domains additional dimensions. There are all kinds of models used in different disciplines which are empirical and not completely understood. Reducing things to the lowest level of physics and building up models from there has never been the only approach. Biology, geology, weather, materials all have models which have hacks in them, known simplifications, statistical approximations, so the result can be calculated. It's just about choosing the best hacks to get the best trade off of time/money/accuracy.
This is a key but secondary concern to many of us working in molecular geneticist who will use AlphaFold 3 to evaluate pair-wise interactions. We often have genetic support for an interaction between proteins A and B. For example, in a study of genetic variation in responses of mice to morphine I currently have two candidate proteins that interact epistatically, suggesting a possible “lock and key” model—-the mu opiate receptor (MOR) and FGF12. I can now evaluate the likelihood of a direct molecular interaction between these proteins and possible amino acids substitutions that account for individuals difference.
In other words I bring a hypothesis to AF3 and ask for it to refute or affirm.
For me the big question is how do we confidently validate the output of this/these model(s).
It's the right question to ask, and the answer is that we will still have to confirm them by experimental structure determination.
You are conflating the whole scientific endeavor to a very specific problem to which this specific approach is effective at producing results that fit with the observable world. This has nothing to do with science as a whole.
My argument is: weather.
I think it is fine & better for society to have applications and models for things we don't fully understand... We can model lots of small aspects of weather, and we have a lot of factors nailed down, but not necessarily all the interactions.. and not all of the factors. (Additional example for the same reason: Gravity)
Used responsibly. Of course. I wouldn't think an AI model designing an airplane that no engineers understand how it works is a good idea :-)
And presumably all of this is followed by people trying to understand the results (expanding potential research areas)
It would be cool to see an airplane made using generative design.
1 reply →
We need to advance mechanistic interpretability (field reverse engineering neural networks) https://www.youtube.com/watch?v=P7sjVMtb5Sg https://www.youtube.com/watch?v=7t9umZ1tFso https://www.youtube.com/watch?v=2Rdp9GvcYOE
To paraphrase Kahan, it's not interesting to me whether a method is accurate enough or not, but whether you can predict how accurate you can be. So, if ML methods can predict that they're right 98% of times then we can build this in our systems, even if we don't understand how they work.
Deterministic methods can predict result with a single run, ML methods will need ensemble of results to show the same confidence. It is possible at the end of day that the difference in cost might not he that high over time.
Science has always given us better, but error prone tooling to see further and make better guesses. There is still a scientific test. In a clinical trial, is this new drug safe and effective.
Perhaps an ai can be made to produce the work as well as a final answer, even if it has to reconstruct or invent the work backwards rather than explain it's own internal inscrutable process.
"produce a process that arrives at this result" should be just another answer it can spit out. We don't necessarily care if the answer it produces is actually the same as what originally happened inside itself. All we need is that the answer checks out when we try it.
No, science doesn't work that way. You can just calculate your way to scientific discoveries, you got to test them in the real world. Learning, both in humans and AI, is based on the signals provided by the environment. There are plenty of things not written anywhere, so the models can't simply train on human text to discover new things. They learn directly from the environment to do that, like AlphaZero did when it beat humans at Go.
In order for that not to happen (uninterpretable ML models) some research on symbolic distillation, aka symbolic regression
https://arxiv.org/abs/2006.11287
https://www.science.org/doi/10.1126/sciadv.aay2631
I think at some point, we will be able to produce models that are able to pass data into a target model and observe its activations and outputs and put together some interpretable pattern or loose set of rules that govern the input-output relationship in the target model. Using this on a model like AlphaFold might enable us to translate inferred chemical laws into natural language.
Even if we don’t understand the models themselves, you can still use them as a basis for understanding
For example, I have no idea how a computer works in every minute detail (ie, exactly the physics and chemistry of every process that happens in real time), but I have enough of an understanding of what to do with it, that I can use it as an incredibly useful tool for many things
Definitely interesting times!
Not the same. There is a difference between "I cannot understand the deeper details of certain model but some others can and there's the possibility of explaining it in detail" and "Nobody can understand it and there's not a clear cause-effect that we know" .
Except for weird cases, computers (or cars, or cameras, or lots of other man made devices) are clearly known and you (or another specialist) can clearly show why a device does X when you input Y on it.
> Does this decouple progress from our current understanding of the scientific process?
Thank God! As a person who uses my brain, I think I can say, pretty definitively, that people are bad at understanding things.
If this actually pans out, it means we will have harnessed knowledge/truth as a fundamental force, like fire or electricity. The "black box" as a building block.
This type of thing is called an "oracle".
We've had stuff like this for a long time.
Notable examples:
- Temple priestesses
- Tea-leaf reading
- Water scrying
- Palmistry
- Clairvoyance
- Feng shui
- Astrology
The only difference is, the ML model is really quite good at it.
1 reply →
I believe it simply tells us that our understanding of mechanical systems, especially chaotic ones, is not as well defined as we thought.
https://journals.aps.org/prresearch/abstract/10.1103/PhysRev...
> What happens when...
I can only assume that existing methods would still be used for verification. At least we understand the logic used behind these methods. The ML models might become more accurate on average but they could still throw out results that are way off occasionally, so their error rate would have to become equal to the existing methods.
I wonder if ML can someday be employed in deciphering such black box problems; a second model that can look under the hood at all the number crunching performed by the predictive model, identify the pattern that resulted in a prediction, and present it in a way we can understand.
That said, I don’t even know if ML is good at finding patterns in data.
> That said, I don’t even know if ML is good at finding patterns in data.
That's the only thing ML does.
The models are learning an encoding based on evolutionary related and known structures. We should be able to derive fundamental properties from those encodings eventually. Or at least our biophysical programmed models should map into that encoding. That might be a reasonable approach to look at the folding energy landscape.
Perhaps related, the first computer-assisted mathematics proof: https://en.wikipedia.org/wiki/Four_color_theorem
I'm sure that similar arguments for and against the proof apply here as well.
In terms of docking, you can call the conventional approaches "physically-based", however, they are rather poor physical models. Namely, they lack proper electrostatics, and, most importantly, basically ignore entropic contributions. There is no reason for concern.
I can only hope the models will be sophisticated enough and willing to explain their reasoning to us.
Might be easier to come up with new models with analytic solutions if you have a probabilistic model at hand. A lot easier to evaluate against data and iterate. Also, I wouldn't be surprised if we develop better tools for introspecting these models over time.
Perhaps for understanding the structure itself, but having the structure available allows us to focus on a coarser level. We also don't want to use quantum mechanics to understand the everyday world, and that's why we have classic mechanics etc.
These processes are both beyond human comprehension because they contain vast layers of tiny interactions and also not practical to simulate. This tech will allow for exploration for accurate simulations to better understand new ideas if needed.
We could be entering a new age of epicycles - high accuracy but very flawed understanding.
I'm not a scientist by any means, but I imagine even accurate opaque models can be useful in moving the knowledge forward. For example, they can allow you to accurately simulate reality, making experiments faster and cheaper to execute.
There will be an iterative process built around curated training datasets - continually improved, top tier models, teams reverse engineering the model's understanding and reasoning, and applying that to improve datasets and training.
This is a neat observation. Slightly terrifying, but still interesting. Seems like there will also be cases where we discover new theories through the uninterpretable models—much easier and faster to experiment endlessly with a computer.
I think it creates new studies, such as diagnosing these models behaviors without the doctor having an intricate understanding of all of the model's processes/states just like with natural organisms
As a tool people will use it as any other tool, by experimenting, testing, tweaking and iterating.
As a scientific theory for fundamentally explaining the nature of the universe, maybe it won't be as useful.
I would assume that given enough hints from AI and if it is deemed important enough humans will come in to figure out the “first principles” required to arrive at the conclusion.
I believe this is the case also. With a well enough performing AI/ML/probabilistic model where you can change the model's input parameters and get a highly accurate prediction basically instantly, we can test theories approximately and extremely fast rather than running completely new experiments, which will always come with it's own set of errors and problems.
every time the two systems disagree, it's an opportunity to learn something. both kinds of models can be improved with new information, done through real-world experiments
Hook the protein model up to an LLM model, have the LLM interpret the results. Problem solved :-) Then we just have to trust the LLM is giving us correct interpretations.
We will get better with understanding black boxes, if a model can be compressed into simple math formula then it's both easier to understand and to compute.
Is it capable of predictions though? Ie can it accurately predict the folding of new molecules? Otherwise how do you distinguish accuracy from overfitting.
What happens if we get to the stage of being able to simulate every chemical and electrical reaction in a human brain, is doing this torture or wrong?
So the Matrix?
1 reply →
Whatever it is if we needed to we could follow each instruction through the black box. It’s never going to be as opaque as something organic.
Next decade we will focus on building out debugging and visualization tools for deep learning , to glance inside the current black box
Some machine learning models might be more interpretable than others. I think the recent "KAN" model might be a step forward.
I suspect that ML will be state-of-the-art at generating human-interpretable theories as well. Just a matter of time.
This is exactly how the physicists felt at the dawn of quantum physics - the loss of meaningful human inquiry to blindly effective statistics. Sobering stuff…
Personally, I’m convinced that human reason is less pure than we think it to be, and that the move to large mathematical models might just be formalizing a lack-of-control that was always there. But that’s less of a philosophy of science discussion and more of a cognitive science one
All I can see anymore is that March of Progress illustration [1] with a GPU being added to the far right. Interesting times indeed.
[1] https://en.m.wikipedia.org/wiki/March_of_Progress
That is not a real concern, just a confusion on how statistics works :(
Engineering often precedes Science. It's just more data.
We already have the absolute best method for accurately predicting the world, and it is by experimentation. In the protein folding case, it works by actually making the protein and analyzing it. For designing airplanes, computer models are no match for building the thing, or even using physical models and wind tunnels.
And despite having these "best method", it didn't prevent progress in theoretical physics, theory and experimentation complement each other.
ML models are just another kind of model that can help both engineering and fundamental research. Their working is close to the old guy in the shop who knows intuitively what is good design, because he has seen it all. That old guys in shops are sometimes better than modeling using physics equations help scientific progress, as scientists can work together with the old guy, combining the strength of intuition and experience with that of scientific reasoning.
The ML models will help us understand that :)
> Stepping back, the high-order bit here is an ML method is beating physically-based methods for accurately predicting the world.
I mean, it's just faster, no? I don't think anyone is claiming it's a more _accurate_ model of the universe.
Collision libraries and fluid libraries have had baked-in memorized look-up tables that were generated with ML methods nearly a decade ago.
World is still here, although the Matrix/metaverse is becoming more attractive daily.
1 reply →
It’s much easier to reverse engineer a solution that you don’t understand (and discover important underlying theories on that journey), than it is to arrive at that same solution and the underlying theories without knowing in advance where you are going.
For this reason, discoveries made by AI will be immensely useful for accelerating scientific progress, even if those discoveries are opaque at first.
A New Kind Of Science?
We should be thankful that we live in the universe that obeys math simple enough to comprehend that we were able to reach that level.
Imagine if optis was complex enough that it would require ML model to predict anything.
We'd be in permanent stone age without a way out.
What would a universe look like that lacked simple things, and somehow only complex things existed?
It makes me think of how Gaussian integers have irreducibles but not prime numbers, where some large things cannot be uniquely expressed as combination of smaller things.
The top HN response to this should be,
what happens is an opportunity has entered the chat.
There is a wave coming—I won't try to predict if it's the next one—where the hot thing in AI/ML is going to be profoundly powerful tools for analyze other such tools and render them intelligible to us,
which will I imagine mean providing something like a zoomable explainer. At every level there are footnotes; if you want to understand why the simplified model is a simplification, you look at the fine print. Which has fine print. Which has...
Which doesn't mean there is not a stable level at which some formal notion of "accurate" cannot be said to exist, which is the minimum viable level of simplification.
Etc.
This sort of thing will of course will the input to many other things.
How do they compare on accuracy per watt?
So it’s okay now to publish a computational paper with no code? I guess Nature’s reporting standards don’t apply to everyone.
> A condition of publication in a Nature Portfolio journal is that authors are required to make materials, data, code, and associated protocols promptly available to readers without undue qualifications.
> Authors must make available upon request, to editors and reviewers, any previously unreported custom computer code or algorithm used to generate results that are reported in the paper and central to its main claims.
https://www.nature.com/nature-portfolio/editorial-policies/r...
Nature has long been willing to break its own rules to be at the forefront of publishing new science.
Are you an editor or reviewer?
If you read the standards it applies broadly beyond reviewers or editors.
> A condition of publication in a Nature Portfolio journal is that authors are required to make materials, data, code, and associated protocols promptly available to readers without undue qualifications.
1 reply →
Good question.
Also makes me wonder -- where's the line? Is it reasonable to have "layperson" reviewers? Is it reasonable to think that regular citizens could review such content?
3 replies →
This reminds me of Google’s claim that another “AI” discovered millions of new materials. The results turned out to be a lot of useless noise but that was only apparent after actual expert spent hundreds of hours reviewed the results[0]
0: https://www.404media.co/google-says-it-discovered-millions-o...
The alphafold work has been used across the industry (successfully, in the sense of blind prediction), and has been replicated independently. The work on alphafold will likely net Demis and John a Nobel prize in the next few years.
(that said, one should always inspect Google publications with a fine-toothed comb and lots of skepticism, as they have a tendency to juice the results)
>The alphafold work has been used across the industry (successfully, in the sense of blind prediction), and has been replicated independently.
This is clearly an overstatement, or at least very incomplete. See for instance https://www.nature.com/articles/s41592-023-02087-4:
"In many cases, AlphaFold predictions matched experimental maps remarkably closely. In other cases, even very high-confidence predictions differed from experimental maps on a global scale through distortion and domain orientation, and on a local scale in backbone and side-chain conformation. We suggest considering AlphaFold predictions as exceptionally useful hypotheses."
7 replies →
Depending on your expected value of quantum computing, the Nobel committee shouldn't wait too long.
1 reply →
> We have yet to find any strikingly novel compounds in the GNoME and Stable Structure listings, although we anticipate that there must be some among the 384,870 compositions. We also note that, while many of the new compositions are trivial adaptations of known materials, the computational approach delivers credible overall compositions, which gives us confidence that the underlying approach is sound.
Doesn't seem outright useless.
[flagged]
s/predicts/attempts to predict
AlphaFold has been widely validated- it's now appreciated that its predictions are pretty damn good, with a few important exceptions, instances of which are addressed with the newer implementation.
"pretty damn good"
So... what percentage of the time? If you made an AI to pilot an airplane, how would you verify its edge conditions, you know, like plummeting out of the sky because it thought it had to nosedive?
Because these AIs are black box neural networks, how do you know they are predicting things correctly for things that aren't in the training dataset?
AI has so many weasel words.
1 reply →
A prediction is a prediction; it's not necessarily a correct prediction.
The weatherman predicts the weather, even if he's sometimes wrong, we don't say "he attempts to predict" the weather.
The title OP gave accurately reflects the title of Google's blog post. Title should not be editorialized.
Unless the title is clickbait, which it appears this is…
Syntax error
Legal without the trailing slash in vi!
Not sure why the first thing they point it at wouldn't be prions.
So after 6 years of this "revolutionary technology", what we have to show for all the hype and breathless press releases is: ....another press release saying how "revolutionary" it is. Fantastic. Thanks DeepMind.
So much hyperbole from recent Google releases.
I wish they didn't hype AI so much, but I guess that's what people want to hear, so they say that.
I don't blame them for hyping their products - if only to fight the sentiment that Google is far behind OpenAI because they were not first to release a LLM.
Very cool, and what’s cooler is this rap about alphafold3 https://heymusic.ai/blog/news/alphafold-3
I’m inclined to ignore such pr fluff until they actually demonstrate a _practical_ result. Eg. cure some form of cancer or some autoimmune disease. All this “prediction of structure” has been in the news for years, and it seems to have resulted in nothing practically usable IRL as far as I can tell. I could be wrong of course, I do not work in this field
the R&D of all major pharma is currently using AlphaFold predictions when they don't have experimentally determined structures. I cannot share further details but the results suggest that we will see future pharmaceuticals based on AF predictions.
The important thing to recognize is that protein structures are primarily hypothesis-generation machines and tools to stimulate ideas, rather that direct targets of computational docking. Currently structures rarely capture the salient details required to identify a molecule that has precisely the biological outcome desired, because the biological outcome is an extremely complex function that incorporates a wide array of other details, such as other proteins, metabolism, and more.
Sure. If/when we see anything practical, that’ll be the right moment to pay attention. This is much like “quantum computing” where everyone who doesn’t know what it is is excited for some reason, and those that do know can’t even articulate any practical applications
1 reply →
There are a few AI-designed drugs in various phases of clinical trials, these things take time.