Alphafold (opens in new tab)

I see an interesting geometric similarity between the two models, namely the attention mechanism that learns relationships between structure in embedded reference frames (i.e. 1D,2D embeddings in RoseTTaFold or "local" frames in Alphafold) and the true structure of the intrinsic space reference frames (3D coordinates in RoseTTaFold and "global" frames in Alphafold)

Very cool! Great to see this competition between academia and industry yielding improvements on all fronts.

Could it be that AlphaFold 2 was open sourced in response to this?

> With RoseTTAFold, a protein structure can be computed in as little as ten minutes on a single gaming computer.

I guess its a little less accurate but the quick compute time makes as much difference too. E.g. research students can have multiple less costly mistakes before achieving what they want with the software.

That is the way, unlike AlphaFold, to publish everything in open source. Kudos to the research team!

>A disturbing thing is that the architecture is much less novel than I originally thought it would be, so this shows perhaps one of the major difficulties was having the resources to try different things on a massive set of multiple alignments.

A similar concern has sparked some worries about "AI overhang" https://www.lesswrong.com/posts/75dnjiD8kv2khe9eQ/measuring-...

Most of the compute in ML research seems to be going into architecture search. Once the architecture is found, training and net finetuning/transfer learning is comparatively cheap, and then inference is cheaper still. This implies we could see 10-100x gains in AI algorithms using today's hardware, or sudden surprising appearance of AI dominance in an unexpected field. (Object grasping in unstructured environments? Art synthesis?) A task could go from totally impossible to trivial in a year. In retrospect, the EfficientNet scaling graph should have alarmed more people than it did: https://learnopencv.com/wp-content/uploads/2019/06/Efficient...

Waymo has been puttering along for years, not announcing much of interest. This may have caused some complacency about self-driving cars, which is a mistake. Algorithms only get better, while humans stay the same. Once Waymo can replace some human drivers some of the time, things will start changing very quickly.

> A disturbing thing is that the architecture is much less novel than I originally thought it would be, so this shows perhaps one of the major difficulties was having the resources to try different things on a massive set of multiple alignments. This is something an industrial lab like DeepMind excels at. Whereas universities tend to suck at anything that requires a directed effort of more than a handful of people.

Yeah, the HN commentary on Alphafold has a high heat-to-light ratio. I'm eager to read the paper because the previous description of the method sounded remarkably similar to methods that have been around for ages, plus a few twists.

The devil is going to be in the details on this one.

I haven't read the full paper but there is certainly some new/exciting developments i'm seeing just from scanning. The “Invariant Point Attention" which is described as a novel, geometry-aware and equivariant attention operation is pretty huge.

Something along these lines was speculated to be to be used by Fabian Fuchs [0] soon after the original CASP competition. Basically, it's a huge win for the geometric deep learning people, and indicates an exciting direction for mainstream academia to move in.

https://fabianfuchsml.github.io/alphafold2/

I'm genuinely curious: could the output of Alphafold be fed into a classical folding algorithm (as a starting point), or is the output of Alphafold too far down the wrong path, in these cases?

Why is it disturbing? Isn't that just a values-neutral outcome, or, are you saying it's disturbing from the perspective of academia?

many of these resources are available, it's mostly that academic scientists don't have the time, money, or expertise to manage large datasets. However, the community has maintained high quality MSA database for decades and that's exactly the work that DM drafted off.

> The total download size is around 428 GB and the total size when unzipped is 2.2 TB. Please make sure you have a large enough hard drive space, bandwidth and time to download.

> This was tested on Google Cloud with a machine using the nvidia-gpu-cloud-image with 12 vCPUs, 85 GB of RAM, a 100 GB boot disk, the databases on an additional 3 TB disk, and an A100 GPU.

This is amazingly detailed for a researcher who wants to follow in the track and also Apache licensed, which is one road-bump out of the way for a commercial enterprise, like an actual drug manufacturer who wants to burn some money trying this out.

edit: said the last part too fast, the code has a "the AlphaFold parameters are made available for non-commercial use only under the terms of the CC BY-NC 4.0 license"

There's a preview paper as well: https://www.nature.com/articles/s41586-021-03819-2

The process is described in Supplementary, but where do you see the code to train the model? The repository is the inference pipeline.

My understanding is that it has been manually tested. I.e. it has produced correct results to previously intractable problems. I'm not sure how much automated testing would add at that point.

Prior to this model, protein folding hadn't seen significant advancements in a decade or more. Worrying about the lack of tests in a first of its kind model is very much akin to complaining about the choice of font in a user manual for the world's first warp drive. I understand you're attempting to frame the problem in terms in things you know, but trying to weigh down pioneering research with professional development ceremony is very much counterproductive. The 'missing' ceremony would not have contributed to the strength of AlphaFold's result, the model's only purpose was to compete within the context of an existing validation framework.

Because it passes the huge number of integration tests.

Research code is highly volatile: the details and architecture changes a lot. It is much more important to invest the time into writing more experimental code and validate it with e2e functional tests that don't need to change, than to constantly having to rewrite the code and the tests.

Artbreeder has some interesting prior art here: nVidia forbid commercial use of StyleGAN, but artbreeder disregarded it and happily sold all the breeding you wanted. No one seemed to care.

I suspect that the clause is there to prevent a startup launching on the basis of “see this trained model? Yeah, that’s literally our business model” though, which is a mildly amusing thought, wot wot.

So basically, a few tens of thousands, sure. A few million, big G might have a problem.

Still, the smart move would be to launch the business anyway, and gamble that you can work out a licensing deal.

If you took their parameters, then trained it for while on a different set of data, it would vary from the original. I wonder how much compute would be required to make the offset far enough to hold up from scrutiny, and in court.

Alternatively, you could manually change the network model, add a few hidden layers, etc... modifying the parameters in step, and result in a new model and new parameters. Some training to vary the parameters, and it's now a new work.

Here's where I think we need to be going: You go to a doctor's office, sick. 1) They take a blood sample. 2) They find the malignant bacteria and DNA sequence it. 3) If it's a known strain, they know what antibiotics to use on it. 4) If not, they solve protein folding on the genes. 5) From that, they see which existing antibiotics would kill it. 6) If none will, then given the proteins, they have to derive a new antibiotic.

1) is easy. 2) might not be - there can be a lot of things in a blood sample, and finding only the interesting (bad) things might not be simple. The sequencing part is pretty much solved. 3) would take a bit of work, but I think it's possible now. 4) we're getting there. 5) might have a fair amount in common with 3), but it probably takes some additional work. 6) is... probably non-trivial.

That's just one research agenda. There are others. You may have to move to related work, but I doubt you're going to be out of a job in this lifetime.

why would it put you out of job? Wouldn't it just become one of the tools you use?

I've worked in bio and drug discovery for some 25 years. That includes building classifiers using gradient descent in the 90s (when algorithms, computers and data were all much worse). I ported DOCK to Linux in ~96 or 97. Since then I built an academic and then industrial career with some emphasis on using computing to solve problems in drug discovery, but I don't play that role any more.

It doesn't look like the models produced by this would immediately turn the challenging problem of finding, approving, and marketing successful pharmaceuticals (IE, it doesn't eliminate any real bottleneck).

There was a long-term dream of structure-based drug discovery based on docking, but IMO, it has never really proved itself (most of the examples of success are cherry picked from a much larger pile of massive failures).

I haven't worked on the drug-side of things, but here my bio perspective: It's kind of out-of-vogue, but consider the "lock and key" model of proteins and small molecules (drugs). For drug design, what you want to do is get a key that fits just one lock (to pull whatever lever) and not others (to avoid side-effects). It's relatively easy to find a molecule that fits a protein, because that protein is what you might spend years researching and probing, but it's tricky to check if it does anything against ~100,000 others in humans. If you could do an in silico computational survey to be like, oh, maybe it'll target this accidentally, you could spot-check those in vitro, and/or stick on some other atoms to your small-molecule to make it not fit that off-target.

Holy grail, IMO, though is being able to design de novo protein sequences (to make "biologics", aka engineered protein drugs) that can a) target (bind/block/enhance) or do (chemical reactions) what you want and only that, b) are easily synthesizeable by bacteria/yeast (cheap to make), and c) are stable (easy to transport/store).

> Could you give an overview of how people can leverage this (or how you might?).

Short answer: nobody knows. Traditionally, protein folding is a solution in search of a problem, but that's largely because the predictions were...unusably bad. This was always more of a super-difficult validation problem for the force fields and simulation methods, which could then be used for other problems of greater value (such as rational protein design, or simulation of the motion of proteins with known structures).

These predictions are better, but still pretty far from the level of precision that you'd want for any kind of rational drug design, where the exact locations of protein side-chains (for example) matter a lot. You'll note that AlphaFold returns structures that are "relaxed" using one of the oldest simulation systems for proteins: AMBER. So it's not exactly a clean-room solution to the problem, and you can't assume that the details (which matter to drug design) are going to be any better than for the older methods.

But that said, if you have a method that can reliably give you a blurry view of the overall shape of a protein, even that could be useful for things like target discovery or inference of biological networks. But this is still a lot closer to pure research than "revolutionizing drug discovery", as is frequently batted around on reddit, HN and the press.

You can't do intelligent drug design if you don't know what the target protein looks like. We've gotten great at solving protein structures with things like crystallography and cryo-EM microscopy. Unfortunately, many interesting drug targets reside in the membrane of a cell, which means you can't easily work with them in a lab because they aren't soluble in anything but a plasma membrane. For instance, this is an issue with the 5HT2A protein, a g coupled protein receptor that is implicated in many serotonin related pathways.

Being able to predict what it would look like would be a huge deal because then you can go about intelligently designing drugs for it.

It can be an aid in drug development, and can perhaps assist a bit in tuning small molecule drugs for more stable binding.

Though I think the major impacts will be two-fold:

(1) The field of structural biology is going to see a change, with much more data available. Some structures of difficult to crystallize proteins will be solved, which may lead to much greater biological understanding. We may enter a time, where once you have a primary sequence, you also have a likely 3d-structure, which will probably change the daily work of quite a few biologists a bit.

(2) Industrial protein design. A tool such as this can potentially have great utility in optimizing proteins as chemical catalysts for various processes in different industries. This includes expanding the conditions under which a protein is active and also making their conformation more stable and so the protein more long-lived in solution.

I work in cancer research with a drug discovery focus in a lab with some structure biologists. My understanding is that if we identified proteins targets suitable for therapeutics then understand its structure to identify secondary binding sites could be crucial for drug discovery. Drugs can then be designed to modulate its biological functions.

> (it's NP-complete after all)

Protein folding is a physical/biological phenomenon. AFAIK we don't currently have a proper exact mathematical formulation of the problem that would let one determine its complexity.

You may be referring to this paper [1]. It only claims that one particular optimization problem, believed to give a solution to protein folding problems, is NP-hard. So, even if a suitable exact formulation exists, it is not yet proven that protein folding is hard, although it for sure seems plausible.

By the way, it is perfectly possible today to solve some very large-scale NP-hard problems (think millions of variables and constraints) in reasonable amounts of time (think minutes or hours). Examples are knapsack problems, SAT problems [2], the Traveling Salesman Problem [3] or more generally Mixed Integer Programming [4].

[1] "Complexity of protein folding", 1993, by Aviezri S. Fraenkel

[2] http://www.satcompetition.org

[3] http://www.math.uwaterloo.ca/tsp/

[4] http://plato.asu.edu/bench.html

The protein folding problem is not NP complete. The "formal" protein folding problem, as posed (find the set of dihedral angles whose resulting structure has the lowest energy) might be, but that bears only a distant resemblance to how people "solve" the problem today. At the very least, the statement is incorrect because many proteins don't actually fold to their energy minimum, they get stuck in kinetic traps, and the formal PF defintion never accomodated that idea.

Not really an answer to your question, but is the problem really NP-complete, or just combinatorially difficult? For example how is this condition of NP-completeness satisfied?

> it is a problem for which the correctness of each solution can be verified quickly [0]

[0] https://en.wikipedia.org/wiki/NP-completeness

I dont know much about protein folding, but for most things in life,exact solutions to NPC problems usually aren't needed for non-contrived problems. In many cases, approximations are good enough.

Besides, this is real life - if predictions and real life match, that's great. If they don't, well you know you went wrong somewhere.

A very-non-expert opinion, if an approach approximates it pretty well and can be improved upon, then it could end up being quite useful. Given that biology exists on a real, tangible scale then perfection in the fold prediction isn't necessary, instead just an approximation that is sufficiently good to be functionally useful.

^ That sounds like word-salad BS but I think there's some truth to it. I know protein folding has been postulated to be useful in terms of understanding basic biology, understanding disease pathology, and drug prediction. While a wide range of approximations are functionally useless, perhaps the Alphafold approach or some improved version of it surpasses the functionally useful threshold.

At least I hope so

Yes, it is still useful. Even structures obtained through traditional means (eg. x-ray crystallography) are approximations to an extent since there are limits to the resolution that you can obtain and oftentimes regions of proteins are "disordered". Additionally, these structures are only snapshots of a protein in a particular state, which may not completely reflect the dynamics of the protein in its native environment.

You want to find a protein that has X structure (since structure determines function to a degree).

If AlphaFold is substantially more accurate at solving proteins, it can mean that drug discovery is faster, assays are faster, etc. etc.

The "unexpected problems" would be caught in the assay stage.

I would expect that once AlphaFold has helped you identify a potential protein (e.g. as a drug) out of a bigger set of potential proteins, there will still be a manual step of traditional cryoEM, NMR, etc. to get an accurate high-resolution structure.

To me, the interesting thing is not the specific results but rather that you can accurately predict crystal structures from sequence alone. This begets the question: what other physical biological properties can we predict?

Is it really np complete? If so we could map other np complete problems onto it and let biology solve it for us.

NP completeness tells you about the hardest cases, not the most useful cases.

AlphaFold is not about solving any kind of NP-complete problem.

Proteins consist of chains of amino acids which spontaneously fold up to form a structure. Understanding how the amino acid chain determines the protein structure is highly challenging, and this is called the "protein folding problem".

People use mathematical models to predict how proteins fold in nature. Many such mathematical models are stated in terms such as "proteins fold into a configuration that minimizes a certain energy function". Even the simplest such models [1] give rise to NP-hard decision problems, which are also known (somewhat confusingly) as "protein folding problems". To make this a bit less confusing, I will call the mathematical decision problems PFPs.

Like all mathematical models, our protein folding models don't correspond exactly to reality. Even if you are somehow able to determine the exact mathematical solution to a mathematical PFP, that _still_ doesn't guarantee that the real protein that you were trying to model behaves like the mathematical solution would indicate. E.g. the protein may fold in such a way that it gets stuck in a local optimum of the energy function you were using.

How do we detect this? We make inferences about how the protein should behave, given the mathematical solution to the Protein Folding Problem, and then we perform experiments, and find out (empirically) that the protein behaves in a manner that is inconsistent with the inferences drawn from the mathematical model. Scientists _do_ do this. And they would have to do it even if they had a fast, exact way to solve NP-complete problems, because the NP-complete problems are still just part of a mathematical model, and need not correspond to reality in any way.

The success of AlphaFold is not measured by how well it solves (or approximates) mathematical PFPs. The success of AlphaFold is measured by making successful predictions about how certain proteins will fold. And this is exactly how it was tested [2]: they threw it at a bunch of problems for which scientists have empirically determined how certain amino acid chains fold, but didn't release the results. And then they compared the solutions predicted by AlphaFold, and found that most of the predictions were consistent with what they knew to be the case.*

[1] https://en.wikipedia.org/wiki/Lattice_protein

[2] https://predictioncenter.org/casp14/index.cgi

* That's an understatement. The solutions were really very good, much better than those produced by any other submission to CASP14.

Unless I'm mistaken, you could train the model yourself, starting with a random set of values. In time, your error rates would be low enough to have a new set of parameters which you could use however you like.

Nevermind

> The simplest way to run AlphaFold is using the provided Docker script. This was tested on Google Cloud with a machine using the nvidia-gpu-cloud-image with 12 vCPUs, 85 GB of RAM, a 100 GB boot disk, the databases on an additional 3 TB disk, and an A100 GPU.

2.2TB data

Working with one of the team providing the uniref dataset used here. Running any kind of torrent stuff in an university network setting is a central policy fight one just does not want to get into at all.

On the other hand outbound network traffic from an university is "free". So the benefit is absolutely minimal from a hosting perspective.

It was tried (https://journals.plos.org/plosone/article?id=10.1371/journal...) but it is gone the way of the dodo for the above reasons.

Where there isn't an available crystal structure, Alphafold can be used to create initial structures for simulation via folding@home, replacing older homology modeling techniques.

Source: former folding@home researcher.

no, it wouldn't make sense to do that. Folding@Home is for ab initio where you don't have any prior info for the structure, this is for homology modelling. F@H probes the dynamics of protein folding, this just makes a static prediction.

> This is a completely new model that was entered in CASP14 and published in Nature.

From the repo:

> This package provides an implementation of the inference pipeline of AlphaFold v2.0

There is a lot of bias in the chat here from a more chemistry and pharma slant. If you ignore this AlphaFold solves in a very meaningful way the problem blocking a lot of science investigation.

For comparative and evolutionary analysis structure is far more conserved than sequence. Especially in things like viruses or anything with a high rate of reproduction like bacteria. Just knowing the general fold or overall structure is enough to do structural alignment and tell if two genes are related on that basis, even if their genomic sequence is completely dissimilar. Large groups of researchers rely on sequence homology built from sequences of known structure.

But AlphaFold works well in new sequence space to far more accuracy than is needed. If we had an AlphaFold prediction for every known sequence suddenly the evolutionary relationships between all genes and even all species would be far clearer. This on its own unlocks a new foundation to reason about function and molecular interaction with a wholistic systems view without gaps in what we can know with some reasonable assurance.

For an analogy think of the difference between having books in different languages describing objects. You know what some of the book in English might say but you dont even know if the book in Spanish is even talking about the same things. AlphaFold is like an AI that transforms all the books into picture books and now we can use image similarity or have one person look at all pictures.