I kind of respect deepmind for simply keeping their nose to the ground and doing good work like this without overhyping it too much. More of the under promise over deliver engineering style than some of their competitors tends to do it
The linked article precisely claims that the have overhyped their results by (inadvertently? ) having their test distribution overlap with their training data, and implies that the result won't generalise to novel proteins.
I'm totally clueless about the topic (protein folding), but this stuff is very interesting.
From the article, it seems that AlphaFold 3 is just a biochem version of GPT or what? From what I've heard, the older AlphaFolds had some special tricks for protein prediction. Am I missing something?
No, it's not a miracle; everything it does works because the information to make those predictions is a collection of latent variables and DM found good ways to convert from sequence space into an embedding that approximates those latent variables.
From what I can tell it still depends heavily on having a good sequence and structure template (or templates). It tells us little to nothing about the specific details of the folding process. To me the only part that seems miraculous is that it seems like we can predict novel structures (previously unknown conformations) using small fragments of templates rather than entire protein domains.
> To me the only part that seems miraculous is that it seems like we can predict novel structures (previously unknown conformations) using small fragments of templates rather than entire protein domains.
Just to add a bit of context: Rosetta’s de novo methods, which had the highest success rates of template-free structure prediction before the arrival of ML-based protocols, use a similar approach. Picking small fragments from protein structure databases reduces the conformational sampling space a lot.
Yes, it was miraculous when they did it, too (a quarter century ago: https://pubmed.ncbi.nlm.nih.gov/10526365/). I believe they called it ab-initio, which led to a lot of complaints because ab-initio means something entirely different (I've published with David Baker,but on a different topic).
The author is making the point that Alphafold 3 is not so impressive - it is simply regurgitating its train set, and it's not so good for inference.
I think his central point is fair and interesting. The test train split is apparently legit, as they used structures released before 2021 for training and the rest for testing. However, there was no real check for duplicates, and the success rate might be inflated by a bunch of "me too", low hanging fruit structures that are very slight variations from what we know.
However, I'm not sure I agree with his skepticism. LLMs suffer from the exact same problems - getting it to write a Snake game in any language is trivial, but it is almost certainly regurgitating - , but can be useful as well. I mean, if for various reasons people are publishing very similar structures out there, there's certainly value in speeding up or reducing that work considerably.
> Different proteins can also be related to each other. Even when the sequence similarity between two proteins is low, because of evolutionary pressures, this similarity tends to be concentrated where it matters, which is the binding site.
It’s a small nitpick, but I think that the author actually meant “sequence identity” here, because his statement would make much more sense then. Sequence similarity is physicochemical in nature, and tends to be concentrated, in addition to functionally relevant sites (usually ligand-binding residues, as he mentioned), at key structural regions of the protein such as the hydrophobic core (where a high frequency of similarly hydrophobic residues is expected).
This is one of the reasons why proteins from the same family can share highly similar structures while having very low sequence identity, with highly conserved motifs (where the sequence identity is concentrated) taking care of the functionality.
Why overlapping molecules would indicate memorizing or over fitting is beyond me. Imagine a mechanic designing linkages. They may collide, but if they could pass through each other, they could work. Then they might reconfigure them. Similarly, overlapping molecules could be a step along the way to understanding if the algorithm is focused on binding structures rather than global physical structures.
15 comments
[ 4.2 ms ] story [ 36.0 ms ] threadFrom what I can tell it still depends heavily on having a good sequence and structure template (or templates). It tells us little to nothing about the specific details of the folding process. To me the only part that seems miraculous is that it seems like we can predict novel structures (previously unknown conformations) using small fragments of templates rather than entire protein domains.
Just to add a bit of context: Rosetta’s de novo methods, which had the highest success rates of template-free structure prediction before the arrival of ML-based protocols, use a similar approach. Picking small fragments from protein structure databases reduces the conformational sampling space a lot.
I agree, and that’s probably why they switched to de novo prediction/design at some point.
I think his central point is fair and interesting. The test train split is apparently legit, as they used structures released before 2021 for training and the rest for testing. However, there was no real check for duplicates, and the success rate might be inflated by a bunch of "me too", low hanging fruit structures that are very slight variations from what we know.
However, I'm not sure I agree with his skepticism. LLMs suffer from the exact same problems - getting it to write a Snake game in any language is trivial, but it is almost certainly regurgitating - , but can be useful as well. I mean, if for various reasons people are publishing very similar structures out there, there's certainly value in speeding up or reducing that work considerably.
AF3 stands as one of the greatest achievments in machine learning/structural biology we've yet seen.
They do remove duplicates by sequence similarity (filtered PDB).
Please assume the DM folks really do know what they are doing.
It’s a small nitpick, but I think that the author actually meant “sequence identity” here, because his statement would make much more sense then. Sequence similarity is physicochemical in nature, and tends to be concentrated, in addition to functionally relevant sites (usually ligand-binding residues, as he mentioned), at key structural regions of the protein such as the hydrophobic core (where a high frequency of similarly hydrophobic residues is expected).
This is one of the reasons why proteins from the same family can share highly similar structures while having very low sequence identity, with highly conserved motifs (where the sequence identity is concentrated) taking care of the functionality.