Hacker News Reader

Comment by COGlory | Hacker News Reader

>This “protein-folding problem,” as it came to be known, baffled scientists until this year, when the papers we’ve deemed the 2021 Breakthrough of the Year were published.

This is a straight lie and the author should be embarrassed. There's not one scientist alive that can explain how proteins fold from their amino acid sequence. There's not one algorithm in existence that can predict it de novo.

I'm so tired of seeing this lie repeated over and over. AlphaFold did not gain us any new information on how proteins fold. It just got really good at modelling them from existing data where existing data is available. Which is great. But still ultimately limited by the same categorical limitations of homology modeling. It simply has not solved protein folding.

Yes it's amazing.

Yes it's worthy of lots of praise.

No, it did not solve protein folding, ESPECIALLY in the way that is described here.

View on HN

> There's not one scientist alive that can explain how proteins fold from their amino acid sequence.

As a biochemist, what I understood from this editorial was totally reasonable:

- scientists understand that electrostatic forces guide protein folding, and steric factors limit the conformations a protein can take

- scientists understand that the primary structure determines the secondary and tertiary structures of proteins

- scientists know that protein folding is aided by chaperone proteins

- scientists understand how thermodynamics and entropy affect protein folding

Despite all of the above, scientists cannot solve protein folding de novo. New tools now allow us to make better, cheaper predictions which are testable.

Sounds reasonable to me.

As a structural biologist, I have to ask, how did you draw that from this?

> seemed possible that the three-dimensional structure of a protein could be predicted based on the sequence of its amino acids.

That simply is not what AlphaFold (or any homology modeling) is doing, and certainly not:

>This is a breakthrough on two fronts. It solves a scientific problem that has been on the to-do list for 50 years. And just like Fermat’s Last Theorem or gravitational waves, scientists kept at it until it was done.

> That simply is not what AlphaFold (or any homology modeling) is doing

Ha! The reality is we have very little idea of about what any neural net AI is doing. Not no idea as provided the raw material these things build themselves out of, but then we throw billions of training points at it with the desired results, and a vast stew of polynomials begins to self organise into something that produces the results.

How those polynomials do it? We certainly don't know from first principles. If we alter compression in a car engine, or moderators in a nuclear reactor, or tension in prestressed concrete beam and an engineer will be able to make a fair stab at the result. Ask the people who built these vats of self organising what happens if they change the polynomials coefficients - and they will have no idea, whatsoever.

Any claim you know what AlphaFold "is doing" is simply fanciful. We've crossed Aruthor C Clark's threshold of "Any sufficiently advanced technology is indistinguishable from magic". It seemed so reasonable when he said it. What he, and we didn't foresee is how it would look if you were the ones who built the technology. "I've build a machine so complex I don't know how it works" crosses a line even SciFi authors had trouble imagining. And now we are confronting it as as our new reality.

I wonder if the all the comments on HN about how AI's don't really understand the subject, or don't fit some definition of "conscious", or can't really "think" are just a form a self denial. It really doesn't matter people. The reality is the only way people could tell it was GPT-3 posting to reddit wasn't because of any of those things. It was because no human could churn out so much error free high quality text indistinguishable from what a human would produce in such a short time.

> That simply is not what AlphaFold (or any homology modeling) is doing

I don't know if we're getting hung up in semantic differences, but how would you describe homology modelling, or what AlphaFold is doing? Because for me, it takes as input an AA sequence and gives as output a 3D structure prediction.

It's not a semantic difference. That's the problem. There is a difference between knowledge of protein folding (which is what the article is talking about), i.e. understanding proteins well enough to predict how they will fold, and modeling a folded structure (which is what AlphaFold is doing, based off of other pre-existing experimental structures).

AlphaFold isn't simulating folding at all. It's modeling what the final structure should look like based off other data, not knowledge of how a protein folds. That's what is frustrating me so much with this topic. It's being sold as "we can predict how proteins will fold" as opposed to "we can model a folded protein based on its sequence". There's a huge difference in our understanding of protein folding, between those two points.

It does sound like a semantic difference. To me, and I believe to most in the field, the protein folding problem was always pretty much about figuring out the final 3D structure (the final state), and not necessarily about the act of folding (ie.: the gerundive) itself.

There is a lot of value in learning more about folding (in the gerundive sense), but the focus has always been on finding the folded structure, because as you very well know, structure is function.

The “protein folding problem” consists of three closely related puzzles: (a) What is the folding code? (b) What is the folding mechanism? (c) Can we predict the native structure of a protein from its amino acid sequence? (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2443096/).

It seems the purist definition of solving protein folding means specifying the sequence of actions to arrive at the native structure for a given input sequence. This implies your solution is grounded in a physical reality. Such approaches (i.e MD simulations) have not been very successful to my knowledge and are computationally intractable for most proteins. Even if you are able to do this, how do you verify that your folding mechanism is in fact correct? We don't have ground truth experimental data at the temporal resolution one would want to verify such a prediction.

For now, the only practical framing of the protein folding problem is to predict a native structure given an input sequence. This is simply because we have the requisite data...

We have a lot of data we don't understand and we train a model to represent that data. The model represents the data very well, but we don't understand the model. So now we have a lot of data we don't understand and a model we don't understand. What new knowledge have we acquired?

The OP is right to complain that there is no scientific knowledge gained by the mere act of modelling a poorly understood process, when the modelling is performed automatically and its results are also poorly understood. The goal of science is to gain knowledge of the processes in nature, not merely to predict their results. That's how science has advanced in the last few thousands of years and it's startling to see Nature and Science standing behind such a different idea of "science".

Reply on news.ycombinator.com