[Rochus]

Not sure why this is downvoted. The comment cuts to the core of the "Intelligence vs. Curve-Fitting" debate. From my humble perspective as a PhD in the molecular biology /biophysics field you are fundamentally correct: AlphaFold is optimization (curve-fitting), not thinking. But calling it "propaganda" might be a slight oversimplification of why that optimization is useful. If you ask AlphaFold to predict a protein that violates the laws of physics (e.g. a designed sequence with impossible steric clashes), it will sometimes still confidently predict a folded structure because it is optimizing for "looking like a protein", not for "obeying physics". The "Propaganda" label likely comes from DeepMind's marketing, which uses words like "Solved"; instead, DeepMind found a way to bypass the protein folding problem.

[dekhn]

If there's one thing I wish DeepMind did less of, it's conflating the protein folding problem with static structure prediction. The former is a grand challenge problem that remains 'unsolved' while the latter is an impressive achievment that really is optimization using a huge collection of prior knowledge. I've told John Moult, the organizer of CASP this (I used to "compete" in these things), and I think most people know he's overstating the significance of static structure prediction.

Also, solving the protein folding problem (or getting to 100% accuracy on structure prediction) would not really move the needle in terms of curing diseases. These sorts of simplifications are great if you're trying to inspire students into a field of science, but get in the way when you are actually trying to rationally allocate a research budget for drug discovery.

[smj-edison]

I'm really curious about this space: what types of simulation/prediction (if any) do you see as being the most useful?

Edit to clarify my question: What useful techniques 1. Exist and are used now, and 2. Theoretically exist but have insurmountable engineering issues?

[dekhn]

Right now techniques that exist and used now are mostly around target discovery (identifying proteins in humans that can be targeted by a drug), protein structure prediction and function prediction. Identifying sites on the protein that can be bound by a drug is also pretty common. I worked on a project recently where our goal was to identify useful mutations to make to an engineered antibody so that it bound to a specific protein in the body that is linked to cancer.

If your goal is to bring a drug to market, the most useful thing is predicting the outcome of the FDA drug approval process before you run all the clinical trials. Nobody has a foolproof method to do this, so failure rates at the clinical stage remain high (and it's unlikely you could create a useful predictive model for this).

Getting even more out there, you could in principle imagine an extremely high fidelity simulation model of humans that gave you detailed explanations of why a drug works but has side effects, and which patients would respond positively to the drug due to their genome or other factors. In principle, if you had that technology, you could iterate over large drug-like molecule libraries and just pick successful drugs (effective, few side effects, works for a large portion of the population). I would describe this as an insurmountable engineering issue because the space and time complexity is very high and we don't really know what level of fidelity is required to make useful predictions.

"Solving the protein folding problem" is really more of an academic exercise to answer a fundamental question; personally, I believe you could create successful drugs without knowing the structure of the target at all.

[smj-edison]

Thank you for the detailed answer! I'm just about to start college, and I've been wanting to research molecular dynamics, as well as building a quantitative pathway database. My hope is to speed up the research pipeline, so it's heartening to know that it's not a complete dead end!

[HarHarVeryFunny]

It seems that to solve the protein folding problem in a fundamental way would require solving chemistry, yet the big lie (or false hope) of reductionism is that discovering the fundamental laws of the universe such as quantum theory doesn't in fact help that much with figuring out the laws/dynamics at higher levels of abstraction such as chemistry.

So, in the meantime (or perhaps for ever), we look for patterns rather than laws, with neural nets being one of the best tools we have available to do this.

Of course ANNs need massive amounts of data to "generalize" well, while protein folding only had a small amount available due to the months of effort needed to experimentally discover how any protein is folded, so DeepMind threw the kitchen sink at the problem, apparently using a diffusion like process in AlphaFold 3 to first determine large scale structure then refine it, and using co-evolution of proteins as another source of data to address the paucity.

So, OK, they found a way around our lack of knowledge of chemistry and managed to get an extremely useful result all the same. The movie, propaganda or not, never suggested anything different, and "at least 90% correct" was always the level at which it was understood the result would be useful, even if 100% based on having solved chemistry / molecular geometry would be better.

[dekhn]

We have seen some suggestion that the classical molecular dynamics force fields are sufficient to predict protein folding (in the case of stable, soluble, globular proteins), in the sense that we don't need to solve chemistry but only need to know a coarse approximation of it.

[DrierCycle]

I'm concerned that coders and the general public will confuse optimization with intelligence. That's the nature of propaganda, substituting sleight of hand to create a false narrative.

btw an excellent explanation, thank you.

[autonomousErwin]

What's the difference between optimisation and intelligence?

[HarHarVeryFunny]

For a start optimization is a process, and intelligence is a capability.

[tim333]

I think if you watch the actual film you'd find they don't claim AlphaFold is thinking.

[BanditDefender]

There is quite a bit of bait-and-switch in AI, isn't there?

"Oh, machine learning certainly is not real learning! It is a purely statistical process, but perhaps you need to take some linear algebra. Okay... Now watch this machine learn some theoretical physics!"

"Of course chain-of-thought is not analogous to real thought. Goodness me, it was a metaphor! Okay... now let's see what ChatGPT is really thinking!"

"Nobody is claiming that LLMs are provably intelligent. We are Serious Scientists. We have a responsibility. Okay... now let's prove this LLM is intelligent by having it take a Putnam exam!"

One day AI researchers will be as honest as other researchers. Until then, Demis Hassabis will continue to tell people that MuZero improves via self-play. (MuZero is not capable of play and never will be)

[tim333]

Maybe but the film is about Hassabis thinking about thinking and working towards general intelligence that can think. It doesn't really make claims about their existing software regarding that.