Comment by COGlory
2 years ago
> Not sure if you should be reminded of how alpha fold started, it started by winning a competition thought un winnable by academics. Top labs working in protein structure prediction have fundamentally changed direction after alpha fold and are working to do the same even better.
Not sure what part of "it does homology modeling 2x better" you didn't see in my comment? AlphaFold scored something like 85% in CASP in 2020, in CASP 2016, I-TASSER had I think 42%? So it's ~2x as good as I-TASSER which is exactly what I said in my comment.
>This is not the first (or even tenth) time I’m seeing an academic trying to undermine genuine progress almost to the level of gaslighting. Comparing alphafold to conventional homology modeling is disingenuous at its most charitable interpretation.
It literally is homology modeling. The deep learning aspect is to boost otherwise unnoticed signal that most homology modeling software couldn't tease out. Also, I don't think I'm gaslighting, but maybe I'm wrong? If anything, I felt gaslit by the language around AlphaFold.
>Not sure what else to say. Structural biology has always been the weirdest field I’ve seen, the way students are abused (crystallize and publish in nature or go bust), and how every nature issue will have three structure papers as if that cures cancer every day. I suppose it warps one’s perception of outsiders after being in such a bubble?
What on earth are you even talking about? The vast, VAST majority of structures go unpublished ENTIRELY, let alone published in nature. There are almost 200,000 structures on deposit in the PDB.
What ramraj is talking about: if you go into a competitive grad program to get a PhD in structural biology, your advisor will probably expect that in 3-4 years you will: crystallize a protein of interest, collect enough data to make a model, and publish that model in a major journal. Many people in my program could not graduate until they had a Nature or Science paper (my advisor was not an asshole, I graduated with just a paper in Biochemistry).
In a sense both of you are right- DeepMind is massively overplaying the value of what they did, trying to expand its impact far beyond what they actually achieved (this is common in competitive biology), but what they did was such an improvement over the state of the art that it's considered a major accomplishment. It also achieved the target of CASP- which was to make predictions whose scores are indistinguishable from experimentally determined structures.
I don't think academics thought CASP was unwinnable but most groups were very surprised that an industrial player using 5 year old tech did so well.
To add to this, the deep learning field has already moved on towards MSA-less structure prediction. None of this would be possible without building on top of the work open sourced by Deepmind.
https://www.biorxiv.org/content/10.1101/2022.07.21.500999v1 https://www.biorxiv.org/content/10.1101/2022.07.20.500902v1
To be overly dismissive is to lack imagination.
How do we know these "MSA-less" models aren't cheating (i.e. learning all MSAs implicitly from their training data)? If they are, they would similarly fail on any "novel" AA sequence (i.e. one without known/learned MSAs)
> What ramraj is talking about: if you go into a competitive grad program to get a PhD in structural biology, your advisor will probably expect that in 3-4 years you will: crystallize a protein of interest, collect enough data to make a model, and publish that model in a major journal.
All of that is generally applicable to molecular biology in general, and I don't see how the field of structural biology is especially egregious, the way ramraj is making it out to be.
Protein crystallization can be very difficult and there is no general solution. Kits that screen for crystal growth conditions usually help but optimization is needed in most cases. Then, that crystal must have certain properties that allow for good data acquisition at the X-ray facility. That’s another problem by itself and months or years can pass until you get a suitable protein crystal and X-ray diffraction dataset where you can model your structure.
9 replies →
I did rotations in multiple types of lab as part of my program and I can't say I ever found that students in regular molecular biology labs had nearly as hard a time as structural biologists; SB is its own class of hell. Given the number of papers published in molecular biology that turn out to be "gel was physically cut and reasssembled to show the results the authors desired" (it's much harder to cheat on a protein structure)...
1 reply →
Hear, hear. This is probably the best take.
> Not sure what part of "it does homology modeling 2x better" you didn't see in my comment? AlphaFold scored something like 85% in CASP in 2020, in CASP 2016, I-TASSER had I think 42%? So it's ~2x as good as I-TASSER which is exactly what I said in my comment.
Wait, stop, I don't know anything about proteins but 84% success is not ~2x better than 42%.
It doesn't really make sense to talk about 2x better in terms of success percentages, but if you want a feel, I would measure 1/error instead (a 99% correct system is 10 times better than a 90% correct system), making AlphaFold around 3.6 times better.
I think odds ratio ( p/(1-p) ) is the thing I'd use here. It gives the right limiting behavior (at p ~= 0, doubling p is twice as good, and at p~=1, halving 1-p is twice as good) and it's the natural way to express Bayes rule, meaning you can say "I'm twice as sure (in odds ratio terms) based on this evidence" and have that be solely a property of the update, not the prior.
For the lazy, this would make alphafold 7.25x better than the previous tools
Excellent comment. I think the issue is that "better" is underspecified and needs some precisification to be useful. The metric you are using here is the proper response to the question "how many times more surprising is it when method A fails than method B?". This is in many cases what we care about. Probably, it's what we care about here. The odds ratio seems to do a good job of capturing the scale of the achievement.
On the other hand, it's not necessarily the only thing we might care about under that description. If I have a manufacturing process that is 99.99% successful (the remaining 0.01% has to be thrown out), it probably does not strike me as a 10x improvement if the process is improved to 99.999% success. What I care about is the cost to produce the average product that can be sent to market, and this "10x improvement" changes that only a very small amount.
TIL, thanks for this.
> AlphaFold scored something like 85% in CASP in 2020, in CASP 2016, I-TASSER had I think 42%? So it's ~2x as good as I-TASSER
As someone who doesn't know proteins, but is decent at math, I would not describe it this way. You are assuming a linear relationship between effort and value, but more often than not, effort has diminishing returns. 80dB is not 2x as loud as 40 dB. An 8K image doesn't have 2x the fidelity of a 4K image. If Toyota unveiled a new engine that was 60% efficient tomorrow, no one in their right mind would say "eh, it's just 2x better". If we came out with a CPU that could clock up to 10Ghz we wouldn't say "meh, that's just 2x what we had".
Without being able to define the relationship here, I could just as well say that 85% is 1000x better than 42%. There's just no way to put a number on it. What we can say is that we completely blew all projections out of the water.
Again, I'm not someone working with proteins, but to me it sounds as revolutionary as a 60%+ efficient engine, or a 10Ghz CPU. No one saw it coming or thought it feasible with current technology.
I think the debate between "does amazing on metric X" versus "doesn't really understand the problem" reappears many places and doesn't have any direct way to be resolved.
That's more or less because "really understands the problem" generally winds-up being a placeholder for things the system can't. Which isn't to say it's not important. One thing that is often included in "understanding" is the system knowing the limits of its approach - current AI systems have a harder time giving a certainty value than giving a prediction. But you could have a system that satisfied a metric for this and other things would pop up - for example, what kind of certainty or uncertainty are we talking about (crucial for decision making under uncertainty).