Good post. My post was the one that kicked this discussion off, so I'll bring my comments from the thread here too.
On that thread, I made this point to halvorz:
> How much of protein folding that people care about falls into the category that AF is good at? I feel like “naturally evolved, stably folding proteins that exist in a single major conformation” was the majority of what people cared about 10 years ago...In general research will always push forward to the areas where we know the least, so I'm not super surprised that you as a researcher spend most of your time in exactly the areas that existing tools struggle most with.
So here. You made the claim on that thread that, on a scale from "0 to 10, where 0 is nothing and 10 is the entire problem is solved. Before AlphaFold we were at 1. Now we’re at 1.5. Yes, it’s a huge improvement over 1, but there’s still 8.5 points to go."
Obviously this is subjective and there's no hard-and-fast rule about these things, but I think I just disagree with you on this.
AF has provided a database of 200 million proteins, and that too arguably the ones in the categories that we most care about. Yes, it is true that there are edge cases -- I spent enough time in molecular biology to know this -- so perhaps it is wrong to say that AF 'completely solved' protein folding. But also, before AF2, we only had 200k protein structures. So AF improved our protein coverage not by 10x, or 100x, but 1000x.
I think you are right that AF does not give us explainability, but that to me is shifting the goal posts. In 2015, it was assumed that being able to predict protein structure from sequence necessarily meant you had a good understanding of the dynamics. In 2025, we realize that you can figure out the former without the latter. It turns out for most *applications* we dont really care about the latter.
And of course, I think very few people now believe that the path to 'solving' the 'rest' of protein folding is going to come from techniques like xray crystallography or whatever. Even though you hint otherwise, imo the protein configurations that we currently *can't* solve -- the limitations of AF -- are more likely to be solved by other AI tools or applications of AI tools, than they are to be solved by a deep rules-based theoretical understanding of electron charges and physics, at least in the short term. So, fine, AF may not have 'completely solved' protein folding semantically. But I think what most people are grasping at when they say things like "AF has completely solved protein folding" is that AI has revolutionized this field that was previously seen as totally intractable, as it has done in many other fields in similar settings.
I think you are severely underrating the importance of those limitations. Disordered regions and conformational changes are “edge cases” in software but not in biology. They are core to how proteins actually function. And, predicting how a mutation affects structure or stability is one of the most obvious uses for a hypothetical perfect structure predictor.
I wouldn’t say that AF is going from 1 to 1.5, maybe like 1 to 7. It’s a huge leap forward, but that last little bit is very important, and OP is correct that there isn’t an obvious way to tackle it with AI until our measurement improves. This is largely why RNA structure predictors are mostly terrible.
Yea, fair, but give it some time to bake. AF is less than 7 years old, AF2 only 5 years. I don't think any of this is easy, but also I don't see any *structural* reason that AI systems can't further improve along these axes. If AI tools can't do all these things yet, to me that's more a function of time than capability. The bitter lesson comes for all of us eventually.
I think the name ‘AlphaFold’ certainly doesn’t help disambiguate the problem being solved (structure prediction) from the bigger problem of protein folding. Great article!
Have you heard of Labbot? I’ve been really impressed by it. It tackles many of the challenges mentioned here, not by predicting structure, but by experimentally tracking how proteins behave in real conditions. Things like conformational changes, aggregation, binding events... even under temperature shifts or pH changes. It’s a great complement to AlphaFold when you need to understand what proteins actually do, not just how they might look.
Good post. My post was the one that kicked this discussion off, so I'll bring my comments from the thread here too.
On that thread, I made this point to halvorz:
> How much of protein folding that people care about falls into the category that AF is good at? I feel like “naturally evolved, stably folding proteins that exist in a single major conformation” was the majority of what people cared about 10 years ago...In general research will always push forward to the areas where we know the least, so I'm not super surprised that you as a researcher spend most of your time in exactly the areas that existing tools struggle most with.
So here. You made the claim on that thread that, on a scale from "0 to 10, where 0 is nothing and 10 is the entire problem is solved. Before AlphaFold we were at 1. Now we’re at 1.5. Yes, it’s a huge improvement over 1, but there’s still 8.5 points to go."
Obviously this is subjective and there's no hard-and-fast rule about these things, but I think I just disagree with you on this.
AF has provided a database of 200 million proteins, and that too arguably the ones in the categories that we most care about. Yes, it is true that there are edge cases -- I spent enough time in molecular biology to know this -- so perhaps it is wrong to say that AF 'completely solved' protein folding. But also, before AF2, we only had 200k protein structures. So AF improved our protein coverage not by 10x, or 100x, but 1000x.
I think you are right that AF does not give us explainability, but that to me is shifting the goal posts. In 2015, it was assumed that being able to predict protein structure from sequence necessarily meant you had a good understanding of the dynamics. In 2025, we realize that you can figure out the former without the latter. It turns out for most *applications* we dont really care about the latter.
And of course, I think very few people now believe that the path to 'solving' the 'rest' of protein folding is going to come from techniques like xray crystallography or whatever. Even though you hint otherwise, imo the protein configurations that we currently *can't* solve -- the limitations of AF -- are more likely to be solved by other AI tools or applications of AI tools, than they are to be solved by a deep rules-based theoretical understanding of electron charges and physics, at least in the short term. So, fine, AF may not have 'completely solved' protein folding semantically. But I think what most people are grasping at when they say things like "AF has completely solved protein folding" is that AI has revolutionized this field that was previously seen as totally intractable, as it has done in many other fields in similar settings.
I think you are severely underrating the importance of those limitations. Disordered regions and conformational changes are “edge cases” in software but not in biology. They are core to how proteins actually function. And, predicting how a mutation affects structure or stability is one of the most obvious uses for a hypothetical perfect structure predictor.
I wouldn’t say that AF is going from 1 to 1.5, maybe like 1 to 7. It’s a huge leap forward, but that last little bit is very important, and OP is correct that there isn’t an obvious way to tackle it with AI until our measurement improves. This is largely why RNA structure predictors are mostly terrible.
Yea, fair, but give it some time to bake. AF is less than 7 years old, AF2 only 5 years. I don't think any of this is easy, but also I don't see any *structural* reason that AI systems can't further improve along these axes. If AI tools can't do all these things yet, to me that's more a function of time than capability. The bitter lesson comes for all of us eventually.
I think the name ‘AlphaFold’ certainly doesn’t help disambiguate the problem being solved (structure prediction) from the bigger problem of protein folding. Great article!
Have you heard of Labbot? I’ve been really impressed by it. It tackles many of the challenges mentioned here, not by predicting structure, but by experimentally tracking how proteins behave in real conditions. Things like conformational changes, aggregation, binding events... even under temperature shifts or pH changes. It’s a great complement to AlphaFold when you need to understand what proteins actually do, not just how they might look.
certainly a brilliant invention, but by no means a complete solution to the protein folding problem