为什么 AlphaFold 没有解决蛋白质折叠问题——Google DeepMind 的 Pushmeet Kohli 与 Biohub 的 Sal Candido

1. AlphaFold: Confidence, Trustworthiness, and Interpretability

So when people say that the problem of protein folding has been solved, supposedly at a conceptual level, yes, perhaps there really are some advances. We don't know what the true ground state of a protein is or what the actual distribution of structures that proteins acquire is. What we're trying to do is essentially recreate a structure that someone has obtained and put into the PDB, and we're trying to replicate it. That's what we did, right? And it just so happens that it's useful. But that doesn't mean we understand all of the dynamics of proteins.

2. Introduction: The Bitter Lesson for Biological Data

Yeah, well, it's great to be here. What an exciting morning—so many cool announcements. I think the future of bioscience is being proclaimed right here. This is a modeling session. We are all modelers, so it is natural for us to talk about data.

To start, I think one of the things I like to think about when it comes to data is: Is there scaling? How do you scale data properly? And this brought me to the question of a “bitter lesson,” but reinterpreted in the context of data. The “bitter lesson,” for those outside of AI, is that methods that scale ultimately win. If you can scale enough, it wins. So my question, starting with Sal, is: Is there a “bitter lesson” for data?

3. Finding Scaling Laws and the Right Data for Biology

Yeah, definitely. You obviously need the right data, right, to make it work. I think one of the misconceptions about scaling laws is that they're supposed to exist everywhere, all the time. A big part of the work is finding this scaling law, right? A big part of our work is trying to understand under what conditions, by putting in more computation and more data, you get a better result.

The reason why this is a great situation is because when this happens, you can just start the process, right? This becomes an engineering problem, which I like. This applies to architecture, but to a large extent it also applies to data. If you don't have data with the right statistics to solve your problem—which I think is pretty obvious if you think about it for a second—then you're not going to get a model that has the capabilities and insights that you need.

So you can actually extract information from just the data you have and then use that to generalize it, right?

Yes.

So, when it comes to data collection, how do you determine what types of data modalities are best? And perhaps a related question: Do you think modelers tend to work on problems for which data is already available, rather than on those that best address the goal or have the greatest impact on translational medicine?

Oh, definitely. I won't speak for all modelers, but I'm lazy, so I'll work with what's at hand. I think there's both a positive and a negative side to this. The upside is that when you do machine learning, or let's say traditional machine learning, you're looking for the most flawless and high-quality data examples you can find.

But if you're anything like me, you're the one digging around in the back room, going through other people's junk in search of what's there. A concrete example of this is that when we train a protein language model, we use metagenomic sequences, which are not the highest-quality data. In fact, I can guarantee that a lot of this data isn't even real, complete proteins. And yet, it increases the effectiveness of the model for designing real, working proteins and for understanding proteins that we know exist.

This is a positive aspect, but it can lead you to a negative outcome when you say, “Well, let's just scale the data that we can easily generate.” I don't think that's necessarily the right way to go.

I think that's one of the things that really excites me about what we're discussing here today at BBI. This is where we really go out and say, “What data do we need to solve this problem?” These are the right resources, but also the right community.

I think one of the things we do, or at least try to do, at Biohub is to work openly, collaborate with the community, and move it forward. This is a critical point, because if we only have those who build models, we will tend to rely on the data that already exists. If we only have those who generate data, they will tend to focus on what can be generated.

4. DeepMind's Bitter Lesson: Start with the Problem

But if we work together as a community at every stage in an open way, we can understand exactly what data we need and find that scaling law. What do you think, Sal? More or less data?

I was at DeepMind when Rich was working with us, thinking about the idea of a “bitter lesson.” What I took away from his lecture, which he gave at DeepMind, wasn't so much about the benefits of data or how to think about machine learning. I think he was talking about something more conceptual.

The conceptual thing was that sometimes, when we look at problems, we treat solutions almost like religion. “I am a modeling specialist,” or, “I am a data-generation specialist.” And I think this is the bitter lesson: If you approach a problem with this mindset, you may not succeed. The problem comes first.

You must be flexible in choosing your solution approaches. It's worth trying both options. You should try to understand what problem you are trying to solve. If you need modeling, do modeling. If you need to collect more data, collect data.

But what happened in machine learning back then was that people said, “We are machine learning researchers. Here is the dataset. Here is the training data. Here is the test data.” And we just optimize the model. This is the wrong path, isn't it?

In the sense that if your ultimate goal is to solve a problem, you must consider both aspects of the process. The process is not just data or modeling. This is also a question of expertise.

5. AlphaFold's Data Limits and Building the Virtual Cell

The example I have is AlphaFold. We looked at what we could do with the existing datasets because we didn't have the expertise or, for that matter, the resources to say, “Let's expand the PDB significantly.” This was simply impossible, as the amount of investment that organizations and scientists around the world had made in creating this data was priceless.

Therefore, the focus here is to get the most return on your modeling investment. But in other areas—for example, if you approach cellular genomics—we took the same approach and asked, “What can be done with CELLxGENE?”

If you look at cellular genomics, it has become quite clear after much effort that the data is not yet sufficient to realize the ambitious goal of creating a virtual cell. This encourages people to come together and focus on real challenges for the progress of science, instead of blindly following only trends in data generation or modeling.

I really like this answer. So, what do you think is the main practical takeaway from this? That is, always first define the problem, and then understand in which solution space to look for the answer.

More broadly, how should the community think about this as we move forward into the next generation of translational medicine research?

The advice I give to anyone starting out in this field is to consider yourself a multidisciplinary expert. First, understand the problem itself. Why are you working on this problem? What are we trying to achieve? Then think about what exactly this will require: modeling, computational scaling, or data generation.

Understanding the problem is extremely important. Of course, then you need to build your own expertise. Of course, you will face certain limitations. For example, there is only a certain amount of data you can generate, or a certain size of model you can afford to train.

Understand these limitations and try to fail quickly. Try to see what approaches will be appropriate in the long term to reach the level of results and the level of impact that you are striving for.

6. Handcrafted Architectures vs. Scaling Compute

Okay, great. That actually brings me to the next question. When I look at the evolution of modeling, let's say AlphaFold 2 was essentially a work of art—a bunch of very carefully handcrafted functions. Every part of its design was very thoughtful and purposeful.

I think some of the work that Google or Alphabet is doing still falls into that realm, but the general consensus seems to be moving toward more scalable, more universal strategies. Do you think there are still many niche, “artisanal” solutions or problems that need them?

Is it better to spend our resources—because resources and funds are limited—on computing, on talent, or on data? Should we focus primarily on scaling?

Well, again, let's approach this from first principles. Even when you think about the “art” of building a better model, I don't think that was an accident, right? We experimented a lot, but behind it was a vision—this whole scientific intuition that came from biophysics and biochemistry: These interactions, these amino acid residues, don't just exist on their own but are influenced by other residues. So let's take that into account.

7. Good Data, Inductive Bias, and Model Design

If you have learned something from the scientific community, why not use it? Use this information, try to encourage the model, and give it that unfair advantage. This essentially makes the model much more efficient at working with the data because it doesn't have to recreate everything from scratch.

I should also point out that curating quality data is an art in itself, right? It's not like you're saying, “Well, give me more data.” If you just duplicate the same amount of data, you won't get anywhere. So it's not just about big data; it's about quality data. Actually understanding what kind of data is needed to make progress on a problem is, I would say, a much more interesting and much more difficult problem in itself.

Yes. You seem to have an opinion on this, so keep going.

Yes, I completely agree. As you say, it really depends on the problem that needs to be solved. If you have less data, then it's important that having inductive bias in the model will help you, right? For example, this is really necessary to get results with smaller amounts of data.

As you get more and more data, you see that sometimes the model can find something that you didn't necessarily know about. Sometimes that inductive bias, if it wasn't quite right, can hold you back. There's definitely a tipping point where things get better, but I also partly object to your question because I think there's actually a lot of skill involved in scaling. There's a lot of algorithmic work to do to take these models, train them on more data, do more computation, and make them bigger.

8. Beyond Transformers: Evolving AI Architectures

We see this not only from an infrastructure perspective, where models are inferring faster, but also in the fact that we're moving beyond transformers to more specific architectures that can—well, I don't think we're in a post-transformer world, but we're modifying those architectures to make them more fit for purpose and more efficient, even on Internet-scale data. So, I guess this is kind of a project, right? As data becomes more abundant, it's not just about helping curate that data and choosing the next piece of information that we actually need. This is more information than data: how do you get that next packet of information that models need?

9. AlphaFold's Limits: Structure, Function, and Dynamics

But also, at every step and at every scale, what architectures are best to get the most out of this data? If you think about the state of protein structure prediction now, the news would say that the protein structure prediction problem has been solved. But if you talk to my friends, they all say, “Well, you still have so much to do here.” So many things, such as function, dynamics, and design, are still simply open problems.

It's been about 5 years since AlphaFold 2 was announced. Progress has been made, but I'm curious, Pushmeet, do you think another big leap is expected here? Are there any obstacles to solving these problems? Maybe we don't have the necessary data yet, or maybe we lack the right algorithms or ideas. Or is this something we will see in the future, and it's just a matter of time?

Yes, I think science works by isolating one thing and moving forward step by step. So when people say that the protein-folding problem has been solved, on a conceptual level, maybe yes—there have been some advances. But I think it also applies to the narrative that proteins are building blocks. Proteins are not blocks; they don't act like blocks, right? I keep saying that proteins are building blocks, but I don't really believe that.

Proteins are extremely complex. They are disordered. Their form may change depending on the context. And what AlphaFold did—in fact, John Jumper and I discussed it at one time—was raise the question: What exactly are we trying to solve? We do not know what the true ground state of a protein is or what the actual distribution of structures it adopts is.

What we're trying to do is this: someone got a structure, uploaded it to the PDB, and we're trying to reproduce and get the same structure. That's what we did, right? It just so happened that it turned out to be useful. But that doesn't mean we understand all of the dynamics of proteins.

I think it's easier at the highest level to say that we've made progress, but the scientists among us, in this audience, know the truth. Isn't that right? I think it's really important for us to understand that and have a shared vision that there's still a lot of work to do. We've made progress, and there's a lot to be excited about, but let's not stop funding protein structure prediction and protein dynamics, because we're just getting started.

10. Cryo-EM and the Next Protein Breakthrough

So what do you think is the biggest obstacle? If there's one thing that you could wave a magic wand at and say, “We have more of this,” something that would accelerate one of these things—function, dynamics, or design—what would this magic wand be like? What would you make a reality?

Oh, basically, when you ask me this question, my background is pretty eclectic. I started as a security researcher, then moved into computer vision, Bayesian theory, and finally got into discriminative machine learning, deep learning, and so on, then AlphaFold programming, and finally science.

When you ask me this question, my inner computer vision researcher was extremely fascinated by cryo-EM micrographs. I was thinking, “What kind of PDB data is this? I should be working with the original source, right? I should be looking at cryo-EM micrographs. I don't need these structures. They probably miss all the data.”

I would have to work directly with cryo-EM micrographs, creating models that can scale to that level with the right amount of data and extract all the dynamics and distributional information captured there. That would be an amazing thing. I tried that, but it takes more work, right?

So, work on it, however you believe this is the way to go.

11. From Individual Proteins to Whole Biological Systems

Yes. I think at some point in time, maybe someone better than me will take this on and we'll achieve something.

What do you think, Sal?

Yes, I think it's really interesting because these models are quite useful, but they're not exactly solving the problem that most people want to solve. They serve a very specific purpose, and you can also use them for other things, like designing new proteins, right? Although these are not what you would start with, they are very useful.

But I think of the models we're building now as if you're trying to understand how a bicycle works and you're only modeling one spoke on the wheel, right? These models can get better over time, but a lot of what really needs to be done is moving from spoke models to wheels, and maybe even whole bikes, because that's what people really want to understand.

In fact, to continue the analogy, it seems like people want to use a bicycle model to design pickup-truck parts or something like that, right? I think that when we implement these models in the specific contexts that they're in, we can learn more about these interactions in a broader sense. I think that's what it's all about.

But I agree with you. I think we should continue to work on folding models, because I think they will get better and better.

Regarding design, there is a famous quote by Feynman: “What I cannot create, I do not understand.” We now live in a world where it is very easy to create things without understanding them at all. So I'm wondering what your feelings and thoughts are on how important it is to have models that can help people understand things, versus magical “black boxes” that can effectively create a picomolar binder the first time or something like that.

I think we were designing things with magical “black boxes” long before AI came along, right? So in a way, I think it's still useful. But I think understanding is really important, and that's actually one of the big things that I think about a lot in terms of AI.

There's so much that these models have to learn, and as intelligence becomes cheaper and more accessible, you can apply it to learn more and more about what's going on in the world. But then how do you actually extract that knowledge from the machine, so to speak, right? It's something I can understand. Maybe it's just my esoteric curiosity.

I think there are so many things worth exploring. Many scientists really want to understand things, and the end results may not be that important. But we're here to solve the problem of translational medicine, right?

12. The Hidden Knowledge Inside Protein Language Models

Yes. I mean, it seems to me that the deeper you get into something, the more you find the right ways to keep moving forward. One thing that's really important to me is that models contain a lot more information than we know.

For example, we have worked a lot on interpretability for our models. You find a lot of information there, because I think people know that protein language models learn some idea of structure in their representations. But we find information about, say, some functions. We find information about certain movements, right?

So I think there's still a lot to discover there, even in the models we have now. I think it's important to understand how these models work as we increase their capabilities. After all, you expect a model of the world to emerge from trying to squeeze all this information into a model, right?

How does the model do its job? How does it design a protein? It compressed all the information of evolution into this model, you know? So, besides the fact that it can reveal something useful for us, we can learn a lot just by looking inside.

What do you think, Pushmeet? Design or understanding?

I have a slightly different perspective on this, as I believe that when we think about models, a certain level of understanding is necessary. So I will tell you what level of understanding is required.

AlphaFold is actually not perfect. It has 90 GDT. AlphaFold 2 had 90 GDT at that time on that dataset. But even if it had 95 GDT, the pLDDT reading would be completely uncalibrated. Who would trust it? It would just magically give good answers, but then suddenly it would say, “Here's the answer; I'm very sure,” and you would work on it for a year and find out it was completely wrong.

The calibration of the uncertainty measure was extremely important. So, in that sense, we really understand, and we've made significant progress in understanding how AlphaFold 2 behaves. But that's different from exactly how it worked and found a solution.

Here I agree with Sal that there is a concept of interpretability. Why did it work? Why did it give that answer? I think, if you think about it, AlphaFold 2 was, at the highest level, interpretable, because if you look at it, it had the ability to generalize, and we didn't really discover that until launch.

When we launched AlphaFold 2 and made the weights available, people found out that it was a great tool for predicting disordered proteins.

13. Could AI Understand Other Models Better Than Humans?

It could determine which elements of the protein were disordered. This just proves that it is generalizing. Interpretability raises the question: interpreted by whom? If you are talking about human interpretability as a rational system with the cognitive and computational limitations of the human brain, then no, AlphaFold 2 is not interpretable.

But if you're asking whether AlphaFold 2 is interpretable for a much larger and more advanced model in terms of how it works, maybe so. We just don't understand it. Isn't that right?

But I think, as users of AlphaFold 2, we need to understand what it can and can't do. That means understanding the behavioral characteristics, strengths, and limitations of the model. I think this is extremely important, and we can't just use these models without this kind of behavioral analysis, because otherwise they can hurt us instead of helping us. Isn't that right?

So your position is that interpretability is not strictly necessary, but ensuring the model is reliable so that people can make effective decisions is what you should focus on.

That's right. I also think that interpretability depends on who is observing. Who exactly interprets it?

If it's a human scientist trying to interpret how the model works and how it will make this prediction, that's a different question from asking whether a much larger LLM could interpret it. If you give it access to the activation levels of one of these big, advanced models of the future and ask, “Okay, tell me: can you predict what AlphaFold will do?” perhaps it can predict and propose a theory of exactly how AlphaFold 2 interpreted and produced these results.

14. When Will AI Transform Drug Discovery and Reach the Clinic?

Okay, this is a 2-minute warning. I have one last question for both of you, a short one. There is a real chance that AI will make significant improvements in human health in the near future. I love quantitative predictions. How long do you think it will be before we see the results of AI in clinics? Pushmeet, do you want to start?

“AI results in the clinic” is an incorrect question. AI is already being used at every stage of the drug development process today. From that perspective, it's already there, right?

But if you ask me when we'll see a 10x or 100x acceleration in time, that's a different question. What exactly are we accelerating? Is it the optimization of conductive compounds? Is it target search? Is it preclinical research or toxicology work?

15. Feynman's Principle: Creating Without Understanding

I think over the next few years—and that's why this initiative that we're announcing today is so important—we need to solve some of these complex biological problems. Only then will we be able to get a real breakthrough that will radically change the drug development process. AI will be present in clinical practice and will constantly influence the things that happen there, but the real large-scale acceleration will only be possible through a better understanding of the biological models that this initiative is trying to create.

Okay, thank you. Quick question—time is almost up, but—

Yes, I'm honestly willing to literally put on my “biology hat” to answer this question. But I think it's extremely important to think about this. As you said, what does it mean to move the industry forward? What we really want to see are these results. When will there be a medicine created entirely with AI, or something like that?

I don't know. It's hard to say for sure. But I really think we're going to see rapid progress in this area very, very soon, because all of these tools are already being actively used.

16. Curing All Disease: Why Biology Needs 10x Breakthroughs

Going back to your point about the basics, we need basic research and fundamental understanding. A long time ago, early in my career at Google, I realized one thing: sometimes it's easier to approach a problem by thinking about what it takes to improve by 10 times, rather than by 10 percent.

This is not at all because that path is easier, but because it allows you to look wider and see solutions that you haven't considered before. It helps you understand what the real path to the goal is by returning to first principles. How do we do this? How do we actually move this forward?

I think both approaches—both 10 percent and 10 times—are valuable, and we should apply both. I'm very glad that at Biohub we have, in my opinion, a beautiful and lofty mission: to cure all diseases. If you want to achieve that, you really need to understand what the 10x approach is all about. I think this is a great opportunity to go down this path.

Perfect. Thank you both. Thank you for experiencing this for yourself, literally. Aha. Thank you again.