MHC Class II Peptide Binding Prediction: Challenges, AlphaFold 3, and ESM2 Explained

In this Pharmacy Times interview, Kamel Lahouel, PhD, and Cristian Tomasetti, PhD, discuss the challenges of predicting peptide binding to MHC class II molecules and why this problem is more complex than MHC class I. They mention how structural predictions from AlphaFold 3 can be integrated with sequence-based models like ESM2, along with key considerations such as model uncertainty, data bias across alleles, and computational trade-offs.

Pharmacy Times: Can you discuss the bigger picture of the findings presented in the presentation?

Cristian Tomasetti, PhD: The reason why this type of work
is important is [because] immunotherapy is a very critical new direction in terms of cancer treatment, a very promising one. And so the basic idea is that you can train your immune cells to recognize the cancer, and after training, you send them back, and the cells now can recognize much better the enemy and attack it. OK, so that's the future, and in fact, it's already happening for some cancer types. As I said, it's one of the most promising directions in cancer research. However, at the core of this approach, there is the issue of understanding among the many things that characterize the bad cells, the cancer cells, which ones we should train the immune system cells, the white blood cells, to recognize? And so this, in technical terminology, comes down to understanding how well the peptides can be, essentially things that are shown by these cancer cells can bind to, essentially, the MHC complex, which is an important element of the immune system. So, how well, essentially, the white blood cells can bind to this particular feature of the cancer cells. Since there is an extremely large number of features that you could theoretically attack, we need to essentially create a priority list and pick the one that give us the highest probability of being successful in being recognized, and an attack, by the way blood cells, and it's really about binding, okay, so that's, that's why this is fundamental work. And we believe that we provide a new approach that makes a substantial improvement in identifying and ranking this; we call it the binding affinity between peptide and MHC complex.

Kamel Lahouel, PhD: This model has very specific binding affinity prediction. It focuses on a very specific one, which is MHC class II with the peptides, as opposed to MHC class I with the peptides. And it's important for CD4 T-cell activation. CD4 T-cell activation is very important for persistent immune response, and so MHC-II is involved in persistent immune response. That's what makes this particular binding prediction problem important in our case.

Pharmacy Times: Why is predicting peptide binding to MHC class II molecules considered harder than MHC class I, and what structural features of the MHC II binding groove make this problem particularly challenging for computational methods?

Tomasetti: First of all, MHC-II is particularly important for the sustained response to the immune system. At the same time, the binding between MHC class 2 and the peptides is harder because of a few issues. The main one is that…where the binding occurs is a variable length in MHC class II, while with MHC-I, you have a very restricted space.… [You] can imagine, basically, the binding groove is blocked…with MHC-II, it's open ends, and so on and so on. In terms of prediction, the problem becomes significantly more complex. So I would say that and this are related also to the effect that this flanking regions can have on the prediction too. So overall, MHC-II binding is well known to be [that] they have the problem, which is, yeah, the answer to your question.

Pharmacy Times: AlphaFold 3 and ESM2 have fundamentally different design philosophies—one predicts 3D structure from sequence, the other learns sequence-level representations. How did your work integrate these two modalities, and why was using both preferable to using either alone?

Lahouel: So first of all, to be fair to ESM2, ESM2 predicts, so it's a language model for proteins, and we use what we call the embeddings, basically the meaning is that you take a sequence, and it encodes the meaning of that sequence, and it's an abstract meaning. It's vectors of numbers, actually, from ESM2; it is possible to create a structure like AlphaFold 3, which we do not use. And in fact, it's because AlphaFold 3 is the best known structure predictor at the moment. So AlphaFold 3 outperforms ESM2 as a structure predictor, and we use the ESM2 just as a language model, as you mentioned. Now that said. So as you said, clearly, the output from AlphaFold 3 and ESM2 that we use are different. One is the structure, the 3D structure of the complex proteins. The other is the embeddings of the text. The proteins that we see as a text. So first of all, they have complementary information, because one is purely sequence-based, just takes the sequence and looks at the meaning of the sequence. These are what we call the embeddings. The other one is more interested in the 3D structure, also, if you think about the sequence model from so the embeddings of the ESM2, they are kind of intermediary features. This is something that is very important. So they are downstream features from which you can generate the AlphaFold 3 structure. But they are more abstract. So the approach consists in combining more abstract features with the 3D structure that makes more sense physically. Now, how do we do this? We build what is called the graph, neural networks. We build the graph where the nodes correspond to the amino acid from alpha 4, 3, and in each node, we put the corresponding embedding from ESM2. So you have a node in a graph. That node has a representation, and that representation corresponds to the embedding from ESM2. So it's an annotated graph, think about it that way. The graph is built with AlphaFold 3, and the annotation is built with the ESM2. That's how we combine them.

Pharmacy Times: AlphaFold 3 produces a confidence-weighted structural prediction, not a ground-truth structure. How did you account for model uncertainty—particularly low pLDDT regions in the peptide—when using structural features as inputs to your binding predictor?

Lahouel: So this is a very important feature that is given by AlphaFold 3, and that we use in our in our model. And so AlphaFold 3, as you mentioned, doesn't give you just the 3D structure. It gives you the 3D structure plus how confident it is about the structure. It turns out that that information is very important for the binding affinity prediction. And…there is a way to look at how confident AlphaFold 3 is in its prediction and how that relates to binding affinity. So we use that information, and it's central, actually, as opposed to just one fixed 3D structure; we use it coupled with the confidence that AlphaFold 3 gives.

Pharmacy Times: Training data for MHC-II binding is heavily biased toward a small number of common alleles and anchor motifs. How did you evaluate whether your model generalizes to rare or novel alleles, and what evidence do you have that performance improvements aren't simply a reflection of better memorization of well-represented alleles?

Lahouel: This was a major concern at the beginning. What we did simply is that we evaluated the performance of our binding predictor on multiple alleles, the common ones are the DRB1 alleles, and then we evaluated the performance on the less common alleles, the DRS, DRB3/4/5 and HLA alleles, TP, DQ, which are rare. And our model didn't train that much on those alleles. We do see a slight drop in performance on rare alleles, but it’s still competitive with much less training data and state-of-the-art methods, and we are largely still in the game on those alleles.

Tomasetti: When you compare it to what has been considered the gold standard in terms of performance, being trained essentially on the same small data set, Kamal just mentioned our AUC—and this is in review and hopefully will be published very soon. It's about 10 points higher, so something like from 65%, if I remember correctly, to see 75%, this is an enormous difference. So we are very excited about these results.

Pharmacy Times: Structure-based features are computationally expensive to generate—AlphaFold 3 inference is nontrivial at scale. For a practical application such as neoantigen prioritization across a patient’s mutanome, how does your approach balance predictive gain with computational cost relative to sequence-only baselines like NetMHCpan?

Lahouel: At the inference stage, AlphaFold3 is a much heavier model than sequence-based models like NetMHCpan. That said, compared to wet lab approaches, the cost is negligible. Also, one needs to keep in mind that the cost here; the computation cost is negligible compared to performance. What we are, what we showed, compared to NetMHCpan here is that. It is very likely that when you see new alleles, new peptides, those AlphaFold 3–based approaches are going to be more performant. That…still needs to be proved; but that is our strong intuition, at least also the inference is heavy because we have big proteins that we are predicting. It turns out that the main changing part when you want to run a new prediction is the peptides, which is the small part of the complex. And we can store predictions of the big proteins so that we don't need to rerun them every time we need to make a prediction. So there are ways to make this much more efficient, and really the additional cost is negligible compared to the benefit that we can get here.

Pharmacy Times: MHC-II presentation is central to CD4-positive T-cell help in both infection and cancer immunotherapy. Looking beyond binding affinity prediction, what additional biological steps between peptide MHC II binding and productive T cell activation does your model not yet capture, and how might those gaps affect downstream use cases?

Lahouel: So I would say there are 3 main steps, three main conditions that need to happen to have a good immune response in the way we think about the model, a good binding between MHC-II and peptides. That's what the model is directly tackling: a good response from the T cell. And so there is a third body that is the T cell that comes and needs to attack the MHC peptide complex. And the last step is the protein needs to be expressed. So the peptides need to be present. We already started with our colleague, John Alton at TGen, looking at the 3 bodies problem, meaning MHC peptide plus T cell binding. So that is a step that needs to be tackled, and the problem is more challenging. There you have a 3-bodies problem, as opposed to 2 bodies. Then there is the last step, which is taking into account protein expression and gene expression level. And some models, like Maria, which is one of the othe…models for this type of predictions, has the merit to take that into account. He was very smart to do it. But what we see is that we very significantly outperform Maria, at least on the first task of MHC peptide binding. So clearly there is a need to redo that part and insert the gene expression. The gene expression as a third step, using these models, one thing we're very interested in is, instead of predicting binding affinity, what if we can synthesize proteins given certain binding affinities, properties that are desirable, and that will be very significant for vaccine creation, drug development, etc. So it's clearly something that we're interested in for the future.