💻 bioinformatics

PepCL: A replay-based continual learning framework for updating peptide-MHC models

The paper introduces PepCL, a replay-based continual learning framework paired with a new state-of-the-art model called MHCPrime, which enables peptide-MHC predictors to integrate new experimental assay data while preserving prior mass spectrometry knowledge to overcome catastrophic forgetting and improve predictive performance across diverse biological contexts.

Original authors: Chati, P. M., Lashkari, V. D., Salhotra, A., Bruno, P. M., Ntranos, V.

Published 2026-07-21

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Chati, P. M., Lashkari, V. D., Salhotra, A., Bruno, P. M., Ntranos, V.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your immune system as a highly trained security force patrolling the body's borders. Its job is to spot intruders—viruses, bacteria, or cancer cells—and sound the alarm. To do this, the security guards (T cells) need to see the intruders' faces. But the guards don't see the intruders directly; they see tiny fragments of them, called peptides, displayed on a special billboard on the surface of every cell. This billboard is called the Major Histocompatibility Complex, or MHC.

The problem is that there are billions of possible peptide fragments, and the billboards come in thousands of different shapes (alleles), each with its own unique taste in what it likes to display. It's like trying to predict which specific puzzle piece fits into which specific slot in a giant, ever-changing jigsaw puzzle. Scientists have been building computer programs to guess these fits, but they've mostly been trained on one specific type of data: mass spectrometry. Think of mass spectrometry as a high-tech camera that takes pictures of the peptides currently hanging on the billboards inside a cell. It's great, but it has blind spots. It misses some types of pieces, and it can't easily see the ones that don't fit. Meanwhile, new, faster experiments are popping up that can test millions of pieces at once, but they see the puzzle from a slightly different angle. The big question is: How do we teach our computer programs to learn from these new, faster experiments without forgetting everything they already learned from the high-tech camera?

Enter PepCL, a clever new framework introduced by Chati and colleagues that acts like a "do-over" button for these computer models. The researchers realized that simply retraining a model on new data is like trying to learn a new language by only speaking that language; you might get good at it, but you'll forget your native tongue. Instead, they developed a method called continual learning. Imagine a student studying for a math test. Usually, if they start studying for a history test, they might forget the math formulas. But with PepCL, the student is given a "cheat sheet" of the old math problems (replay data) while studying history. Every time they answer a history question, they also glance at the math cheat sheet to make sure they haven't forgotten the old rules.

The team built a new, super-flexible model called MHCPrime to serve as the student. They then used PepCL to update this model with data from various new experiments, including yeast display and EpiScan. The results were impressive: the model learned the new, specific details of these experiments (like how to spot tricky, hydrophobic peptides that the old camera missed) but didn't lose its ability to predict what happens inside a real human cell. In fact, when they tried to update the model using the old, standard method (called fine-tuning), the model got "confused," becoming great at the new experiment but terrible at the old one—a phenomenon known as "catastrophic forgetting." PepCL, however, kept the balance, suggesting that it's possible to upgrade our immune system's "face recognition" software with new data without wiping the hard drive clean. This approach could help scientists design better vaccines and cancer treatments by making these prediction tools smarter and more adaptable as new scientific tools are invented.

Technical Summary: PepCL – A Replay-Based Continual Learning Framework for Peptide-MHC Models

Problem Statement
Predicting peptide-major histocompatibility complex (MHC) class I binding is critical for vaccine design and immunotherapy but faces a combinatorial challenge due to over 40,000 MHC alleles. While mass spectrometry (MS) has provided the largest dataset for training predictors, MS-derived ligandomes are incomplete and biased, underrepresenting specific sequence classes (e.g., cysteine-containing or highly hydrophobic peptides) and failing to capture non-binders. Emerging experimental assays (e.g., yeast display, EpiScan, binding affinity, ESCAPE-seq) offer complementary data, including experimentally validated non-binders and continuous readouts. However, existing peptide-MHC predictors are typically distributed as fixed, versioned models. Incorporating these new, orthogonal data sources usually requires full retraining or conventional fine-tuning (FT). FT often leads to "catastrophic forgetting," where the model over-specializes to the new assay's specific biases and distribution, overwriting the valuable, biologically grounded knowledge of endogenous peptide presentation learned from MS data.

Methodology
To address the trade-off between adapting to new data and retaining prior knowledge, the authors introduce PepCL (Peptide-MHC Continual Learning), a replay-based continual learning framework.

MHCPrime Backbone: The authors developed MHCPrime, a new state-of-the-art, pan-allelic, transformer-based model trained on MS eluted ligand (EL) data. Unlike closed-source ensembled models (e.g., HLApollo), MHCPrime is open-source, GPU-optimized, and designed with direct access to parameters, logits, and gradients to facilitate continual learning. It utilizes a dual-encoder architecture (peptide and MHC pseudosequence) incorporating biochemical features and local context via Conformer and Transformer blocks.
PepCL Framework: PepCL updates the pretrained MHCPrime model using a combination of three mechanisms:
- New Data Update: The model is trained on incoming assay data (e.g., yeast display, EpiScan).
- Data Replay: Batches of the original MS training data are interleaved during the update to provide gradients that counterbalance the new data distribution.
- Drift Regularization: A frozen copy of the pretrained model is maintained. The updated model is penalized (via Mean Squared Error on logits) if its predictions on the replay data drift significantly from the frozen model's predictions. This penalty is controlled by a hyperparameter $\lambda$ .
Training Objective: The framework combines classification losses (using a differentiable Smooth-AP loss for ranking) from both new and replay data with the drift regularization loss. For assays with continuous readouts (e.g., Binding Affinity, ESCAPE-seq), a differentiable soft rank loss is used for the new data while maintaining the classification loss for the replay set.

Key Results
The authors evaluated PepCL and MHCPrime across diverse biological contexts, including infectious disease and cancer, using multiple assay modalities.

Base Model Performance: MHCPrime achieved state-of-the-art performance on held-out MS benchmarks, matching or exceeding closed-source ensembled models like HLApollo while being a single, non-ensembled model. However, like other MS-trained models, it underperformed on non-MS datasets.
Mitigation of Catastrophic Forgetting: In controlled experiments where models were updated with new MS subsets or specific motif clusters, conventional FT caused severe degradation in performance on the original data (forgetting). PepCL successfully incorporated the new information while preserving performance on the prior MS distribution.
Assay Adaptation: When updated with non-MS data (Yeast Display, EpiScan, Binding Affinity, ESCAPE-seq):
- FT improved performance on the target assay but caused disproportionate losses on MS benchmarks and related domains.
- PepCL achieved comparable gains on the new assays while retaining or even improving performance on MS benchmarks.
- PepCL demonstrated superior ability to learn from "hard negatives" (e.g., in EpiScan retraining libraries) and continuous readouts without overfitting to assay-specific artifacts.
Cross-Allele Transfer: A significant finding was that PepCL enabled the transfer of assay-derived signals to alleles not present in the update set. By leveraging the pan-allelic representations of MHCPrime, the framework improved prediction for unseen alleles, whereas FT often degraded performance on the prior MS domain for these alleles.
Conservative Adaptation: Analysis of sequence motifs revealed that while FT fully absorbed assay-specific biases (e.g., relaxed C-terminal constraints in Yeast Display due to specific mutations), PepCL adopted these signals more conservatively, remaining anchored to the biologically relevant MS-derived binding preferences.

Significance and Claims
The paper claims that PepCL establishes a flexible framework for extending the utility of peptide-MHC models as immunopeptidomics assays evolve. The authors argue that:

Continual Learning is Necessary: The field must shift from static, one-time training to updatable frameworks that can integrate new experimental evidence without discarding prior knowledge.
Balanced Adaptation: PepCL provides a controlled mechanism to absorb complementary information from diverse assays (which often have different biases and endpoints) while preventing the overspecialization and catastrophic forgetting associated with conventional fine-tuning.
Clinical Relevance: By preserving the MS prior (which reflects endogenous cellular presentation) while integrating new signals, PepCL offers a more robust path for clinical applications like neoantigen prioritization and vaccine design, where accurate prediction of natural presentation is paramount.
Scalability: The framework allows for the propagation of limited-allele assay data across the broader pan-allelic space, addressing the challenge of sparse data for many clinically relevant alleles.

The authors conclude that progress in peptide-MHC prediction will increasingly depend not just on model architecture, but on frameworks like PepCL that provide a practical path for expanding predictor coverage as new datasets emerge.

Technical Summary: PepCL – A Replay-Based Continual Learning Framework for Peptide-MHC Models

More like this