Separating selection from mutation in antibody language models

Reviewer #1 (Public review):

Summary:

Matsen et al. describe an approach for training an antibody language model that explicitly tries to remove effects of "neutral mutation" from the language model training task, e.g. learning the codon table, which they claim results in biased functional predictions. They do so by modeling empirical sequence-derived likelihoods through a combination of a "mutation" model and a "selection" model; the mutation model is a non-neural Thrifty model previously developed by the authors, and the selection model is a small Transformer that is trained via gradient descent. The sequence likelihoods themselves are obtained from analyzing parent-child relationships in natural SHM datasets. The authors validate their method on several standard benchmark datasets and demonstrate its favorable computational cost. They discuss how deep learning models explicitly designed to capture selection and not mutation, trained on parent-child pairs, could potentially apply to other domains such as viral evolution or protein evolution at large.

Strengths:

Overall, we think the idea behind this manuscript is really clever and shows promising empirical results. Two aspects of the study are conceptually interesting: the first is factorizing the training likelihood objective to learn properties that are not explained by simple neutral mutation rules, and the second is training not on self-supervised sequence statistics but on the differences between sequences along an antibody evolutionary trajectory. If this approach generalizes to other domains of life, it could offer a new paradigm for training sequence-to-fitness models that is less biased by phylogeny or other aspects of the underlying mutation process.

Weaknesses:

Some claims made in the paper are weakly or indirectly supported by the data. In particular, the claim that learning the codon table contributes to biased functional effect predictions may be true, but requires more justification. Additionally, the paper could benefit from additional benchmarking and comparison to enhanced versions of existing methods, such as AbLang plus a multi-hit correction. Further descriptions of model components and validation metrics could help make the manuscript more readable.

Reviewer #2 (Public review):

Summary:

Endowing protein language models with the ability to predict the function of antibodies would open a world of translational possibilities. However, antibody language models have yet to achieve breakthrough success, which large language models have achieved for the understanding and generation of natural language. This paper elegantly demonstrates how training objectives imported from natural language applications lead antibody language models astray on function prediction tasks. Training models to predict masked amino acids teaches models to exploit biases of nucleotide-level mutational processes, rather than protein biophysics. Taking the underlying biology of antibody diversification and selection seriously allows for disentangling these processes through what the authors call deep amino acid selection models. These models extend previous work by the authors (Matsen MBE 2025) by providing predictions not only for the selection strength at individual sites, but also for individual amino acid substitutions. This represents a practically important advance.

Strengths:

The paper is based on a deep conceptual insight, the existence of a multitude of biological processes that affect antibody maturation trajectories. The figures and writing a very clear, which should help make the broader field aware of this important but sometimes overlooked insight. The paper adds to a growing literature proposing biology-informed tweaks for training protein language models, and should thus be of interest to a wide readership interested in the application of machine learning to protein sequence understanding and design.

Weaknesses:

Proponents of the state-of-the-art protein language models might counter the claims of the paper by appealing to the ability of fine-tuning to deconvolve selection and mutation-related signatures in their high-dimensional representation spaces. Leaving the exercise of assessing this claim entirely to future work somewhat diminishes the heft of the (otherwise good!) argument. In the context of predicting antibody binding affinity, the modeling strategy only allows prediction of mutations that improve affinity on average, but not those which improve binding to specific epitopes.

Reviewer #3 (Public review):

Summary:

This work proposes DASM, a new transformer-based approach to learning the distribution of antibody sequences which outperforms current foundational models at the task of predicting mutation propensities under selected phenotypes, such as protein expression levels and target binding affinity. The key ingredient is the disentanglement, by construction, of selection-induced mutational effects and biases intrinsic to the somatic hypermutation process (which are embedded in a pre-trained model).

Strengths:

The approach is benchmarked on a variety of available datasets and for two different phenotypes (expression and binding affinity). The biologically informed logic for model construction implemented is compelling, and the advantage, in terms of mutational effects prediction, is clearly demonstrated via comparisons to state-of-the-art models.

Weaknesses:

The gain in interpretability is only mentioned but not really elaborated upon or leveraged for gaining insight. The following aspects could have been better documented: the hyperparametric search to establish the optimal model; the predictive performance of baseline approaches, to fully showcase the gain yielded by DASM.