29 September 2025 in Nature Methods
9 August 2024 in Nature Methods
15 November 2024 in bioRxiv
A bstract: Protein language models (PLMs) have demonstrated remarkable success in protein modeling and design, yet their internal mechanisms for predicting structure and function remain poorly understood. Here we present a systematic approach to extract and analyze interpretable features from PLMs using sparse autoencoders (SAEs). By training SAEs on embeddings from the PLM ESM-2, we identify thousands of human-interpretable latent features per layer that highlight hundreds of known biological concepts such as binding sites, structural motifs, and functional domains. In contrast, examining individual neurons in ESM-2 reveals significantly less conceptual alignment, suggesting that PLMs represent most concepts in superposition. We further demonstrate that a larger PLM (ESM-2 with 650M parameters) captures substantially more interpretable concepts than a smaller PLM (ESM-2 with 8M parameters). Beyond capturing known annotations, we show that ESM-2 learns coherent concepts that do not map onto existing annotations and propose a pipeline using language models to automatically interpret novel latent features learned by the SAEs. As practical applications, we demonstrate how these latent features can fill in missing annotations in protein databases and enable targeted steering of protein sequence generation. Our results demonstrate that PLMs encode rich, interpretable representations of protein biology and we propose a systematic framework to uncover and understand these latent features. In the process, we recover both known biology and potentially new protein motifs. As community resources, we introduce InterPLM (interPLM.ai), an interactive visualization platform for investigating learned PLM features, and release code for training and analysis at github.com/ElanaPearl/interPLM.
28 July 2025 in Nature Genetics
Human leukocyte antigens (HLAs) are encoded by the most polymorphic genes in the human genome. HLA class I alleles control antigen presentation for T cell recognition, which is pivotal for autoimmunity, infectious diseases and cancer. Current knowledge of HLA-bound peptides is limited, skewed and falls short of population-wide HLA binding profiles for high-value targets. Here we present ESCAPE-seq (enhanced single-chain antigen presentation sequencing), a massively parallel platform for comprehensive screening of class I HLA–peptide combinations for antigen presentation via deep DNA sequencing. ESCAPE-seq demonstrates programmability, high throughput, sensitivity and nominated viral and cancer epitopes. We simultaneously assessed over 75,000 peptide–HLA combinations, revealing broadly presented epitopes from oncogenic driver mutations and fusions across diverse HLA-A, HLA-B and HLA-C alleles that cover 90% of the human population. We further identified epitopes that are differentially presented, comparing oncogenic hotspot mutations versus wild type. ESCAPE-seq enables one-shot population-wide antigen presentation discovery, offering insights into HLA specificity and immune recognition of genomic mutations.