{
  "id": 5391119,
  "title": "Towards Sparse Causal Features for Zero-shot Mutation Effect Prediction in a Protein Language Model",
  "url": "https://urgent.news/2026/09/03/towards-sparse-causal-features-for-zero-shot-mutation-effect",
  "topic": "science",
  "section": "Science",
  "published": "2026-09-03T00:00:00.000Z",
  "source": {
    "name": "bioRxiv",
    "slug": "biorxiv",
    "url": "https://www.biorxiv.org/content/10.64898/2026.08.28.747907v1?rss=1"
  },
  "original_language": "en",
  "account": "Protein language models, like ESM-2, demonstrate impressive ability to predict the effects of mutations in proteins, but the underlying mechanisms driving these predictions are not well understood. A new study introduces a sparse feature circuit framework that aims to uncover the latent features responsible for the model's zero-shot mutation effect prediction in ESM-2 650M. The researchers evaluated their framework on 67 mutations, ranging from highly deleterious to mildly deleterious, within the DNAJA1 J-domain. The results showed that ESM-2's predictions were strongly aligned with measurements from deep mutational scanning. The study found that circuits identified through indirect effect analysis were more effective in reproducing the model's predictions and offering meaningful biological explanations compared to circuits derived from raw activation changes. This suggests that activation magnitude alone does not necessarily correspond to causal significance. The research also revealed that similar substitutions often share significant portions, ranging from 40% to 75%, of their respective feature circuits. Additionally, the shared features frequently involve residues that are in close proximity in three-dimensional space to the mutation site. To the best of the researchers' knowledge, this work represents the first causal, feature-level explanation for zero-shot mutation effect prediction in a protein language model.",
  "summary": "Protein language models (pLMs) such as ESM-2 achieve strong zero-shot mutation-effect prediction, yet the internal computations supporting these predictions remain poorly understood. We introduce a sparse feature circuit framework that combines sparse autoencoders, integrated-gradients attribution, and activation patching to identify the latent features that causally mediate zero-shot mutation…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}