Skip to content

S2 — Biological Molecular Function

Full session title: Biological Molecular Function: Methods and Benchmarks for Finding Function in Biological Dark Matter Papers: 7 (pp. 425–519) Session introduction: McDermott J.E. et al., pp. 417–424 · PDF · PMID 41758157

Organizers:

  • Jason E. McDermott — Computational Biology Group, Pacific Northwest National Laboratory; Molecular Microbiology and Immunology, Oregon Health & Science University
  • Yana Bromberg — Computer Science and Biology, Emory University
  • Hannah Carter — Genomics and Precision Medicine, UC San Diego
  • Travis Wheeler — College of Pharmacy, University of Arizona

Invited talk: Martin Steinegger (Sun Jan 4, 3:40 p.m.) — affiliation not printed in the schedule.

Reading this session

"Biological dark matter" means sequence and structure whose function we cannot assign. The seven papers split cleanly into structure-and-design methods and LLM-based function assignment, and the second group is the one that makes this session interesting for the corpus.

Protein structure, surfaces, and design. Scott et al. represent protein surfaces and interfaces implicitly and differentiably — a representation-learning contribution rather than a predictor. Viggiano et al. steer protein generative models at test time to guide AAV2 capsid design, which is directly relevant to the gene-therapy delivery problem. Ahmed et al.'s HALO is a hybrid-attention subcellular localization model.

LLMs assigned to function prediction — three independent attempts. Shringarpure et al. (23andMe) use LLMs to identify causal genes in complex-trait GWAS; Wang Z. et al.'s Gene-R1 applies data-augmented lightweight reasoning LLMs to gene-set analysis; Zhapa-Camacho et al. build an LLM agent for protein function prediction. That three groups arrived at the same session with the same instinct is itself the finding: function annotation is being reframed as a language problem, and this session is where that got argued in public.

Perturbation modeling. Chang et al.'s PertSpectra is the session's — and arguably the meeting's — most corpus-relevant paper: interpretable guided matrix factorization for predicting the functional impact of genetic perturbation experiments, evaluated on three single-cell RNA-seq Perturb-seq-style datasets with both single and combinatorial perturbations, using a gene–gene interaction graph as prior.

Papers

Title First author Pages PDF PMID
Implicitly and Differentiably Representing Protein Surfaces and Interfaces Cory B. Scott 425–437 scott.pdf 41758158
Steering Protein Generative Models at Test-Time for Guided AAV2 Capsid Design Ben Viggiano 438–451 viggiano.pdf 41758159
HALO: Hybrid Attention Model for Subcellular Localization Shafayat Ahmed 452–464 ahmed.pdf 41758160
PertSpectra: Interpretable Matrix Factorization for Predicting Functional Impact of Genetic Perturbation Experiments Seowon Chang 465–479 chang_s.pdf 41758161
Large Language Models Identify Causal Genes in Complex Trait GWAS Suyash S. Shringarpure 480–493 shringarpure.pdf 41758162
Gene-R1: Reasoning with Data-Augmented Lightweight LLMs for Gene Set Analysis Zhizheng Wang 494–507 wang_z.pdf 41758163
LLM Agent Based Protein Function Prediction Fernando Zhapa-Camacho 508–519 zhapa-camacho.pdf 41758164

Connections to the corpus

Sources