Yarin Gal

This paper evaluates sampling methods for protein language models, showing that multi-mutant strategies outperform single-mutant approaches in capturing epistatic interactions for functional protein design.

ProteinGym unifies over 250 deep mutational scans and ~65k clinical variants to benchmark over 70 models in zero-shot and supervised settings; TranceptEVE ranks best overall.

This paper introduces ProteinNPT, a non-parametric transformer for protein property prediction and design, showing improved performance in label-scarce and multi-task settings.

Blending an alignment-based family model with an autoregressive protein language model, weighted by how deep the alignment is, reportedly improves mutation-effect prediction against experimental assays and ClinVar.

This paper introduces DyNeMo, a model for dynamic functional brain networks that outperforms HMM in capturing non-exclusive network modes in MEG data.

The paper predicts the pathogenicity of 11 million genetic variants across 1,081 disease genes using a deep learning model that leverages evolutionary data instead of clinical labels.