protein-language-models

PFArena introduces a benchmark for evaluating language models on protein modification tasks, comparing protein language models and large language models to assess their capabilities in biological applications.

The paper identifies a failure mode called off-manifold collapse in guided protein language models, where generated sequences degenerate, and proposes Mahalanobis filtering as a training-free post-hoc method to improve sequence quality and structural plausibility.

Research indicates that discarding singleton proteins in metagenomic datasets during the training of protein language models is likely incorrect, based on findings from a collaborative study.

Goodfire's Silico research platform is now publicly available, and a demo shows how protein language models like ESM-C represent structural information from sequence alone, enabling frontier-scale model interpretation and training.

This paper proposes a two-stage fine-tuning pipeline combining domain-adaptive fine-tuning and reinforcement learning to generate protein sequences that match a desired amino-acid composition profile while maintaining sequence quality.

Pepti-drift is a toxicity-aware latent refinement framework for generating antigen-specific peptides that avoids toxicity while maintaining binding affinity. It achieves large speedups over existing methods and produces diverse, valid, and low-toxicity peptide candidates.

ProtoCol applies late-interaction retrieval to protein homology search, representing proteins as sets of residue embeddings and using MaxSim for scoring, outperforming pooled and alignment-based methods on remote homology benchmarks.

This paper proposes SoftBlobGIN, a framework that enhances the interpretability of protein language model representations by projecting them onto contact graphs for structure-aware message passing. It demonstrates improved performance on enzyme classification and binding-site detection while providing auditable structural explanations.

This paper introduces a discrete diffusion model with a novel 'germline absorbing' modification to improve conditional antibody sequence generation. It addresses germline bias in protein language models and demonstrates superior performance in optimizing antibody binding affinity and developability compared to existing methods like EvoProtGrad.