Julius Schlensok

This paper argues for evaluating protein language models through prospective experimental success rather than retrospective predictive accuracy.

A fine-tuned ProtT5 protein language model predicts per-residue energetic frustration from sequence (Pearson r 0.86 on unseen CATH topologies) and profiles the human proteome in ~17 minutes on one GPU.

UdonPred predicts per-residue protein disorder from sequence using embeddings from ProstT5, showing that annotation quality, not model capacity, drives performance in continuous disorder prediction.

VespaG is a fast missense variant effect predictor using protein Language Models, achieving high accuracy on the ProteinGym benchmark.