#34 · ML / AI methods

Protein language models and LLMs predict mutation, editability and PAM specificity

The dominant thread is evaluating learned sequence and structure models against defined biological targets, often without task-specific training. VenusRAR ranks and audits protein language model zero-shot mutation predictions under biophysical constraints, reporting ProteinGym Spearman 0.551. DETANGO decomposes protein language model mutation effects into stability and functional components to flag stable-but-inactive variants. Protein2PAM predicts Cas PAM specificity from sequence and guides computational evolution. Beyond protein sequence, an open-weights LLM benchmark on three CRISPR screens reaches 0.715 AUROC for zero-shot synthetic lethal prediction, and AdarEdit applies structure-explicit graph attention to A-to-I RNA editability across GTEx tissues. What is new is the explicit structuring of objectives: biophysical constraints, separated stability and function terms, and zero-shot reasoning over screen data.