To summarise, PLM-interact extends single-protein PLMs to jointly encode interacting partners. It achieves state-of-the-art performance in cross-species and virus–host PPI prediction tasks and can be fine-tuned to predict mutation effects in human PPIs.

PLM-interact was applied to virus–human PPI prediction. The model outperforms existing approaches, achieving 5.7%, 10.9%, and 11.9% gains in AUPR, F1, and MCC, respectively — effectively capturing virus–host interactions at the protein level.

We further demonstrate examples where PLM-interact correctly predicts the effects of mutations on PPIs associated with human diseases. These results highlight its potential to identify whether disease-associated mutations weaken or strengthen protein interactions.

We fine-tuned PLM-interact to predict the effects of mutations on protein interactions — identifying whether mutations increase or decrease interaction strength. The fine-tuned model significantly outperforms zero-shot PPI models in the mutation-effect prediction task.

PLM-interact achieves state-of-the-art performance on a widely adopted cross-species PPI prediction benchmark — trained on human data and tested on mouse, fly, worm, yeast, and E. coli.

Existing PPI models use pre-trained PLMs to embed each protein separately, ignoring amino acid interactions between proteins. PLM-interact goes beyond single-protein encoding by jointly representing protein pairs to learn their relationships.

Protein language models trained on massive protein sequence datasets capture evolutionary, sequence and structural features — becoming the method of choice for representing proteins in state-of-the-art PPI predictors.

This work is a part of my viroinf PhD research, carried out under the supervision of Craig Macdonald, @davidlrobertson.bsky.social, and Ke Yuan, with HPC support from DiRAC (www.dirac.ac.uk).

🌊 Coronaviruses likely acquired their spike proteins from aquatic herpesviruses. 4/5

🔍 Searches against structural network are extremely sensitive and allow to recover more RdRps than iterative profile-based searches. 3/5

🦠 The diversity of viral proteins can be reduced to 19,000 structural clusters. A structure-similarity network captures relationships between them. 2/5