PhageFlare: deep-learning tool to annotate phage proteins that structure-similarity searches miss (bioRxiv preprint)

A Hong Kong Baptist University team (corresponding author Lu Zhang) posted the bioRxiv preprint "Illuminating phage dark matter beyond the reach of structural alignments with PhageFlare" on 29 October 2026, not yet peer reviewed. PhageFlare is a deep-learning framework (frozen ESM-C 600M protein language model, 45 functional labels, optional genomic context) for predicting the functions of phage proteins that have no significant structural match to annotated references. The authors report macro-averaged F1 of 0.871, 0.889 and 0.738 on three structure-disjoint benchmarks, beating the strongest baseline, and say it flagged lysogenic potential in 1,866 prophage-designated genomes that BACPHLIP had classed as lytic. Code is on GitHub (MIT licence) and the model checkpoint is on Hugging Face.