Proteasome-derived antimicrobial peptides discovered via deep learning.

Recent computational discoveries have identified numerous bioactive peptides within the human proteome, as well as across the broader tree of life, which were previously unrecognized for their roles in host immunity. These findings have led us to propose the "cross-talk hypothesis", suggesting that many molecules, such as proteins and peptides, traditionally viewed as extraneous to immune function may in fact actively contribute to immunity. Building on our earlier studies, which uncovered proteasome-derived peptides with putative antimicrobial activity in the human proteome, here we systematically interrogated the proteasome, a large protein complex responsible for degrading and recycling damaged or surplus proteins, for additional antimicrobial peptides. Using deep learning, we systematically mined ProteasomeDB, a curated repository of proteasomal cleavage and splicing events, to predict antibiotic activity against 11 clinically relevant pathogens. This deep learning approach uncovered 59 candidate peptides ("proteasomins") with a median minimum inhibitory concentration (MIC) of ≤64 μmol L