abstract:
Antimicrobial resistance threatens to become a leading cause of death by 2050,
and antimicrobial peptides (AMPs) offer a promising alternative because bacteria
acquire resistance to them more slowly than to conventional antibiotics.
Deep generative models design AMP candidates at scale, but translating this into
clinical leads requires three capabilities: controllable generation toward
specified properties, evaluation of computational predictors against biological
activity rather than database membership, and extensive wet-lab validation. This
talk presents three contributions addressing these capabilities: HydrAMP,
BATTLE-AMP, and OmegAMP. HydrAMP is a conditional variational autoencoder that
disentangles peptide sequence from antimicrobial conditions and supports
unconstrained and analog generation with parameter-controlled creativity.
In experiments across five bacterial strains, nine analogs of clinically relevant
active prototypes and six analogs of an inactive prototype showed high
antimicrobial activity. BATTLE-AMP benchmarks AMP predictors against experimentally
measured minimum inhibitory concentrations (MICs) rather than database membership.
We surveyed 48 methods, found fewer than 25% reproducible, and evaluated 21 model
variants. MIC-trained models consistently outperform binary classifiers; most rely
on amino acid composition rather than residue order; and activity cliffs remain
unresolved by both machine learning and molecular dynamics. OmegAMP extends
controllable generation through a diffusion model with biologically informed
encoding, combining three generation modes (de novo, analog, motif-guided) with
physicochemical and species-specific conditioning. In experiments across
20 bacterial strains including multidrug-resistant clinical isolates,
86% of 95 de novo peptides achieved MIC $\leq 2 \mu M$, 93% of analogs
converted inactive prototypes into active antimicrobials, and motif-guided
generation introduced antimicrobial activity while preserving LPS- or DNA-binding
function.