Accelerating antibiotic discovery from Earth's isolated habitats through protein language models and microfluidics

Abstract

The antibiotic potential of Earth's isolated habitat microbiomes remains largely unexplored. Accessing this resource requires reliable prioritization across divergent peptide sequences and efficient experimental validation. Here we present MicroSPEAR, an end-to-end framework for antimicrobial peptide (AMP) discovery integrating SPEAR, a sequence-structure-aware ensemble predictor, with MicroGrad, a microfluidic testing platform. SPEAR generalized to low-homology peptides and prioritized potent candidates despite training on binary activity labels. MicroGrad simplified liquid handling and shortened susceptibility testing from approximately 18 h to 4 h. Applied to a catalogue derived from 4,480 metagenomes spanning 15 isolated habitat types, SPEAR identified 383,027 non-redundant candidate AMPs, 63.1% of which lacked detectable homologues in reference databases. Of 85 synthesized candidates, 79 exhibited antibacterial activity at [≤]64 M;. Mechanistic analyses revealed distinct modes of membrane disruption. Four broad-spectrum candidates with low toxicity significantly reduced bacterial burdens in a mouse thigh infection model. MicroSPEAR connects underexplored microbial diversity with experimentally validated antibiotic leads and provides a foundation for lab-in-the-loop discovery.