MPI-NAT Seminar: Accelerating antibiotic discovery using AI

For a century, antibiotic discovery has depended on labor-intensive “dirt mining”: scientists collected soil or water samples and painstakingly screening them for active compounds. That trial-and-error paradigm is too slow to keep pace with escalating antimicrobial resistance. Over the past decade, our laboratory has replaced it with digital discovery, using artificial-intelligence (AI) models to mine the world’s biological data sets. While AI already outperforms humans in image and text recognition, its impact on biology and medicine is only beginning to emerge. We pioneered the design of antibiotics using AI, producing molecules that show strong efficacy in preclinical animal models. By systematically mining the human proteome, we uncovered thousands of novel antimicrobial peptides that appear to form a distinct branch of peptide-based immunity, shaping host defense and broader physiological functions. Subsequently, we hypothesized that similar compounds could be found throughout evolution. By expanding our efforts to ancient biology, our AI-driven approach led to the discovery of therapeutic molecules from organisms such as Neanderthals and the woolly mammoth, a milestone that launched the field of molecular de-extinction and yielded preclinical candidates such as neanderthalin, mammuthusin, and elephasin. Determined to explore the full tree of life, we then turned to its other two major domains, Bacteria and Archaea. By computationally analyzing the global microbiome, we identified nearly one million new antibiotic molecules, all of which have been made freely available and open access to encourage worldwide collaboration. Through machine learning, this collaborative effort explored the vast diversity of the microbial world by analyzing many thousands of microbial metagenomes and genomes. Additionally, by examining thousands of human microbiomes, we and our collaborators discovered a myriad of new antimicrobial agents, including prevotellin-2 from the gut microbe Prevotella copri. Most recently, we have also digitally mined archaea, an underexplored domain of life, identifying a new class of antibiotics called archaeasins. Finally, I will introduce our latest AI models APEX, APEXGO, and APEXDUO, which respectively enable sequence-to-function prediction, computer-based antibiotic optimization, and the generation of multimodal therapeutics. Collectively, our efforts have dramatically accelerated antibiotic discovery, reducing the time required to identify preclinical candidates from years to just a few hours. I believe we are on the cusp of a new era in science where advances enabled by AI will help control antibiotic resistance, infectious disease outbreaks, future pandemics, and accelerate discoveries across biology.