Google DeepMind researchers have demonstrated an agentic development platform named Google Antigravity, using it to identify a novel antimicrobial peptide candidate with predicted activity against respiratory tract infections. The candidate molecule emerged from a computational search in Streptococcus pseudopneumoniae and is now undergoing wet-lab validation.
Antimicrobial resistance threatens to cause an estimated 10 million deaths annually across the world by 2050. To counter this risk, the project focused on bacteriocins, natural antimicrobial peptides that target fundamental structural components such as bacterial cell envelopes and essential intracellular machinery. Natural sequence databases have previously yielded therapeutics such as Nisin, Darobactin, Dynobactin, and Lariocidin, yet millions of sequenced proteins remain uncharacterized in public archives like UniProt.
ProtNLM, a deep learning model developed by Google, reads raw amino acid sequences to predict biological functions based on evolutionary patterns. Antigravity integrates this model to parse unannotated UniProt sequences during active research sessions, replacing fragmented workflows that previously required custom query scripts and separate genomic tools.
Engineers embedded a 3D visualization tool called the Biomolecular Structure Viewer inside the workbench. The viewer creates a shared visual state between the human researcher and the artificial intelligence agent, retrieving structures from the AlphaFold Database or user predictions. When a scientist selects residues or notes features on a binding pocket, the agent updates its structural reasoning simultaneously.
The demonstration was authored by Pengfei Tian, Maxwell L. Bileschi, Zoi Tsangalidou, Sukhdeep Singh, and Lucy J. Colwell, with feedback from EMBL-EBI researcher Stephanie Lo. DeepMind reported that scientists can broaden this search method beyond human commensals to discover peptide candidates in environmental microbes.
