Drug-resistant infections could cause an estimated 10 million deaths a year by 2050, and the search for new antibiotics is slow, partly because promising molecules are buried in databases nobody has properly read. UniProt, the leading public protein resource, holds millions of sequences whose function is unknown. Google DeepMind has shown how biologists can use Google Antigravity, its agentic development platform, to dig through that material in one working session, and how the session produced a new antimicrobial peptide candidate that is now being tested in the lab.
The workflow pairs a researcher with an AI agent. The starting point is ProtNLM, a Google model that predicts what a protein does from its amino acid sequence and has added functional labels to millions of unnamed UniProt entries. In the walkthrough, the researcher sets a goal, finding new antimicrobial peptides in the bacterium Streptococcus pseudopneumoniae, and the agent filters millions of entries down to a handful. It settles on a 97-amino-acid protein from the Lactococcin 972 family, then checks its novelty with BLAST searches and its predicted 3D structure from the AlphaFold database, which closely matches a known bacteriocin. It inspects neighbouring genes and finds transport and regulatory genes that support the idea, reviews the literature to suggest a mechanism targeting bacterial cell division, and drafts a bench protocol covering expression, susceptibility testing and toxicity.
The value DeepMind claims is not a single molecule but time. What used to mean custom scripts, exported files, separate modelling programs and several experts becomes one collaborative session, with a shared 3D structure viewer that both the scientist and the agent can mark up. Because lab validation is expensive, narrowing the field to the strongest leads before any experiment is where the saving lands.
For readers following Antigravity, which we review as a coding platform, this shows Google extending it into an agentic science workbench. It also fits a pattern across labs: OpenAI's thousands of GPT-Rosalind agents mining a mouse brain atlas, Anthropic's guest research projects in biology. All make the same bet that agents can read the long tail of biological data faster than people.
The limits are explicit in DeepMind's own text. The candidate is a biologically plausible lead backed by structural and genomic evidence, nothing more; its activity against respiratory pathogens is predicted, not measured, and it is undergoing wet-lab validation along with several others from the same pipeline. The post gives no hit rates, timings or comparison with how long the same search takes without the agent. Until lab results arrive, this demonstrates a workflow rather than a discovery.