DeepMind details Antigravity search for antimicrobial peptide candidate

Google DeepMind says its Antigravity platform helped prioritize an antimicrobial peptide candidate from protein databases. The molecule is still undergoing wet-lab validation.

The technical walkthrough describes a human-agent workflow combining database filtering, protein-sequence searches, predicted-structure comparisons, genomic-context checks and literature synthesis in one session. The result was a candidate for experiments—not evidence of antimicrobial activity or a validated treatment. Related protein research includes DeepMind’s SynthID Bio work on AI-designed proteins.

A researcher set the goal of finding and computationally validating novel antimicrobial peptides in Streptococcus pseudopneumoniae, then chose which questions to pursue and what evidence to collect. According to DeepMind, the agent queried databases, analyzed structures and synthesized literature as the researcher steered the process.

DeepMind said the search yielded a small group of candidates, including UniProt entry A0A0U0D591, a 97-amino-acid protein assigned to the Lactococcin 972 family. The company reported that a search of the manually reviewed Swiss-Prot collection found no meaningful sequence homolog. It did not publish the underlying search output, significance thresholds or full ranking criteria.

That distinction between database annotations and laboratory evidence matters. The UniProt Consortium’s 2025 database paper says expert biocurators review Swiss-Prot records, while automated systems computationally annotate TrEMBL entries. ProtNLM, the model used in the workflow, predicts protein functions directly from amino-acid sequences and supplies annotations for previously uncharacterized entries.

DeepMind said the agent then retrieved an AlphaFold Database model for the candidate and compared it with experimentally determined structures. According to the company, the predicted fold closely matched an experimentally determined bacteriocin. The walkthrough did not identify that structure or provide alignment metrics. The AlphaFold Database disclaimer cautions that its structures are predictions with varying confidence and should be interpreted carefully.

The agent also examined neighboring genes. DeepMind said the candidate sits six base pairs upstream of an ABC transporter permease, near a possible ATP-binding partner, another annotated bacteriocin and an XRE transcriptional regulator. The company treated that genomic neighborhood as support for the hypothesis that the protein belongs to a bacteriocin system.

A literature review generated hypotheses about a cell-division mechanism and possible activity against drug-resistant pathogens. Those properties were predicted, not experimentally demonstrated for the candidate. The agent also produced a proposed protocol covering protein expression, antimicrobial susceptibility testing, target engagement and in vitro toxicity.

DeepMind said A0A0U0D591 and several other candidates from the pipeline are now in wet-lab validation. The company disclosed no assay results, activity measurements, toxicity data or target-engagement findings, and did not identify the laboratory or timing for the experiments.