Google's AI agent found an antimicrobial peptide candidate. Nobody has tested it in a lab yet.

On October 15, 2026, Google DeepMind published a technical walkthrough in which a biologist and its Antigravity agent flagged a possible new antimicrobial peptide from a bacterium called Streptococcus pseudopneumoniae. Everything in the post is computational. The authors say the candidate still needs wet-lab validation, and no potency, toxicity or activity data are reported.

What the agent did

The walkthrough describes a multi-turn session between a biologist and Antigravity, Google's agentic development platform. Per the post, the agent:

• Filtered UniProt entries and flagged a candidate bacteriocin, UniProt ID A0A0U0D591, from the Lactococcin 972 family
• Ran a BLAST search against Swiss-Prot, the manually reviewed UniProt subset, and found no meaningful sequence homolog
• Analyzed the AlphaFold-predicted structure with Foldseek and matched the fold to an experimentally determined bacteriocin
• Checked the genomic context, finding the candidate near an ABC transporter, a partner ATP-binding protein, another bacteriocin and an XRE transcriptional regulator
• Reviewed literature to propose a mechanism, targeting bacterial cell division, and a range of drug-resistant pathogens
• Drafted a wet-lab plan covering protein expression, susceptibility testing, target engagement and in vitro toxicity

The authors also introduced the Biomolecular Structure Viewer, a 3D tool embedded in Antigravity. The post gives antimicrobial resistance as the motivation, citing an estimated 10 million deaths per year by 2050.

What the database entry says

I checked the candidate in UniProt. The entry A0A0U0D591 is for Streptococcus pseudopneumoniae strain Spain2270. It is 97 amino acids long, with a molecular weight of 10,408 daltons. It is unreviewed (TrEMBL), meaning it is not manually curated, and its annotation score is 1.0. Protein existence is listed as "predicted." The protein name, Lactococcin 972 family bacteriocin, was assigned by ProtNLM, an automated model. The entry carries no function comment, and it comes from a preliminary whole-genome shotgun sequence. A signal peptide at residues 1 to 28 is predicted, and the mature chain would span residues 29 to 97.

That is not a criticism of the work. It is what a screening starting point looks like. Unreviewed entries are the bulk of the database, and mining them is the point of the exercise. But it means the chain of evidence is predicted function, then predicted structure, then predicted mechanism. Each step leans on the previous one. The agent's mechanism and pathogen spectrum, the post says, come from the literature, not from measurement.

Why the careful wording matters

The authors are measured. They call the result "a biologically plausible lead," and say the walkthrough covers one narrow sequence window in one species. That honesty is useful, because headlines about AI and drug discovery tend to jump from "flagged" to "found."

Compare it with what Google says elsewhere. The Gemini 4 Argon announcement from October 13 says the model produced five novel serine esterase designs with confirmed catalytic activity in lab tests. That is a different claim: designs made, then measured. The peptide post sits a rung lower on the evidence ladder. It is a hypothesis generated quickly, which has real value, but it is not yet a finding.

There is also a dual-use question that Google's own safety write-up for Argon acknowledges, since it reports elevated capabilities in the biological domain and says external red teamers unlocked under 1% of tested bio sub-tasks. A tool that triages protein function at speed serves defenders and, in principle, others. We looked at the wider debate in our piece on Claude's gene-editing discovery and lab oversight.

What would count as a result

For a reader trying to judge whether this becomes news later, the markers are concrete:

• A measured minimum inhibitory concentration against named pathogens
• Toxicity against human cells in vitro
• Whether the peptide is actually secreted and active, since a predicted bacteriocin may not kill anything
• Independent replication outside Google

The post says the candidate and several others from the same pipeline are in wet-lab validation, with results pending. No date is given. Peptide candidates often fail at the first assay, which is why the base rate matters more than the demo.

What it says about the tool

The more durable claim here is about workflow, not the molecule. An agent that chains database filters, sequence search, structure comparison and literature review in one session can compress days of triage. That is plausible and checkable, because the steps are standard tools: UniProt, BLAST, AlphaFold and Foldseek, plus thanks to EMBL-EBI for the BLAST service. A skeptic can rerun the steps.

What cannot be judged from the post is hit rate. One walkthrough on one window says nothing about how many candidates the pipeline surfaces, how many are false leads, or how the agent behaves on harder questions.

What this does not change about drug discovery

Finding a plausible candidate is the cheap part of antibiotic development. The expensive parts come after: showing the molecule kills bacteria at a safe dose, that it survives in the body, that it can be manufactured, and that it works in people. Most candidates fail at one of those steps. A faster triage step still helps, because it widens the funnel at low cost, but the walkthrough covers only the very top of it. A reader should therefore hold two thoughts at once: the workflow is genuinely useful, and the candidate is one of many unproven leads.

Our take

This is a competent demo with honest caveats and no result yet. Credit the authors for labeling it plausible rather than proven. The sensible stance is to watch for the wet-lab data. If the peptide shows activity at a useful concentration and low toxicity, then the story is that an agent shortened discovery. If it fails, the story is still that triage got cheaper, but the claim to have found a lead goes away.