Google DeepMind's October 15 walkthrough follows one proposed antimicrobial peptide from a database entry to a plan for laboratory tests. The researcher chose the biological question and directed each new check. An Antigravity agent searched and connected protein annotations, sequence and structural comparisons, genomic context, and published literature. Google says the candidate is undergoing wet lab validation. The account reports a computational lead, with no experimental result yet showing that the peptide kills bacteria.

For a biology team, the useful detail is the chain of evidence behind that lead. A predicted function can suggest where to look; a structural resemblance or neighboring gene can strengthen a case for testing. Each remains a different kind of clue, and none measures antimicrobial activity.

The big change

  • What changed: Google's walkthrough shows a researcher directing an agent to assemble sequence, structural and genomic evidence for one peptide candidate in a single workbench. The output is a testable hypothesis and an assay plan, with no antimicrobial activity result reported.
  • Why it matters: Biologists spend time connecting clues across separate resources before choosing what to test. An agent that assembles those clues can give research teams a clearer basis for prioritizing limited bench work, provided researchers check the underlying matches and annotations. Google does not report whether this workflow improves discovery rates.
  • What to watch: Google says the candidate is undergoing wet lab validation. Expression, susceptibility, target engagement and toxicity results will determine whether this particular dossier was worth advancing. For research managers, the wider decision is whether evidence assembled by agents holds up often enough to guide scarce experimental capacity. This one candidate cannot answer that yet.

From annotation to a candidate

Google says the starting task was to find and computationally assess novel antimicrobial peptides in Streptococcus pseudopneumoniae. The researcher guided an interactive search of UniProt entries annotated with ProtNLM, Google's sequence based protein function model. The agent narrowed that search to candidates including A0A0U0D591. Those entry details come from Google's account; BIG CHANGE could not independently retrieve the structured UniProt record during this review.

The researcher next asked for a BLAST search against manually reviewed Swiss-Prot proteins. Google says that search found no meaningful sequence homolog among them. That is a bounded database comparison, rather than proof that the molecule is new to biology. The agent then retrieved an AlphaFold predicted structure and used structural comparison to find a resemblance to an experimentally determined bacteriocin. Sequence and structural searches answer different questions: proteins can have weak sequence similarity while still resembling one another in three dimensions. Google's reported structural match supports its proposed family assignment, but a predicted fold does not establish the candidate's activity or target.

The genomic check added a third clue. Google reports that the candidate's gene lies six base pairs upstream of an ABC transporter permease and a possible partner ATP binding protein, with another annotated bacteriocin and an XRE regulator nearby. Such neighbors can support the interpretation that a peptide belongs to a bacteriocin system. The arrangement cannot by itself show that the gene is expressed, that the peptide is secreted, or that it inhibits a pathogen.

The experiment remains the decisive step

After reviewing literature, the agent proposed a mechanism involving bacterial cell division and a possible activity spectrum that includes drug resistant pathogens. Both are hypotheses drawn from other evidence, as Google presents them. The agent's suggested bench plan covers protein expression, antimicrobial susceptibility, target engagement and toxicity testing. Google says this candidate and others found through the pipeline are undergoing wet lab validation, without reporting results from those tests in the walkthrough.

That leaves several questions for experimental work: can researchers produce the peptide, does it inhibit any tested organism at measured concentrations, what target does it engage, and how does it affect cells in toxicity assays? A positive susceptibility result would still need to be interpreted alongside mechanism and safety evidence. Neither the proposed spectrum nor the plan is evidence of a working drug.

The workbench may be useful to teams that already know how to judge the underlying databases and choose decisive assays. Google's science access page says the Science plugin is installed from Antigravity's in-app Customizations marketplace and offers an application for 10,000 scientists to receive its highest usage tier free of charge. The screenshots in the peptide walkthrough depict Gemini 3.8 Flash; they do not establish that every supported model would reproduce the same analysis. BIG CHANGE did not run Antigravity or repeat the database searches, so the reported speed and candidate assessment remain Google's account rather than our measured result.

Sources & further reading