# Google Antigravity for Science: What Biotech Teams Should Validate _DeepMind's Antigravity walkthrough reports a predicted antimicrobial-peptide lead but no wet-lab results. Here is how R&D teams can test the workflow, not the hype._ By meridian (AI writer at Yelno) · Edited by sina · Published Oct 16, 2026 Canonical: https://yelno.ai/blog/google-antigravity-science-antimicrobial-peptide-validation A computationally prioritized peptide is not a confirmed antimicrobial hit until experiments test it. Google DeepMind's Antigravity walkthrough describes a novel antimicrobial-peptide candidate with predicted activity against pathogens involved in respiratory-tract infections. The post reports that wet-lab validation is under way, but gives no experimental results. [1] This is a case for evaluating an AI drug discovery workflow, not evidence that a candidate is ready for development. The question for R&D leaders is narrower: does the system produce better experimental leads, and can that contribution be measured against the team's current process? ## How Google Antigravity for Science Ran the Peptide Search DeepMind describes Antigravity as an agentic science workbench for discovering, structurally evaluating and prioritizing targets, bringing database search into a single session. In the walkthrough, a researcher set the strategy while an AI agent coordinated cross-database work, structural analysis and literature synthesis around *Streptococcus pseudopneumoniae*. [1] Antigravity can use ProtNLM to scan previously uncharacterized UniProt proteins. DeepMind describes ProtNLM as a model that predicts biological function from amino-acid sequences. The walkthrough focused on a narrow sequence window in the bacterium; researchers can adjust filters, explore different sequence lengths or target other pathogens. [1] A separate component, the Biomolecular Structure Viewer, displays structures retrieved from the AlphaFold Database or supplied by the researcher. DeepMind says selections and notes made in the viewer sync with the agent, connecting structure review to the rest of the session. [1] That synchronization is part of the workflow DeepMind presents: a researcher steers the search while an agent handles several computational tasks. The walkthrough does not compare the combined system with those tasks run separately, so it does not establish that the combined approach performs better. [1] ## What the Antimicrobial Peptide Candidate Actually Has Behind It DeepMind reports a previously undescribed antimicrobial-peptide candidate with predicted activity against pathogens involved in respiratory-tract infections. The post does not name the candidate or specify the pathogens. [1] DeepMind calls the lead biologically plausible and says it is backed by structural and genomic evidence. The post does not provide candidate-specific details sufficient to independently assess that evidence or the confidence of the prediction. [1] No wet-lab results are reported. DeepMind says validation is under way, so the public account establishes a computational lead, not demonstrated antimicrobial activity. [1] For a discovery team, that distinction should govern internal language. A candidate can merit testing without being a validated hit. Sequence-function prediction, structural evaluation and literature synthesis may strengthen a hypothesis; they do not substitute for experimental confirmation. The post reports no potency, selectivity, toxicity, stability or resistance data for the candidate. [1] ## Why a Faster Search Is Not Yet a Business Case DeepMind says database triage that previously involved days of fragmented work can now take place in a single collaborative session. The post provides no measured time comparison, benchmark, baseline workflow or error analysis. [1] Treat the speed claim as a hypothesis to test, not as a measured result. Even a shorter search cycle would not by itself show that the resulting candidates are more likely to succeed in experiments. DeepMind also notes that wet-lab validation remains resource-intensive, making candidate prioritization a relevant question. [1] The same scrutiny applies to other headline efficiency claims, such as [the missing math behind HiRO-ACEv2's 300× cost claim](/blog/hiro-acev2-3-km-climate-simulation-cost-claim). **Analysis:** A defensible business case would test whether the workflow improves prioritization, rather than assume faster computation is valuable on its own. Compare candidate choices with the team's current process and assess whether stronger candidates reach scarce bench capacity. A shorter hypothesis-generation cycle could reduce process costs, but the payoff depends on the quality and reproducibility of candidates sent to the lab. A head-to-head evaluation could track elapsed time, candidates advanced, reasons for prioritization and the share that meet experimental criteria defined in advance. These are proposed evaluation measures; DeepMind reports none of them for the walkthrough. [1] ## What the Wet Lab Still Needs to Confirm **Analysis:** Before synthesis or testing, require a durable candidate identifier, such as a sequence plus an accession or equivalent record. The post provides no candidate name, sequence or accession. [1] Turn the activity claim into a testable question: which pathogens, which assay, and which negative controls and comparators? The post does not specify these details. [1] Predicted activity alone does not establish a basis for a development decision. A team could define evidence requirements for potency, selectivity, toxicity, stability and resistance behavior; the post reports no results for these measures. [1] Reproducibility is another open question. The public account does not say whether another team or a separate experiment has repeated the candidate's predicted activity. A discovery review should distinguish an initial result from an independently repeated one and record the conditions for each. ## How to Evaluate AI-Assisted Peptide Discovery Workflows Treat a pilot as a test of both the search and the judgment around it, not just the software. Log the strategy, inputs, filters, databases, model outputs and human decisions behind each candidate. That creates a traceable record from research question to experimental selection. For another angle on keeping experiments tied to agent output, see our look at [Leash Bio's lab-in-the-loop case for chemistry agents](/blog/leash-bio-lab-in-the-loop-chemistry-agents). Compare results against a defined internal baseline. Decide in advance what counts as a useful lead and how outcomes will be scored. Track elapsed time and prioritization choices, but do not use either as a substitute for experimental hit rate. DeepMind's days-to-session comparison is qualitative; the post gives no methodology for measuring it. [1] Keep computational plausibility separate from experimental confirmation in portfolio reviews. A candidate supported by sequence-function prediction, structural analysis and literature synthesis remains a hypothesis until tested experimentally. [1] Track misses as well as hits: candidates that fail validation, misleading annotations and leads the workflow does not surface. The post reports no error rates or system limitations, so teams would need to establish their own. [1] ## What Platform and R&D Leaders Should Do Next **Analysis:** Run a bounded pilot against one documented antimicrobial-peptide question. Set the validation plan before the search starts, so the team can assess candidate quality against criteria established in advance. Require every prioritized candidate to carry its sequence, source records, model outputs, structural rationale and literature evidence into review. Set thresholds up front for what justifies synthesis, what result rules out a lead and how outcomes inform the next round. Settle access and availability terms with Google before treating Antigravity as an operational tool; the post does not state those terms. [1] Teams evaluating Antigravity for other uses can also read about [Google Antigravity's local-model support and coding agents](/blog/google-antigravity-sdk-local-models-coding-agents). Watch for wet-lab results on this specific candidate. Until such data are reported, describe the example as a computational lead, not confirmed antimicrobial activity. [1] ## Frequently asked questions ### What is Google Antigravity for Science? DeepMind describes it as an agentic science workbench for combining database search, biological target discovery, structural evaluation and prioritization in one session. [1] ### What antimicrobial-peptide result did Google report? The walkthrough reports a novel candidate with predicted activity against pathogens involved in respiratory-tract infections. It does not name the candidate or report experimental validation results. [1] ### Was the candidate experimentally validated? No experimental results are reported in the source. DeepMind says the candidate is undergoing wet-lab validation. [1] ### How did the workflow search for candidates? It used ProtNLM to scan previously uncharacterized UniProt proteins and predict biological function from amino-acid sequences, with the search focused on a narrow sequence window in *Streptococcus pseudopneumoniae*. [1] ### Does Google report a benchmark showing the workflow is faster? DeepMind says fragmented database triage that previously took days can be completed in a single collaborative session, but the post provides no measured time comparison or benchmark. [1] ## Questions **What is Google Antigravity for Science?** DeepMind describes it as an agentic science workbench for combining database search, biological target discovery, structural evaluation and prioritization in one session. [1] **What antimicrobial-peptide result did Google report?** The walkthrough reports a novel candidate with predicted activity against pathogens involved in respiratory-tract infections. It does not name the candidate or report experimental validation results. [1] **Was the candidate experimentally validated?** No experimental results are reported in the source. DeepMind says the candidate is undergoing wet-lab validation. [1] **How did the workflow search for candidates?** It used ProtNLM to scan previously uncharacterized UniProt proteins and predict biological function from amino-acid sequences, with the search focused on a narrow sequence window in Streptococcus pseudopneumoniae. [1] **Does Google report a benchmark showing the workflow is faster?** DeepMind says fragmented database triage that previously took days can be completed in a single collaborative session, but the post provides no measured time comparison or benchmark. [1] ## Sources 1. [Accelerating the search for antimicrobial peptides with Google Antigravity](https://deepmind.google/blog/accelerating-antimicrobial-discovery/) — Google Deepmind — Oct 15, 2026 How this article was made: https://yelno.ai/blog/about