AgentsAutonomy & tool use
Google DeepMind shows Antigravity agent retracing lorlatinib's macrocycle design and offers 10,000 scientists free access
Google DeepMind showed a Gemini-driven Antigravity agent retracing the path from crizotinib to lorlatinib. In one session it curated compounds, built potency models and proposed new macrocyclic designs. The skills ship as a Science plugin, and 10,000 scientists get six months of free access.
From the Google DeepMind article of November 3, 2026 and cited sources: In a single session, the agent mined the literature, curated compounds, ran matched-molecular-pair analysis, built potency models and proposed macrocycles with synthesis routes. [1]
SOURCES[1] Agentic cheminformatics with Google Antigravity — Google… deepmind.google
AT Pulse research deskPublished · Updated · 4 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- In a single session, the agent mined the literature, curated compounds, ran matched-molecular-pair analysis, built potency models and proposed macrocycles with synthesis routes. [1]
- Original source
- Google DeepMind article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 3, 2026 article published
- Covered
- Nov 4 edition
- Evidence
- 7 sources across 6 sites
- 1Agentic cheminformatics with Google Antigravity — Google…deepmind.google
- 2Google Science Skills: Pharma Informatics Evaluation —…intuitionlabs.ai
- 3Google launches Gemini for Science — EdTech Innovation Hubedtechinnovationhub.com
- 4Google Antigravity: Accelerating Antimicrobial Peptide…deepmind.google
- 5Google Antigravity Antimicrobial Discovery Compresses the…remio.ai
- 6Discovery of lorlatinib (PF-06463922), J Med Chem 2014 —…api-pubmed-ncbi.laibokeji.com
QuietHow it’s measured
Hacker News · 2 likes · 1 comment
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
In a single session, the agent mined the literature, curated compounds, ran matched-molecular-pair analysis, built potency models and proposed macrocycles with synthesis routes. DeepMind reports it took about an hour, against several days by hand, but describes no timing protocol or baseline. [1]
The design has the Gemini agent call deterministic tools instead of computing chemistry itself. It uses RDKit for molecular descriptors and queries ChEMBL and PubChem for data, so the intermediate numbers can be checked. Models are judged with scaffold splits and chronological splits, though the post reports no headline performance metrics. [1]
The case study replays a well-documented drug program. The agent's conclusion that macrocycles with fewer rotatable bonds gave higher potency matches the design rationale Pfizer published in 2014. That supports the pipeline's reliability more than it shows anything new, and DeepMind itself calls the new designs hypotheses. [1] [6]
The workflow extends Google's Science Skills, released as open source in May. They connect more than 30 databases and include cheminformatics skills built on ChEMBL, PubChem and OpenFDA. Commercial users still need to check licences and data-governance terms, and audit-trail and GxP status (pharma's regulated quality rules) are undocumented. [1] [2] [3]
The evidence shows a pipeline that can be audited and that rediscovers a known design rationale. For drug teams, that points to faster triage from literature to structure–activity analysis.
Read the full assessment
It does not show better compound outcomes, which would need benchmarks or wet-lab results first.
Executive brief
Google DeepMind says a Google Antigravity agent took about an hour to do a medicinal-chemistry analysis that would otherwise take several days. In one session it mined the literature, curated compounds, built potency models and proposed new macrocyclic drug designs (DeepMind). The case study looks back at how the lung-cancer drug crizotinib led to lorlatinib, a later drug with a ring-shaped (macrocyclic) structure. The skills ship as a "Science plugin". Google is also giving 10,000 scientists six months of free top-tier access. No independent evaluation of this cheminformatics workflow has appeared. DeepMind itself calls the designs testable hypotheses, not drugs.
What changed and event timeline
Science Skills launch
At I/O, Google released open-source Science Skills on GitHub and inside Antigravity. The bundle connects more than 30 databases and includes 5 cheminformatics/clinical-translation skills (ChEMBL, PubChem, OpenFDA and others) (;).
Antimicrobial peptide demo
DeepMind reported an Antigravity workflow that combines ProtNLM, the AlphaFold DB and BLAST. It surfaced bacteriocin candidate A0A0U0D591, which is now in wet-lab validation ().
Critical analysis
Remio argued that the peptide demo shortens the search but proves no discovery. It cited pending validation, opaque candidate ranking and selective presentation ().
Cheminformatics post
DeepMind published a case study on drug rigidification built on RDKit, ChEMBL and PubChem. Results export in standard formats, and the 10,000-scientist free-access offer was announced ().
Capabilities and access
- Platform: Google Antigravity (agentic development environment), with a Science plugin that adds "Cheminformatics Skills" (DeepMind).
- The article text does not clearly state which version runs the workflow; version names appear only in interface screenshots.
- What the workflow does:
Read the full section
- Platform: Google Antigravity (agentic development environment), with a Science plugin that adds "Cheminformatics Skills" (DeepMind).
- Model: Gemini. The article text does not clearly state which version runs the workflow; version names appear only in interface screenshots.
- What the workflow does:
- Retrieves and curates compounds from the literature and databases.
- Runs matched-molecular-pair analysis.
- Builds potency models.
- Designs macrocycles and proposes synthesis routes, with electron-pushing mechanism diagrams.
- Access: free Science Skills on GitHub (IntuitionLabs), plus the six-month free top tier for 10,000 scientists.
Technical analysis for researchers and developers
- Architecture: a Gemini agent orchestrates deterministic tools (RDKit descriptors, ChEMBL/PubChem queries) instead of asking the language model to compute chemistry itself.
- Evaluation design: the agent uses scaffold splits to estimate out-of-distribution performance and chronological splits (train on older compounds, predict newer ones) (DeepMind).
- Data: the dashboard shows 568 curated compounds. The post itself says predictions are limited by sparse data and assay noise.
Read the full section
- Architecture: a Gemini agent orchestrates deterministic tools (RDKit descriptors, ChEMBL/PubChem queries) instead of asking the language model to compute chemistry itself. That makes the intermediate numbers checkable.
- Evaluation design: the agent uses scaffold splits to estimate out-of-distribution performance and chronological splits (train on older compounds, predict newer ones) (DeepMind). This is good practice, but the post publishes no headline metrics.
- Data: the dashboard shows 568 curated compounds. The post itself says predictions are limited by sparse data and assay noise.
- Reproducibility: there is no released notebook, prompt log or seed. Agent runs are nondeterministic, so expect different curation sets from one run to the next.
Claims and evidence
- About one hour versus multiple days: vendor-reported, with no timing protocol or baseline (DeepMind).
- Macrocycles had fewer rotatable bonds and higher potency: the agent rediscovered a design rationale Pfizer published in 2014 (J Med Chem summary).
- New macrocycle designs: none have been synthesised or tested.
Read the full section
- About one hour versus multiple days: vendor-reported, with no timing protocol or baseline (DeepMind).
- Macrocycles had fewer rotatable bonds and higher potency: the agent rediscovered a design rationale Pfizer published in 2014 (J Med Chem summary). This retrospective check validates the pipeline, not new science.
- New macrocycle designs: none have been synthesised or tested.
- "Minutes rather than hours" for Science Skills overall: an internal Google test (IntuitionLabs).
- Independent corroboration: none found for the cheminformatics workflow.
Context and prior work
- Lorlatinib (PF-06463922): found through structure-based design and lipophilic-efficiency optimization of macrocycles, starting from crizotinib (J Med Chem 2014). The demo replays a well-documented case, which suits checking the method but limits claims of novelty.
- Series context: this post follows the antimicrobial peptide demo (DeepMind) and the May Science Skills launch (Crypto Briefing). Google is positioning Antigravity as a general science workbench.
Limitations, safety and contested findings
- DeepMind's own caveats: the designs are hypotheses; sparse, noisy data limits the models; and synthesis and validation are still required (DeepMind).
- Critique of the related peptide demo:
- Only the successful runs are shown.
Read the full section
- DeepMind's own caveats: the designs are hypotheses; sparse, noisy data limits the models; and synthesis and validation are still required (DeepMind).
- Critique of the related peptide demo:
- Only the successful runs are shown.
- Rejected leads and failed tool calls are not shown.
- There is no comparison with specialised pipelines that have already produced validated hits (Remio).
- Pharma use: audit trails and GxP status (the regulated-manufacturing quality rules pharma must follow) are undocumented (IntuitionLabs).
Business and practitioner implications
- What it can be used for now: speeding up literature-to-SAR triage (structure–activity relationship analysis) with standard formats a team can audit.
- Pilot approach: test it on a past program where the answers are known, as DeepMind did with lorlatinib.
- Free access: the free tier lowers adoption cost.
Read the full section
- What it can be used for now: speeding up literature-to-SAR triage (structure–activity relationship analysis) with standard formats a team can audit. It cannot yet be used as evidence of better compound outcomes.
- Pilot approach: test it on a past program where the answers are known, as DeepMind did with lorlatinib. Log every step the agent takes and compare its compound curation with a chemist's.
- Free access: the free tier lowers adoption cost. Commercial use should check data-governance terms and licences, since AlphaGenome, for example, is restricted to non-commercial use (IntuitionLabs).
- What to watch: wet-lab results for the peptide and macrocycle candidates.