Antigravity cheminformatics is the better choice today for a team that wants an agent doing real chemistry work this week. On November 3, 2026, Google DeepMind's new skills for Google Antigravity claim an agent can search thousands of compounds, run matched molecular pair analysis and propose macrocycles with synthetic routes in about an hour, a loop the post says typically takes multiple days. El Agente Fármaco, an academic preprint submitted the day before, is the stronger design to watch if you need open, inspectable infrastructure, but it is not yet a product. Neither produces a finished drug; both produce hypotheses that still need synthesis and lab validation.
TL;DR
- Winner for immediate use: Google Antigravity with the Cheminformatics Skills, because it ships inside an app you can install now.
- Winner for openness and research reuse: El Agente Fármaco, with the caveat that it is an unreviewed preprint.
- Both split the work the same way: a language model plans, deterministic chemistry code computes.
- Outputs from either tool are testable hypotheses, not drug candidates.
- Last verified: 4 November 2026, against the DeepMind post and the arXiv listing.
What is Antigravity cheminformatics?
Antigravity cheminformatics is a set of skills that lets the Google Antigravity agent do computational chemistry. Antigravity began as an agentic coding environment; this release, described in DeepMind's announcement of 3 November 2026, repositions it as a science workbench with local tooling for both computational and bench chemists.
The skills are grounded in RDKit, the open-source cheminformatics toolkit, and connect to curated data from ChEMBL and PubChem. It follows DeepMind's earlier antimicrobial peptide release for Antigravity, which makes this the second science-skills pack in roughly a month. For the wider pattern of agents as shared workbenches, see our piece on Claude and multi-agent science workbenches.
Why does the language model not do the chemistry itself?
Because molecular representation leaves no room for error. DeepMind notes that the same molecule can appear as an IUPAC name in a paper, a Markush structure in a patent or an SDF file in a catalogue, and that changing a single atom or bond describes a different compound. So the model handles orchestration, database navigation and reasoning, while RDKit and related tools execute the calculations deterministically.
This is the same trust pattern we describe in agentic AI architecture in 2026: let the model decide what to run, and let tested code produce the numbers. Tool access of this kind is also a natural fit for the approach covered in MCP in agentic AI.
What did the agent actually do in the demo?
DeepMind's worked example uses lung cancer kinase inhibitors. Crizotinib, a first-generation tyrosine kinase inhibitor, has five major rotatable bonds; lorlatinib, a cyclised third-generation analogue, reduces that count to zero and gains potency. The idea is rigidification: a floppy molecule pays an entropic cost when it locks into a binding shape, so chemists cyclise it to pay that cost in advance.
The agent worked through four steps:
- Literature mining. It searched academic and patent sources, extracted molecules, filtered out buffers and solvents, and logged provenance from ChEMBL, PubChem and papers. An interactive viewer in the post, as reported by explainx.ai, shows 568 curated compounds across four charts.
- Matched molecular pair analysis. The agent suggested this on its own and found that cyclisation reduces rotatable bonds while generally raising potency.
- Model building. It trained several potency models, preferring chronological splits and falling back to scaffold splits, and benchmarked architectures across four experimental read-outs.
- Design. It proposed new rigidified macrocycles with a medicinal chemistry rationale, candidate reactions and synthesis routes, plus electron-pushing diagrams for the cyclisation step.
The report also flagged the macrocycles as largely predicted to be brain-penetrant, which matters because this cancer often spreads to the brain. That is a model prediction, not a clinical finding.
What is El Agente Fármaco, and how does it differ?
El Agente Fármaco is an LLM-based drug discovery agent from a group including Alán Aspuru-Guzik, described in arXiv 2611.03832. It turns goals written in plain language into computational screening workflows, selects tools, reviews results and adjusts later steps.
Its central idea is Fármaco Atlas, a shared memory that stores molecules, structures and results with provenance. One request launches calculations across a whole compound collection, so the model does not make a separate decision per compound. The authors report, per the preprint, that against an expert-led campaign it proposed chemically distinct candidates with competitive docking scores and higher calculated drug-likeness, and that in a 22-task benchmark it had the highest mean model-judged score with lower median token use than plain tool calls and Codex. These are the authors' own results and the paper has not been peer reviewed.
Which one should you choose?
| Question | Antigravity Cheminformatics Skills | El Agente Fármaco |
|---|---|---|
| Can you use it now? | Yes, via the Science plugin in the Antigravity app | Research preprint |
| Chemistry engine | RDKit plus ChEMBL and PubChem connectors | Tool-driven screening with shared memory |
| Evidence | Vendor demonstration | Author-reported benchmarks, not peer reviewed |
| Best for | Medicinal chemistry teams iterating on a series | Groups building reproducible screening infrastructure |
Pick Antigravity if your constraint is time. Enterprise users can sign in with Google Cloud credentials, and DeepMind is offering 10,000 scientists its highest Antigravity usage tier free for six months, subject to eligibility and capacity. Pick El Agente Fármaco if your constraint is control, and you are willing to read a long paper and assemble the pieces yourself. Teams in regulated settings should also read our comparison of agentic AI in health care before connecting any of this to clinical data.
What are the limits?
DeepMind states them plainly: sparse data, inter-lab assay noise and narrow dynamic ranges limit the predictive precision of even strong models, and every proposed design must be synthesised and validated experimentally. The one-hour figure is a vendor claim about one demonstration on well-studied compounds. Expect less on a new target with thin public data. Nothing here is medical advice or evidence that any compound is safe or effective.
FAQ
Q: What is Antigravity cheminformatics?
A: It is a set of Cheminformatics Skills for Google Antigravity, released by Google DeepMind on 3 November 2026, that lets an agent run RDKit-based chemistry workflows with ChEMBL and PubChem data.
Q: Does the AI design drugs on its own?
A: No. It proposes candidate molecules and synthetic routes as hypotheses. Each design still has to be made in a lab and tested.
Q: Why use deterministic tools instead of the model for calculations?
A: A single wrong atom or bond means a different compound, so the model orchestrates while tools such as RDKit perform the exact calculations.
Q: Is El Agente Fármaco peer reviewed?
A: No. It is a preprint on arXiv (2611.03832) submitted on 2 November 2026, so its benchmark claims are the authors' own and unverified by independent review.
Q: How do I get access to the Antigravity science skills?
A: Install the Science plugin in the Google Antigravity app. Enterprise users choose "Use Business Account" and sign in with Google Cloud credentials. DeepMind also runs a rolling application programme for free usage.
Sources
- Agentic cheminformatics with Google Antigravity, Google DeepMind
- El Agente Fármaco, arXiv 2611.03832
- RDKit, ChEMBL, PubChem
- Third-party corroboration: explainx.ai walkthrough, globai.org analysis, relvehq.com summary
Last verified: 4 November 2026. Corrections log: none yet. This article was prepared with AI assistance and checked against the primary sources listed above.
Top comments (0)