Google DeepMind cuts molecule design cycle to one hour

Key Points

1. Google DeepMind released Cheminformatics Skills for its Antigravity platform on November 3, 2026.
2. An agent proposed novel macrocycles in one hour, a workflow the company says took days.
3. The tool is available now, but every design still requires lab synthesis and validation.

The latest:

A molecule-design workflow that previously ran for several days and demanded deep computational expertise was completed in a single hour, according to Google DeepMind, which released Cheminformatics Skills for its Antigravity platform on November 3, 2026. The package turns Antigravity into a working bench for computational and experimental chemists, with full access to local tools. The company says every proposed design must still be synthesized and tested experimentally.

Details:

• The split architecture: The skills divide the work between two layers, according to Google DeepMind. The language model orchestrates the workflow, navigates databases and handles scientific reasoning, while deterministic chemistry tools run the molecular calculations themselves — keeping the numeric work out of the model’s hands.
• The underlying stack: The system is grounded in RDKit, the open-source cheminformatics library, with connectors into two public chemistry databases: ChEMBL and PubChem. That gives the agent direct reach into published compound and bioactivity records rather than relying on the model’s own stored knowledge.
• The demonstration: In a lung cancer case, the agent started from two real drugs, crizotinib and lorlatinib, and searched thousands of compounds drawn from the literature, according to the company. It then ran matched molecular pair analysis, a standard technique for isolating how single structural changes shift a molecule’s properties.
• What it produced: The agent trained and compared predictive models, then proposed new rigidified macrocycles — large ring-shaped molecules — complete with synthetic routes and potency predictions. Google DeepMind says the full sequence, from literature search to candidate designs, closed in one hour.
• The time claim: The company frames the one-hour figure against a baseline of several days for the same workflow, carried out by someone with substantial computational chemistry experience. The comparison is Google DeepMind’s own; no independent benchmark was published alongside the release.
• The stated limit: Google DeepMind stresses that the output is a hypothesis, not a result: every design the agent proposes must be synthesized in a lab and validated experimentally before it means anything. Nothing in the release shortens that step.
• How to get it: The skills are live now. Individual users install the Science plugin inside the Antigravity app; enterprise customers access them by signing in through Google Cloud accounts. The company did not disclose pricing or usage limits for either tier.
• Who it targets: The positioning is aimed at both computational chemists and bench scientists, with the company presenting Antigravity as a single scientific workbench rather than a specialist tool. Full access to local tooling is part of that pitch, keeping proprietary data and scripts on the user’s own machine.

Background:

Cheminformatics has long depended on specialists who can write code against libraries like RDKit and query compound databases directly. The skill barrier, not the software cost, has been the practical bottleneck for smaller research groups.

Between the lines:

The architecture choice is the substance here. By routing calculations through deterministic tools rather than the language model, Google DeepMind is addressing the credibility problem that has limited AI in chemistry: models that produce plausible numbers. The one-hour figure measures how fast a hypothesis arrives, not how often it survives the lab — and the company’s own insistence on experimental validation concedes exactly that.

What’s next

Watch for published results from labs that synthesize agent-proposed macrocycles, the first real test of hit rates. Also watch whether Google DeepMind discloses pricing, usage limits or independent benchmarks for the enterprise tier.