Antigravity cheminformatics: the early medchem bottleneck moves to the wet lab, not to better binders

On 3 November 2026, Google DeepMind published Agentic cheminformatics with Google Antigravity. The post shows a science workbench in which Cheminformatics Skills let an agent harvest literature compounds, run matched molecular pair analysis (MMPA), train potency models and propose rigidified macrocycles with synthetic routes. The authors say a workflow that once took multiple days for one iteration can finish in about an hour on a lung-cancer tyrosine kinase inhibitor (TKI) case.

The falsifiable claim is narrow. Antigravity’s Cheminformatics Skills move the early medicinal-chemistry bottleneck from data wrangling and MMPA scaffolding to experimental validation throughput. The product value is language-model orchestration plus deterministic RDKit tooling, not a claim that the model designs better binders. Sparse assay data, inter-lab noise and narrow dynamic ranges still cap predictive precision. If independent users get equal cycle time when the RDKit/ChEMBL/PubChem tool layer is removed, or if potency models systematically beat held-out assay labels beyond inter-lab noise without new wet-lab data, that claim fails.

What shipped

Molecules arrive as IUPAC names, Markush structures, structure-data files and figures. Changing one atom changes the compound. The skills encode domain practice into the agent’s workflow and ground calculations in RDKit, with connectors to curated ChEMBL and PubChem resources. The blog is explicit about division of labour: the model handles orchestration, database navigation and scientific reasoning; deterministic chemistry tools execute the molecular calculations.

The technical walkthrough uses a classic rigidification story. Crizotinib, a first-generation TKI for certain non-small cell lung carcinomas, has five major rotatable bonds. Lorlatinib, a third-generation analogue, is cyclized so that the effective rotatable-bond count falls to zero, which raises potency among other effects. Rigidification “pays the entropic cost in advance” by locking a floppy binder into the shape needed at the target.

What the agent actually did

With the skills enabled, the agent searched literature for these drugs and analogues, extracted molecules and associated assay numbers, filtered out buffers and solvents, and wrote a source report. Within minutes it retrieved thousands of compounds. An interactive viewer let scientists plot cellular potency against rotatable bonds, coloured by macrocycle status; the highest-potency compounds in the view were the most rigid, and many macrocycles were predicted brain-penetrant, useful because the cancer often metastasizes to brain.

The agent then ran MMPA on pairs that differ by cyclization. Results matched the textbook expectation: cyclization cut rotatable bonds and generally raised potency. It also trained many potency models across feature sets and architectures on four experimental readouts, preferring chronological splits when dates were available and falling back to scaffold splits for out-of-distribution checks. Finally it proposed new macrocycles, explained medicinal-chemistry rationale, generated electron-pushing diagrams for cyclization reactions, looked up similar reactions in public databases, and scored designs with the models it had just trained. Exports target industry chemistry packages; the post mentions further analysis with AlphaFold or molecular dynamics.

Steelman and gaps

Steelman. A day of expert chemist plus RDKit scripts already does literature harvest, MMPA and scaffold splits. If Antigravity mostly wraps familiar tooling in chat, the “hour vs days” gain is UI convenience, not a new scientific capability. Predictive models trained on sparse public assays have long been noisy; calling designs “hypotheses” is the correct default, so the post may oversell acceleration relative to a well-tooled computational chemist.

That objection has force. What it does not erase is the stated architecture: orchestration stays with the model, bond-level truth with deterministic tools. The demo’s measurable claim is throughput of a full early loop of extract, clean, MMPA, model, propose and route, not a superiority claim on binding affinity. The blog itself lists the hard ceiling: sparse data, inter-lab assay noise, narrow dynamic ranges. Computational designs remain testable hypotheses that must be synthesised and validated experimentally.

Limits the post does not close: the walkthrough is one vendor case study, not a controlled A/B of agent versus human chemist on matched tasks; “about an hour” and “thousands of compounds” are author-reported; no external potency numbers for the newly proposed macrocycles are given; and the free tier for 10,000 scientists, at the highest Antigravity usage for six months on a rolling and eligibility-limited basis, is a distribution offer, not evidence of wet-lab success.

Who should care, and what to watch

Computational and medicinal chemists who still spend early cycle time on format conversion and pair enumeration should ask whether the skills cut time-to-first-testable design on their own ChEMBL/PubChem-scale projects. Buyers of “AI drug design” demos should treat potency plots and proposed routes as hypothesis generators until synthesis and assay close the loop. Platform teams should note the product bet: encode RDKit-grade practice as agent skills, keep calculation deterministic, leave sparse-data limits visible.

Watch next: whether third-party labs publish cycle-time comparisons with the tool layer ablated; whether proposed macrocycles from Antigravity sessions reach measured potency that beats the training distribution after synthesis; and whether chronological or scaffold-split model cards become standard attachments when agent-designed compounds enter internal review.

Close

DeepMind’s usable sentence is not that Antigravity designs better drugs. It is that an agent with RDKit-grounded cheminformatics skills can collapse early literature-to-design scaffolding into about an hour, while the binding claim still waits on the wet lab. Six months from now, either independent cycle-time and assay follow-ups will show the bottleneck really moved, or the product will look like a faster front-end on the same sparse-data ceiling.