◎THE FACTS AT A GLANCE
  • Google DeepMind's 3 November 2026 walkthrough says its Google Antigravity Cheminformatics Skills took roughly an hour to search literature, analyze molecular structures, train potency predictors and propose candidate macrocycles and synthetic routes. This is a demonstration, not a controlled productivity trial.
  • The article describes a curated interactive set of 568 compounds while an agent drew from thousands of initially retrieved records. The workflow uses deterministic chemistry tools including RDKit and resources such as ChEMBL and PubChem.
  • Google explicitly says generated designs are testable hypotheses, not finished drugs. No evidence is presented that its newly proposed macrocycles were synthesized, showed measured potency, cleared toxicity checks or improved patient outcomes.
01

The Signal

Drug discovery contains long sequences of operations that look suitable for an agent: gathering public chemical records, resolving different molecule representations, normalizing assay measurements, finding related compounds and training predictive models. But these steps mix deterministic calculations with assumptions about noisy biological evidence.

On 3 November Google DeepMind demonstrated new Cheminformatics Skills for Antigravity. In a walkthrough based on lung cancer treatments crizotinib and lorlatinib, an agent retrieved compounds from papers and catalogs, curated a set presented in an interactive viewer, studied molecular rigidity and potency, trained prediction models and proposed new macrocycles with candidate synthetic routes. Google describes the end-to-end demonstration as taking about one hour.

The project uses specialist building blocks: RDKit for molecular operations and links to sources including ChEMBL and PubChem. That separation matters. A language model may plan and navigate a workflow while rule-based chemistry software checks or performs operations whose correctness should not depend on free-form generation.

02

What the evidence changes

The demonstration highlights a real bottleneck. A molecule can appear as a name in a paper, an image in a patent or a machine-readable chemical structure. Conflating two representations or modifying an atom can invalidate the downstream analysis. Automated curation with traceable source links could reduce manual transcription and integration work, provided the extracted records survive expert validation.

The reported viewer contains 568 curated compounds from a wider search. The agent performs matched molecular pair analysis, comparing closely related molecules that differ by cyclization, and uses held-out validation for predictive models. Google describes scaffold and chronological splits as ways to reduce misleading generalization from highly similar chemistry. The publication, however, does not provide a controlled, independently reproduced speedup estimate across a representative suite of drug projects.

The final step is the sharpest evidence boundary. The agent proposes candidate macrocycles and plausible synthetic routes, including structures that predictive models estimate could improve desired properties. These are computational suggestions. Novel synthesis, successful product isolation, binding affinity, permeability, toxicity and clinical utility would each require their own physical measurements.

03

The Cortist perspective

Cortist would not count a one-hour agent demo as a one-hour drug discovery. The demonstrated result is the production of an interpretable computational analysis and a set of hypotheses. It is not evidence that any proposed new treatment works, or that the workflow systematically outperforms medicinal chemists on the same targets.

A rigorous comparison would pre-register input datasets, task success criteria, operating costs and error-handling rules. Experts should assess molecular-identity accuracy, traceability to assay conditions, leakage in train/test splits, invalid reactions and whether subsequent laboratory work confirms the claimed predicted property. The time spent correcting agent output belongs in any productivity estimate.

This work also connects to the gap reported in Anthropic's recent study of scientific AI use. Digitally available data analysis and coding are easier to accelerate than collecting reliable physical evidence. Google illustrates a bridge from computation toward experiment, not the disappearance of the experimental gate.

04 / WHAT TO WATCH
  • Release of runnable workflow recipes, provenance exports and representative comparison cases for independent reproduction.
  • Blind prospective evaluation against human cheminformatics teams, including correction time and actual candidate quality.
  • Physical synthesis and potency assays for newly proposed molecules, with failed designs counted alongside successes.
  • Evidence that model-suggested chemistry improves downstream decision quality under real project budgets and data-quality constraints.
CORTIST TAKE

The most credible near-term benefit is faster, more traceable computational preparation for medicinal chemistry. Google demonstrates an organized agent workflow; it has not demonstrated an experimentally validated new drug. The next evidence is wet-lab confirmation and independent task benchmarking.

05

Related research, distinct evidence

These related articles provide context, not additional independent replications or proven claims. Explore the original sources and remaining uncertainty.

Primary sources and verification boundaries

The original publications describe developer-run benchmarks, engineering observations and workflow demonstrations. Cortist has not replicated private test datasets, reproduced proprietary experiments or validated downstream laboratory outcomes. Source attribution does not imply endorsement.

Read our editorial standards and research methodology.

Google DeepMind: Agentic cheminformatics with Google AntigravityOriginal technical demonstration, 3 November 2026Cortist: usage of AI in science and the discovery gapRelated source-attributed Cortist scientific adoption analysis