Agentic cheminformatics with Google Antigravity

What it claims

Google DeepMind's Cheminformatics Skills add a chemistry workflow to the Google Antigravity agent app. In the post's demo, the agent curates kinase inhibitor data from literature, runs matched molecular pair analysis (MMPA), trains potency models and proposes rigidified macrocycles with synthetic routes in about an hour. The post puts the same workflow at multiple days done by hand. The test case is crizotinib, a first-generation ALK/ROS1-class tyrosine kinase inhibitor with five major rotatable bonds, versus lorlatinib, a third-generation macrocyclic analogue with effectively zero.

How it works

The language model handles orchestration, database navigation and reasoning. RDKit computes the molecular properties, and connectors pull data from ChEMBL and PubChem. The skills encode practice rules into the agent. For potency models, the agent uses chronological splits (train on older compounds, predict newer ones) when dates are available, and falls back to scaffold splits to estimate generalization. It trains many models across feature sets and architectures and validates them on a held-out set.

The agent also opens an interactive viewer with the curated compounds and configurable scatter plots. Plotting potency against rotatable bonds, colored by macrocycle status, shows macrocycles with fewer rotatable bonds and higher potency. The MMPA compared pairs that differ only by cyclization: cyclization reduced rotatable bonds and generally raised potency. The agent proposed MMPA itself before the user asked. Designed compounds can be prepared for AlphaFold or molecular dynamics, and results export in formats that standard chemistry packages read.

Evidence

The post reports no model metrics, no dataset size, no MMPA effect sizes and no validation of any proposed macrocycle. The one-hour figure comes from a single demo on a well-known retrospective case, where the answer (lorlatinib) is already published.

How to use it

• Install the Antigravity app and add the Science plugin.
• Enterprise users choose "Use Business Account" and sign in with Google Cloud credentials.
• DeepMind offers the highest Antigravity usage tier free for 6 months to 10,000 scientists. Onboarding is rolling, subject to eligibility, capacity and selection, and requires an application.

The post does not state pricing, the underlying model, or whether the skills run outside Antigravity.

Limits

The authors name sparse data, inter-lab assay noise and narrow dynamic ranges as limits on predictive precision. They call the designs testable hypotheses that must be synthesized and validated experimentally. The post does not test the workflow on a prospective target or on any chemotype besides crizotinib/lorlatinib analogues.