Google DeepMind Unveils Antigravity Cheminformatics Skills

Details

Google DeepMind introduced Cheminformatics Skills for Antigravity, unveiling a demo that executes drug design iterations, which previously took days, in approximately 1 hour.
By applying a division-of-labor architecture between language models and deterministic chemistry tools like RDKit, they analyzed cases of preventing potency loss through rigidification for crizotinib and lorlatinib.
The average CNS MPO for 113 macrocyclic compounds and 22 ether derivatives was 4.34 and 4.43, respectively, exceeding the baseline of 4.0, confirming the relationship between the reduction of rotatable bonds and the prevention of potency loss.
The holdout R² of the potency prediction model was low at -0.16 to 0.38, indicating it can be utilized as an approximate ranking auxiliary signal rather than an accurate predictor.
They proposed four 12-membered ring ALK macrocycles (LOR-DES-1~4) and a synthesis route (Williamson etherification), suggesting docking and binding conformational energy calculations as subsequent steps.
The RDKit-based skills are not included in the public repository (github.com/google-deepmind/science-skills) as of November 4, 2026, and are accessible by installing the Science plugin in the Customizations marketplace within the Antigravity app.

Key point