Google Antigravity adds Cheminformatics Skills for agentic drug discovery

What are Cheminformatics Skills?

Google DeepMind has added a set of Cheminformatics Skills to Antigravity, turning the general-purpose AI agent into a specialized drug discovery workbench. The skills combine large language model orchestration with deterministic chemistry tools, letting the agent navigate databases, extract molecular structures from literature, and run computational chemistry analyses.

The system is built on RDKit, the open-source cheminformatics toolkit, and connects to curated resources from ChEMBL and PubChem. This lets the agent work with molecules represented as IUPAC names, SMILES strings, structure-data files (SDF), or images, unifying formats that often trip up manual workflows.

As DeepMind explains in the technical walkthrough, the goal is to let scientists focus on hypotheses rather than data wrangling. The underlying language model handles workflow orchestration and scientific reasoning, while deterministic chemistry tools execute molecular calculations with no margin for error.

The architecture reflects a deliberate design choice: language models handle the strategic orchestration and scientific reasoning, while deterministic chemistry engines execute the molecular operations. This division of labor ensures that the high-level reasoning benefits from LLM flexibility, while the low-level calculations maintain the rigor and reproducibility expected in pharmaceutical research.

How does the agent accelerate drug discovery?

DeepMind demonstrated the pipeline using a real medicinal chemistry challenge: rigidification of drug molecules. The case study focused on crizotinib, a first-generation lung cancer drug with five major rotatable bonds, and lorlatinib, a third-generation drug that reduced those bonds to zero through cyclization — a structural change that significantly improved potency.

The agent extracted these molecules and their analogues from literature, associated them with reported biological data, and prepared them for analysis. It then launched an interactive viewer with scatter plots showing potency against rotatable bonds, colored by macrocycle status. The visualization immediately revealed that macrocycles like lorlatinib tend to have fewer rotatable bonds and higher potency.

Next, the agent performed matched molecular pair analysis (MMPA), comparing pairs of molecules that differed only by cyclization. The analysis confirmed that cyclization reduces rotatable bonds while generally increasing potency. The agent also built predictive potency models using various feature sets and model architectures, validated on held-out data using chronological splits where available.

Can the agent design new molecules?

Yes. After completing the analysis, the agent proposed novel rigidified macrocycles based on its learnings and recommended multiple synthesis routes involving different compounds. DeepMind says the agent can prepare these designs for further analysis with AlphaFold or physics-based techniques like molecular dynamics simulation.

The results can be exported in formats compatible with industry-standard chemistry software packages, making the transition from computational design to wet-lab synthesis straightforward.

What are the limitations?

DeepMind is upfront about what the tools cannot do. Sparse data, inter-lab assay noise, and narrow dynamic ranges limit the predictive precision of even the most powerful models. Computational designs remain testable hypotheses, not finished drugs. Every proposed molecule must be synthesized and validated experimentally.

The company emphasizes that the speed of meaningful scientific decisions is accelerating, but the fundamental requirement of experimental validation has not changed. What used to take multiple days of computational expertise can now be done in about an hour, but the wet lab work remains essential.

Who can access Cheminformatics Skills?

The Cheminformatics Skills are available in Antigravity today. Users can install the Science plugin from the Antigravity App to get started. Enterprise users can access the skills by choosing 'Use Business Account' and logging in with Google Cloud credentials.

DeepMind is also offering 10,000 scientists the highest usage tier in Antigravity free of charge for six months. Applications are open now, with onboarding on a rolling basis, subject to eligibility, capacity, and selection.

How does Antigravity compare to other AI drug discovery platforms?

Google DeepMind's Antigravity platform is not the only AI agent tackling drug discovery, but its Cheminformatics Skills occupy a specific niche. Unlike end-to-end protein-folding tools like AlphaFold, which predict 3D structures, or generative models that propose new molecules from scratch, Antigravity's skills handle the full cheminformatics pipeline — from literature mining and data extraction to matched molecular pair analysis and predictive modeling — in a single agentic workflow.

The integration of deterministic tools (RDKit, ChEMBL, PubChem) with LLM orchestration means the agent can execute precise molecular calculations while reasoning about the scientific context. This hybrid approach is similar in spirit to what the OpenBioML project has attempted, but DeepMind benefits from its proprietary Gemini model and access to Google's computational infrastructure.

Competing platforms like Insilico Medicine's Pharma.AI and Recursion Pharmaceuticals offer specialized drug discovery pipelines, but they typically require users to navigate separate tools for each stage. Antigravity's agentic approach consolidates these steps into a conversational interface, which lowers the barrier for bench scientists who may not have extensive computational chemistry training.