Google Antigravity can design new drug candidates in an hour

Reducing the data burden in drug discovery

Developing new drugs requires researchers to gather data from scientific papers, figures and charts, train predictive models, and then suggest molecular candidates that can be synthesized. According to Google DeepMind, a single such iteration typically takes several days.

Google Antigravity's new Cheminformatics Skills support this process. The system combines molecular data from different sources while accounting for the fact that the same compound can appear in multiple forms. In a publication, for example, it may be listed by its IUPAC name, as a Markush structure in a patent, and as an SDF file in a catalogue.

The skills rely on the open-source RDKit cheminformatics toolkit, as well as connectors for selected resources from ChEMBL and PubChem. The language model organizes the workflow, navigates databases and draws scientific conclusions, while deterministic chemistry tools perform the molecular calculations.

Crizotinib and lorlatinib provided the example

The demonstration focused on rigidifying molecules. Rotatable bonds in drug molecules allow them to adopt multiple three-dimensional shapes. Binding to a target, however, requires a specific active shape, which is why medicinal chemistry often restricts a molecule's movement by adding further bonds.

Google DeepMind used crizotinib and lorlatinib as examples. Crizotinib is a first-generation small-molecule tyrosine kinase inhibitor used to treat certain forms of non-small-cell lung cancer, and has five main rotatable bonds. Third-generation lorlatinib is similar, but one end of the molecule was converted into a ring, reducing the number of effectively rotatable bonds to zero. According to the source, this is associated, among other things, with greater potency.

Antigravity searched the scientific literature for these drugs and their analogues, extracted the molecules and associated data, and then filtered out materials irrelevant to the analysis, such as buffers and solvents. The resulting report also documented the sources.

Analysis, machine learning and new synthesis routes

The system created an interactive view of the cleaned compound set. Researchers could use it to examine individual molecules' metadata and visualize relationships. Comparing potency with the number of rotatable bonds showed that macrocycles, such as lorlatinib, generally have fewer rotatable bonds and greater potency. The analysis indicated that most of the macrocycles studied could potentially cross the blood-brain barrier, according to predictions, an important property because of lung cancer's brain metastases.

Antigravity also performed matched molecular pair analysis. This involves comparing pairs of molecules that differ only by a small, defined change. Analysis of cyclized and non-cyclized molecules suggested that cyclization reduces the number of rotatable bonds and generally increases potency.

The system built potency prediction models using multiple feature sets and model architectures, then evaluated them on held-out data. The skills provide guidance for building molecular machine learning models. When a temporal order is available, the system often trains on older compounds and makes predictions on newer ones. In other cases, it uses a scaffold-based split to estimate how well the models generalize.

As a next step, the agent suggested several new, improved macrocycles and possible reactions and synthesis routes for them. It can also verify and modify reactions, create reaction schemes showing electron movement, and search public databases for similar reactions. The proposed compounds can be exported to AlphaFold or molecular dynamics simulations for further analysis.

The designs are still hypotheses to be tested

Google DeepMind stresses that computational designs are not finished drugs. Sparse data, measurement noise between laboratories and a narrow measurement range can limit even advanced models' prediction accuracy.

Google Antigravity's results can be exported in various formats, allowing the data and designs it creates to be used in multiple industry-standard chemistry software packages. According to the source, the tool can primarily accelerate literature processing, data organization, modeling and the preparation of molecular candidates that can be tested experimentally.