Google Antigravity: Macrocyles Generated in One Hour

Google's Antigravity equips Computational Chemistry Skills capable of scouring literature, analyzing thousands of compounds, and proposing macrocycles with synthesis pathways in just one hour. The entire system relies on RDKit and resources like ChEMBL and PubChem, while emphasizing that these designs remain hypotheses to be validated and that predictive accuracy is limited by data quality.

Required Validations and Export to Industry Tools

The outputs from the agent can be saved in several formats that are compatible with the main chemistry software used in the industry. Google notes that the limitation of available data, variations in noise between laboratories, and restricted dynamic ranges hinder predictive performance, even for the most sophisticated models. The proposals generated by the agent are treated as hypotheses that need to be verified, requiring synthesis and experimental confirmation before any use. According to Google, these skills allow for faster generation of better-supported design hypotheses. The workflow presented would have previously required several days and significant computational expertise; Antigravity aims to accelerate the tasks of computational and experimental chemists and refocus them on essential questions. The speed of scientific decision-making is thus enhanced.

A Complete Discovery Cycle Compressed to One Hour

A discovery cycle that typically took several days is now presented as reduced to one hour thanks to Antigravity and its Computational Chemistry Skills. Based on real cases of lung cancer treatments, the agent explores thousands of compounds, performs paired molecular analysis, and proposes rigid macrocycles with synthesis pathways. In just a few minutes, it massively retrieves compounds, filters out irrelevant elements like solvents and buffers, and documents sources in a report.

Integrated Tools and Task Sharing Between AI and Deterministic Engines

With the Computational Chemistry Skills, Antigravity becomes a shared workspace with access to local tools, where computational and experimental chemists can collaborate on the same platform. The language model coordinates the different phases, queries databases, and supports scientific reasoning, while deterministic engines perform calculations on the molecules. The architecture is based on RDKit, which has been in development for over twenty years, as well as connectors to selected databases like ChEMBL and PubChem, allowing for the integration of industry standards. The agent can also launch an interactive viewer combining curated compounds, associated data, and customizable graphics, fostering smooth collaboration with the user.

Literature Curation and Signals on Macrocycles

After extracting molecules from documents and associating them with available measurements, a report indicates that macrocycles are largely predicted to penetrate the brain, an important criterion since the targeted lung cancer frequently spreads to the brain. A visualization of potency based on rotatable bonds, colored according to macrocycle status, highlights that macrocycles such as lorlatinib exhibit fewer rotatable bonds and higher potency.

Rigidifying to Gain Potency: Principle and Application

Rigidification is presented as a classic medicinal chemistry technique to increase potency. Molecules, composed of substructures connected by rotatable bonds, adopt numerous three-dimensional conformations that make them flexible in solution; binding to a target requires locking into a unique active form, with an entropic penalty that reduces affinity and potency. By adding a second bond between two substructures initially connected by a single rotatable bond, chemists significantly limit movements and can lock in the desired geometry. This approach, often summarized by the idea of paying the entropic cost upfront, can notably increase potency.

Advanced Analyses and Synthesis Generation by the Agent

Once the data is structured, the agent conducts a paired molecular analysis, which it suggested itself given the many informative pairs expected. The comparison of molecules that differ by cyclization shows that cyclization reduces the number of rotatable bonds and tends to increase potency. Meanwhile, the agent builds numerous potency models with different families of variables and architectures, then validates them on a reserved dataset. Recommendations guide the construction of models, frequently employing chronological separations—training on older compounds to predict newer ones when information exists—or, if not possible, structural separations to better estimate generalizability. Based on these elements, the agent imagines new macrocycles and submits various reactions and synthesis routes, while preparing the molecules for further examinations via AlphaFold or molecular dynamics simulations. The integrated tools allow for the manipulation and validation of reactions, the creation of electron-pushing mechanisms, and the search for precedents in public databases. The agent details its reasoning in medicinal chemistry, proposes routes compatible with already described methods, and uses its previous models to estimate the potency of each proposal. It can also present a cyclization reaction by providing electron-pushing mechanism diagrams.