Pharma
- 01Google DeepMind's AI agent can search thousands of compounds and propose new molecular structures in about an hour, a process that typically takes days.
- 02The AI demonstration utilized existing lung cancer medicines, crizotinib and lorlatinib, to examine related compounds and suggest new designs based on their properties.
- 03The agent employs cheminformatics tools and databases, such as RDKit, ChEMBL, and PubChem, to analyze molecular structures and their effectiveness.
Show all 5Show less
- 04Alphabet's Isomorphic Labs is applying AI models to drug research, developing a Drug Design Engine called IsoDDE, and collaborating with Johnson & Johnson on drug development.
- 05The AI-generated designs are considered testable hypotheses and have not yet been validated in laboratory settings, leaving the responsibility of evidence evaluation to scientists.
Follow Deployed on Google
Google DeepMind showed an AI agent search thousands of compounds and propose new molecular structures with routes to make them, all in about an hour. The demonstration completed a computational design workflow that Google said would typically take days.
The starting point was two existing lung cancer medicines, crizotinib and lorlatinib. The task was to examine related compounds and use that evidence to suggest what a chemist might make next.
That puts the agent into the preparation behind an experiment: connecting published structures to measured activity, then turning the analysis into design proposals.
Finding the Evidence Behind a Design
The example examined a familiar medicinal chemistry strategy: restricting how freely a molecule moves so it can bind more effectively to its target.
Crizotinib has five major rotatable bonds. Lorlatinib’s ring structure reduces the effective number to zero, contributing to greater potency.
To investigate that relationship, the agent gathered compounds from the literature and connected them to reported data. An interactive viewer allowed scientists to inspect individual records and explore relationships between molecular properties and potency.
The difficulty begins before the analysis. A compound can appear as a chemical name in one document and a drawn structure in another, requiring the system to preserve its identity as it assembles the evidence.
“Alter a single atom or bond and you have described a different compound altogether,” the DeepMind authors wrote.
Antigravity’s Cheminformatics Skills coordinate existing resources, including RDKit for molecular calculations and the ChEMBL and PubChem databases. The language model directs the workflow, while chemistry tools execute the calculations.
The agent compared closely related molecules to examine the effect of structural changes and built predictive models to inform subsequent proposals. Scientists could review the outputs before deciding which designs deserved further work.
Isomorphic Takes Selected Designs Into the Lab
The week's drug discovery, trials and health-system deployments. Wednesdays.
The demonstration comes as Alphabet applies its models to drug research through Isomorphic Labs.
“Isomorphic is using Argon to support drug design research,” said Sundar Pichai, CEO of Alphabet and Google, during the company’s latest earnings call.
Isomorphic develops its own Drug Design Engine, known as IsoDDE. Pichai did not identify Antigravity as part of that work.
In September, Isomorphic described agents refining molecular designs against requirements set by drug designers, including the intended target and desired properties. The company then selected compounds for synthesis and testing, reporting an example that modulated its intended biological target.
Those are company-reported laboratory results from a separate workflow. They illustrate the next step beyond a computational proposal: making the molecule and measuring what it does.
Isomorphic’s research collaboration with Johnson & Johnson, announced earlier this year, also places its technology within a drugmaker partnership covering multiple modalities, including biologics.
OpenAI and Anthropic Connect Agents to Research Work
OpenAI is pursuing a similar connection between scientific reasoning and execution. Its GPT-Rosalind tools bring sourced research and biological-data analysis into a shared workspace, with outputs available for expert review. Its latest acquisition of Orchestra Bio extends that effort into coordinating drug programs, including tracking how a failed manufacturing batch affects a trial’s schedule.
Anthropic has reported work that connects proposed designs to laboratory feedback. Last month, it described Claude working with a researcher through BigHat Biosciences’ lab-data system. BigHat made and tested about 1,010 antibody variants over roughly 10 weeks, with results informing subsequent designs.
The team also revisited unreliable stability and clumping measurements, using raw instrument data to develop revised metrics that BigHat tested experimentally. That work involved improving how candidates were assessed as well as proposing changes to the molecules.
What the Hour Leaves Unanswered
Google’s timing measures the computational exercise. The company did not report laboratory validation of the proposed molecules or a measured reduction in a drug-development program’s costs.
“Computational designs are testable hypotheses, not finished drugs,” the DeepMind authors wrote.
The authors also acknowledged limits from sparse data and variation between laboratory assays. Completing the analysis faster leaves scientists responsible for judging whether the evidence supports the proposed experiment.
For a pharmaceutical team, the operational question is how much preparation time remains saved after that review, including any corrections. Antigravity is available through its Science plugin, with enterprise access using Google Cloud credentials; results from those users would provide evidence beyond the demonstration.
Last updated
