Most drug discovery delays don’t come from the lab bench. They come from the gap between reading the literature and deciding which molecule to try next. Google DeepMind says it has built a tool aimed squarely at that gap, and that it ran a full example in about an hour.
The company calls it agentic cheminformatics with Google Antigravity. Cheminformatics is software-assisted analysis of molecular information, such as structures, properties and how compounds relate to one another. The agentic part means an AI system carries out a chain of steps on its own rather than answering a single prompt. In the demonstration, the agent moved from published compounds to proposed molecular designs.
What Google showed
The November 3 announcement centers on a lung cancer research example involving crizotinib and lorlatinib. Both are kinase inhibitors used against ALK-positive lung cancer. Lorlatinib is the later drug, developed in part to deal with resistance that emerges on earlier treatment. That makes the pair a sensible test case: there is a published trail of chemistry and evidence linking one compound to the next.
Google says its Cheminformatics Skills connect the literature to computational chemistry work. The source material available to me cuts off at that point, so I can’t describe exactly how those skills operate or what the agent produced. That matters, and it is worth keeping in view as you read the claim.
Why the in-between work matters
Medicinal chemists spend a lot of time on tasks that are necessary but not creative. They pull compounds from papers, standardize the data, compare structures, look for patterns, and then propose analogs. Each step is manageable. Together they eat days or weeks. If an agent can handle that stretch reliably, chemists get more time for judgment calls, which is where their expertise counts.
It also fits a broader trend. AI in drug discovery has largely focused on predicting protein structures or scoring individual molecules. Agents that string together many small analytical tasks are a different bet. They treat the workflow, not the model, as the product.
What to ask before getting excited
An hour is a striking number, but speed is the easy part to measure. The harder questions are about quality.
- Were the proposed designs chemically sensible and synthesizable?
- How did they compare with what an experienced chemist would suggest from the same literature?
- Was the example chosen because the evidence was unusually clean?
- Has anyone made and tested any of the molecules?
Nothing in the announcement as summarized here answers those. A proposed design is a hypothesis, not a result. Drug candidates fail for reasons that software analysis of published data can’t always see, including toxicity, metabolism and plain bad luck in biology.
The practical takeaway
For clinicians and patients, this is not near-term news about new lung cancer drugs. For health entrepreneurs and research teams, it is a signal that the major AI labs now see the research workflow itself as a target. The useful test will be independent evaluation: do chemists using these agents reach better candidates, or just reach the same ones faster? Both outcomes have value. Only the first changes how drugs get discovered.