Google Research has published a round-up of its disease work, and the useful way to read it is by stage: a cancer-screening evaluation on 125,000 women, one clinical trial in genetic diagnosis, and a long tail of lab and computational results. Yossi Matias, who heads Google Research, wrote the post on Google’s blog on 12 October, with Pushmeet Kohli of Google DeepMind listed as a second author. It collects efforts across the company rather than announcing a single product, and the company’s own framing is that it wants to help “ensure discoveries make it out of the lab.”
The strongest item is the mammography work. Imperial College London and the UK’s National Health Service worked with Google on two Nature Cancer studies. They tested the screening software on 125,000 women. Google reports it flagged 25 percent of “historically missed” interval cancers, meaning cancers that surface between routine screens, and, according to Google’s post, the studies showed it might be possible for specialists to spend up to 40 percent less time reading scans. The post does not say that saving was observed in a working screening service. This is an evaluation, not a rollout. The post does not say what the system was compared against, how many false alarms it raised, or whether any health service has adopted it, and a model that finds a quarter of missed cancers on a study population has not yet replaced a radiologist’s judgment.
The only item the post calls a clinical trial is with Stanford University School of Medicine. Google’s DeepVariant and DeepConsensus, two tools that clean up genome-sequencing data, were used to help find the genetic cause of disease in time-critical cases, reaching a diagnosis from whole genome sequencing in under eight hours. Google says the result earned a Guinness World Record and that it helped enable life-saving interventions. The post gives no patient count and no comparison with standard turnaround times, so the size of the benefit is not something a reader can check from this page.
Several items are real but narrower. Children’s Mercy Hospital ran DeepSomatic, Google’s tool for finding tumor mutations, on previously sequenced childhood leukemia cases. It confirmed every known variant and flagged 10 mutations that had been missed. Those were old samples, and the post does not say whether the 10 were verified afterward or how many cases were tested. DeepSomatic is publicly available, which makes it checkable by outside labs.
The flu forecast is the one result judged by an outside party. The CDC ranked a Google model first for predicting flu hospital admissions in the 2025 to 2026 season, ahead of 38 other systems. That is a live, public comparison, which is rarer in this field than a benchmark the developer chose.
The drug and biology items are earlier still. Researchers at the Institute for Global Health in Barcelona used AlphaFold to screen 30,000 compounds virtually against Chagas disease and identified two FDA-approved drugs that “could be repurposed.” That is a shortlist from a computer screen; the post does not describe lab or patient testing. The University of Exeter used AlphaGenome Atlas on data from over 54,000 people to flag variants tied to circulating proteins. Google’s AlphaProtein Novo designed enzymes in lab tests, with no clinical use claimed. In Congo, work with WHO AFRO and the country’s INRB used geospatial models to speed up Ebola risk analysis, and the post says responders could deploy mobile laboratories sooner, without saying how often that happened.
The post is also a consolidation. Google counts 90 peer-reviewed papers in five years and says its philanthropic arm has given more than $60 million in 2026 to groups applying these models, including Makerere University and the University of Pennsylvania. The weight sits on molecular biology and screening support rather than on any product sold to patients. Nothing here is a diagnosis or treatment anyone can request today, and health is the field where the distance between a published result and a deployed service is widest. The number to track is how many of these projects appear next year with a health system’s name on a live deployment, not another paper.
Based on Google’s blog post by Yossi Matias and Pushmeet Kohli of 12 October 2026.
