Google Expands Work with Scientists on AI-Powered Disease Detection and Treatment

Google has published a post on its corporate blog about how it is working with scientists and research partners to make diseases more detectable, treatable and preventable. The ambition is large, but the available version of the announcement contains no numbers: no model names, no metrics, no list of clinical partners. So we will go through what is stated directly and compare it with what Google has already done in biology and medicine. For scale: the AlphaFold database holds more than 200 million predicted protein structures, and Demis Hassabis and John Jumper received the 2024 Nobel Prize in Chemistry for that work.

What the announcement says

The wording is simple: the company works with scientists to improve human health in three areas. The first is disease detection, meaning diagnostics and screening. The second is treatment, from understanding the biology of a disease to finding therapeutic hypotheses. The third is prevention. Without specifics, this is more a statement of direction than a product release. The text contains not a single metric that would let you compare the new work with earlier efforts.

What the announcement builds on

Google has accumulated a substantial body of work in healthcare and biology. These are the projects most often mentioned in this context:

• AlphaFold: protein structure prediction, with an open database of more than 200 million entries;
• Med-PaLM 2: in 2023 the model scored 86.5% on questions in the format of the USMLE medical licensing exam;
• AMIE: a research conversational system for diagnostic conversations with patients;
• AlphaGenome: a model that analyzes regulatory regions of DNA on sequences of up to 1 million base pairs;
• Cell2Sentence-Scale: a family of models for analyzing single-cell data, developed jointly with Yale University, with the largest version at 27 billion parameters;
• ARDA: screening for diabetic retinopathy from fundus images.

It is unclear which of these will feature in the new announcement. But the logic is visible: structural biology, genomics, cell data and clinical diagnostics form a chain from laboratory hypothesis to patient.

Where such projects usually hit a wall

A model that does well on a test is not yet working in a clinic. Three barriers lie between a benchmark and a treatment decision. The first is data: medical datasets are fragmented, skewed by population and locked behind privacy rules. The second is validation: it requires prospective studies, not a retrospective review of archives. The third is regulation: the FDA, the EMA and their counterparts assess software that affects diagnosis in different ways.

There is also an organizational side. The phrase "in collaboration with scientists" means that some of the results will be published in peer-reviewed journals and will gain independent verification. That is a good sign, but it does not replace clinical data. A drug candidate found by a model in simulation still has to pass preclinical testing and three phases of trials, which takes years and involves a high attrition rate.

How to read announcements like this

Until the blog provides details, it is more useful to ask questions than to draw conclusions. If you work in medtech or pharma, or invest in the sector, check four things when details are released:

• whether there is a peer-reviewed publication describing the method and its limitations;
• which populations and which countries the model was tested on;
• whether weights, an API or data are available for independent reproduction;
• what the regulatory status is: a research tool, approved software or a clinical pilot.

For developers and startups, the main practical question is whether Google will open access to the models and data. If it does, an ecosystem will quickly grow around them, as happened with AlphaFold. If it does not, the announcement will remain a demonstration of the company's interest in the health market. We will return to the topic when documents and numbers appear.