From 30 Million Neuronal Images to a Starting Point for Drug Discovery: Recursion and Genentech Advance Their First AI-Identified Target

The most common stumbling block in neuroscience drug development is often not the inability to find candidate molecules, but whether the right disease biology was selected from the outset. Recursion and Genentech have now advanced a novel target identified using large-scale neuronal data and AI models into a joint early-stage small-molecule discovery program, creating the first neuroscience R&D project for this approach of “mapping biology first, then generating hypotheses.”

This neuronal phenotypic map was built on large-scale cell production and genome-wide experiments. The partners produced more than one trillion neurons from human induced pluripotent stem cells, performed knockouts and other perturbations targeting more than 17,000 genes, and captured more than 30 million cell images. Specially trained foundation models then compared the high-dimensional phenotypes caused by different gene perturbations, searching for genes that are biologically linked to known drivers of neurological diseases but have not yet been sufficiently explored.

Model ranking was only the starting point. Recursion said the team first examined whether candidate targets did indeed affect the predicted biological pathways, then assessed whether manipulating the targets could alter the health, metabolism, and electrophysiological function of human neurons, and finally tested whether they were sufficient to change disease-related phenotypes. After the first target passed these three levels of validation, Genentech exercised its “validated target” option and brought it into joint development.

The next step is the generation and validation of small-molecule hits. The companies will attempt to design compounds capable of modulating the target, then progressively optimize their potency, selectivity, and drug-like properties. This means the program has moved beyond computational prediction alone, but multiple hurdles remain before reaching a drug candidate, proof of concept in animals, or even human trials. Drugs for the nervous system must also address challenges including brain exposure, long-term safety, and clinical endpoints.

This development also reflects a shift in how AI drug discovery is evaluated: data scale alone is insufficient to demonstrate value. The key question is whether relationships proposed by a model can be reproduced through independent, layered experiments and ultimately translated into a mechanism that can be modulated by a drug. A reusable phenotypic map may generate multiple research hypotheses at once, but each target must still undergo individual testing using conventional pharmacology and disease models.

The largest information gap at present is that the companies have not disclosed the target’s name, the specific disease, experimental effect sizes, or validation models, nor have they released compound data. Current claims come primarily from Recursion and lack corroboration from external research reports or peer-reviewed data on the same development. This is therefore an early R&D milestone with methodological significance, but it cannot yet be regarded as evidence that AI has improved the clinical success rate of drugs for neurological diseases.