Ron Alfa

Ron Alfa is a physician-scientist and co-founder and CEO of Noetik, where he focuses on foundation models for cancer therapeutics. His central aim is to improve translation: moving treatments from early discovery and preclinical models into successful use in patients. In his February 2025 interview with Heather Couture, he argued that discovering a drug target and identifying the specific patient population likely to respond must go together.[2][3]

Alfa and a collaborator he named as Jacob spent roughly six months considering what data a model would need to understand biology, from tissue architecture down to the genome. Noetik then chose to build its initial dataset from scratch, brought in experienced histopathology staff and established a laboratory in South San Francisco. Alfa described this as a response to shortcomings they saw in public datasets and a way to retain control over data quality and future measurements.[2]

That decision included designing the dataset around machine-learning failure modes. Alfa described giving models multiple representations of patient samples processed in different experimental batches, to reduce shortcuts such as recognizing a sample’s origin or the scanner used to image it. In February 2025, he reported banking more than 1,000 primary non-small cell lung cancer samples and processing only a small fraction of each, leaving material available for additional assays as technology evolved.[2]

The initial stack combined four kinds of measurements. H&E imaging captured tissue architecture; multiplex immunofluorescence measured roughly 16 proteins on the same slide to identify tumor, immune and other cell types. Spatial transcriptomics came from an adjacent section four microns away, while whole-exome sequencing supplied tumor genotypes. Alfa’s rationale was that models should see cells in their spatial context and learn relationships across these biological layers.[2]

In the same interview, Alfa described masked image modeling as Noetik’s basic training approach, with choices about masking, modality layering, patient annotations and staining-batch metadata. He reported that its last model had 1.5 billion parameters and had trained on 256 H100 GPUs over weeks. Compute, he explained, was becoming a substantial budget consideration alongside the experimental costs traditionally associated with building a biology company.[2]

Alfa distinguished reconstructing an image accurately from learning useful biology: pixel-level reconstruction quality could miss relationships at the tissue or cell-interaction level. His evaluation approach began with established biological facts, then added measurable cross-modal tasks, such as inferring protein or T-cell markers from H&E. Noetik started with non-small cell lung cancer partly because its extensive trial history and known treatment-response patterns supplied biological checks against which to assess the models.[2]

His ambition extends beyond automating existing pathology tasks. As he put it in his February 2025 interview, “The goal here is to train models that can do what humans cannot do, that can understand biology that we haven't discovered yet.” He anticipated a further wave of precision oncology incorporating immune context and more complex genetics. Therapeutic discovery and better clinical-trial enrollment were near-term goals; helping clinicians choose subsequent treatments was a longer-term possibility contingent on success.[2]

Alfa authored Noetik’s January 8, 2026 announcement of a five-year foundation-model licensing partnership with GSK. The company publication archive identifies the licensed technology as OCTO-Virtual Cell foundation models for cancer therapeutics research and development.[5]

Abhishaike Mahajan’s May 2026 account of Perturb-MARS credits Alfa among a broader team of contributors, including Daniel Bear and Maxime Dhainaut. The system combines multiplexed mouse genetic-perturbation experiments with TARIO-2, a model trained on human cancer tissue, to infer corresponding human tumor states from mouse H&E images. Its bridge is morphological resemblance rather than direct conversion of mouse genes into human genes. Mahajan credits Dhainaut and colleagues with the original Perturb-Map work and Dhainaut with subsequent platform leadership; using experimental feedback for model post-training remained ongoing research.[6]

Before Noetik, Alfa spent six years at Recursion, which he recalled spanning its seed stage through the period after its IPO. That earlier work gave him experience building a machine-learning-enabled drug-discovery organization and integrating scientific and technical teams. He served as Senior Vice President, Head of Research and acting chief scientific officer, leading scientific organizations and portfolio strategy. His responsibilities included platform and portfolio development across rare disease, neuroscience, oncology and immunology, and advancing multiple therapeutic programs from discovery to clinical development.[2][3]

Alfa earned an MD and PhD at Stanford, with doctoral work in neuroscience, and an MA in history of medicine. His earlier education began at a Los Angeles community college before a summa cum laude bachelor’s degree in animal physiology and neurosciences at UC San Diego. He received a Paul & Daisy Soros Fellowship in 2009 to pursue his Stanford MD. The fellowship profile also credits his essay on the origins of the placebo effect with receiving the American Association of the History of Medicine’s William Osler Medal in 2008.[3][4]