Noetik trains transformers to match cancer patients to treatments
Noetik, an AI biotech startup, is building transformer-based models to identify which patients will respond to which cancer treatments — arguing that the 95% clinical trial failure rate is a patient-selection problem, not a drug-discovery problem — and recently closed a $50M deal with GSK.
Score breakdown
Developers and practitioners building AI for life sciences should note that Noetik's platform-licensing deal with GSK signals that pharma companies are beginning to pay for biotech AI as software infrastructure, not just as a path to drug co-development — validating a pure-tools business model in the space.
- 0195% of cancer drugs fail in clinical trials, which Noetik attributes to poor patient selection rather than poor drug development.
- 02Noetik uses three core data modalities: pathology (H&E), spatial transcriptomics, and genomic alterations.
- 03The company built its own data pipelines and tumor-processing infrastructure from scratch, spending roughly 18 months generating data before training models.
Noetik was founded on a contrarian thesis: the reason 95% of cancer drugs fail in clinical trials is not that the drugs themselves are ineffective, but that they are tested on the wrong patients. Co-founder and CEO Ron Alfa, a physician-scientist by training, and VP of AI Dan Bear, who has a background in neuroscience, computational neuroscience, and self-supervised learning, appeared on the Latent Space podcast to explain how the company is building AI models to solve this patient-matching problem. Their core argument is that some patients do respond to drugs that fail in broad trials — and that identifying the biological subtypes of patients who respond is the key to dramatically improving clinical success rates using treatments that already exist.
Their core data modalities include H&E pathology slides, spatial transcriptomics, and genomic alterations.
To build these models, Noetik generated its own high-quality, intentional datasets from scratch, sourcing human tumor samples and building custom processing pipelines over roughly 18 months before they could even begin training models. Their core data modalities include H&E pathology slides, spatial transcriptomics, and genomic alterations. The team employs self-supervised learning and deliberately avoids bias from electronic health records. They also use a tool called PerturbMap and in-vivo mouse models to validate predictions made by their human-data AI models.
GSK recently signed a $50M deal for Noetik's technology, which also includes an undisclosed long-term licensing agreement for their models. The deal is structured as a software platform license — not a drug co-development partnership — which distinguishes it from most big AI plays in biotech, where tool companies typically end up becoming drug companies. The Latent Space hosts note this signals a growing appetite among large pharma companies for biotech AI tools as a software category, alongside other developments in the space such as Boltz and Isomorphic.
Key facts
- 0195% of cancer drugs fail in clinical trials, which Noetik attributes to poor patient selection rather than poor drug development.
- 02Noetik uses three core data modalities: pathology (H&E), spatial transcriptomics, and genomic alterations.
- 03The company built its own data pipelines and tumor-processing infrastructure from scratch, spending roughly 18 months generating data before training models.
- 04Noetik employs self-supervised learning and avoids using electronic health records to reduce bias.
- 05A tool called PerturbMap and in-vivo mouse models are used to validate the company's AI predictions.
- 06GSK signed a $50M deal with Noetik that includes an undisclosed long-term licensing agreement for their models.
- 07The GSK deal is structured as a software platform license, not a drug development partnership — unusual in the biotech AI space.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Apr 22, 2026 · 11:07 UTC. How this works →