Noetik trains transformers to fix cancer trial matching problem
Noetik's TARIO-2, an autoregressive transformer trained on large-scale tumor spatial transcriptomics data, aims to solve the 95% clinical trial failure rate by better matching patients to treatments — and GSK has signed a $50M licensing deal for the technology.
Score breakdown
Practitioners building AI tools for biotech should note that TARIO-2's ability to extract rich tumor biology from a universally available assay (H&E) — and GSK's willingness to license it as a platform — signals a viable commercial path for AI software in drug development beyond the typical pivot to in-house drug discovery.
- 0195% of cancer treatments fail to pass clinical trials, which Noetik frames as a patient-matching problem rather than a drug efficacy problem.
- 02Noetik's TARIO-2 is an autoregressive transformer trained on one of the largest tumor spatial transcriptomics datasets in the world.
- 03TARIO-2 predicts an ~19,000-gene spatial map from the H&E assay that nearly every cancer patient already receives.
A Latent Space podcast episode hosted by Brandon Anderson and RJ Honicky features Noetik co-founders Ron Alfa and Daniel Bear, who contend that the widely cited 95% cancer clinical trial failure rate is not primarily a drug efficacy problem — it's a patient-matching problem. Their thesis is that many treatments that "failed" in trials actually do work, but only for patients with specific tumor biologies that current standard-of-care diagnostics cannot identify. If the right patients could be matched to the right treatments, success rates could improve dramatically using therapies that already exist.
Noetik's flagship model, TARIO-2, is an autoregressive transformer trained on one of the largest collections of tumor spatial transcriptomics datasets in the world.
Noetik's flagship model, TARIO-2, is an autoregressive transformer trained on one of the largest collections of tumor spatial transcriptomics datasets in the world. Whole-plex spatial transcriptomics is described as the richest method for reading a tumor, yet approximately 0% of cancer patients in standard care ever receive one. TARIO-2 addresses this gap by predicting an ~19,000-gene spatial map directly from the H&E histology assay that virtually every cancer patient already undergoes. To build this capability, Noetik spent nearly two years collecting thousands of actual human tumors and assembling a large multimodal dataset of hundreds of millions of images.
The commercial traction is notable: GSK signed a $50M deal that includes an undisclosed long-term licensing arrangement for Noetik's models. Crucially, this is structured as a software licensing deal rather than an in-house drug development partnership — a distinction the episode highlights as meaningful, since most big AI plays in biotech tend to result in tools companies pivoting to become drug companies. The episode frames this alongside other software-focused biotech tools (referencing Boltz and Isomorphic as comparisons) as evidence that pharma's appetite for AI platform licensing is finally growing.
Key facts
- 0195% of cancer treatments fail to pass clinical trials, which Noetik frames as a patient-matching problem rather than a drug efficacy problem.
- 02Noetik's TARIO-2 is an autoregressive transformer trained on one of the largest tumor spatial transcriptomics datasets in the world.
- 03TARIO-2 predicts an ~19,000-gene spatial map from the H&E assay that nearly every cancer patient already receives.
- 04Whole-plex spatial transcriptomics is described as the richest way to read a tumor, yet ~0% of standard-care patients receive one.