🔬 Training Transformers to solve 95% failure rate of Cancer Trials — Ron Alfa & Daniel Bear, Noetik

The podcast begins with an exploration of the high failure rate in cancer drug trials, with around 95% not passing clinical trials. This failure is largely attributed to poor patient selection and the lack of detailed understanding of tumor biology at the individual level. The speakers emphasize that cancer is not a single disease but a collection of many subtypes, each requiring precise identification to improve treatment success rates.

Noetik has committed to comprehensive data collection, sourcing thousands of real human tumor samples and generating billions of multimodal images, including spatial transcriptomics and protein data. This extensive dataset feeds into large self-supervised models, creating what they call a ‘virtual cell’, which deeply understands cancer biology and can predict tumor responses to treatments. They stress the importance of building quality-controlled, purpose-designed datasets for training effective models.

The episode describes how Noetik’s foundation models, such as TARIO-2, use advanced transformer architectures trained on massive multimodal datasets to generate detailed spatial maps of tumors from standard histology slides. These models can distinguish subtypes of cancers, predict patient-specific drug responses, and guide clinical decision-making, potentially reducing the need for traditional drug development cycles.

A significant part of the discussion covers Noetik’s innovative approaches like in vivo tumor models with genetic barcoding in mice, which simulate human tumor biology more accurately than traditional cell lines. These models are used to validate predictions made by AI and facilitate understanding of how genetic perturbations influence tumor behavior, ultimately aiming to bridge the gap between preclinical and clinical results.

There is a highlighted shift in the biotech industry, with major pharma companies investing in platform licensing rather than just specific drugs. This includes broad use of foundational AI models trained on vast datasets to guide drug discovery, trial design, and patient cohort selection. The discussion emphasizes the importance of scale, quality data, and strategic model development in making AI-driven precision medicine a reality.

The speakers share insights on how they built their data sets—from sourcing tumor samples to designing experiments to control for batch effects. They stress the importance of generating purposefully designed, high-quality data at scale, and the challenges faced in collecting enough diverse datasets to train reliable models. Lessons include patience, iteration, and a focus on problem-specific data collection.