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

Ron Alfa and Dan Baer of Noetic join the Latent Space podcast to discuss their mission of using AI to solve the 95% failure rate of cancer drug trials by fundamentally rethinking patient selection. They argue that most drug failures stem not from poor pharmacology, but from flawed patient cohort selection—driven by outdated models based on immortalized cell lines and animal studies that poorly reflect human biology. Noetic’s solution is to build a human-centric, multimodal foundation model trained on massive, high-quality data generated in-house, including spatial transcriptomics, H&E pathology, protein stains, and DNA data. Their approach uses self-supervised learning on over 100 million spatially resolved cells to discover biologically meaningful patient subtypes invisible to traditional biomarkers. The company has developed novel transformer architectures like Tario and OctoVirtualCell to simulate drug responses in silico, enabling virtual clinical trials and repurposing of existing drugs. A landmark $50M deal with GSK validates their model licensing strategy, marking a shift in pharma toward broad access to foundational AI models. The episode underscores the necessity of large-scale, intentional data generation—comparable to PDB or ImageNet—as the bedrock for transformative progress in AI for biology.