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

What It Covers

Noetik co-founders Ron Alfa and Daniel Bear explain how 90-95% of cancer drug trial failures stem from poor patient selection rather than bad pharmacology. They describe building multimodal foundation models trained on spatially-resolved human tumor data — combining H&E pathology, multiplex protein imaging, and 20,000-gene spatial transcriptomics — to match drugs to the right patient subpopulations.

Key Questions Answered

• •Patient selection as root cause: 90-95% of cancer drugs fail in clinical trials not because of poor pharmacology or target selection, but because trials enroll the wrong patients. Drugs often work in a subset of patients, but without models to identify that subset upfront, trials run on broad populations, diluting signal and leading to cancellation of molecules that could help specific subgroups.
• •Spatial transcriptomics as training data: Noetik generates multimodal tissue data stacking H&E pathology images, multiplex fluorescence protein stains, and spatial transcriptomics capturing up to 20,000 genes per spatial location. Each data point functions like a 20,000-channel image rather than a standard RGB image. Over 100 million spatially-resolved cells have been generated, representing at least one order of magnitude more paired data than any known public dataset.
• •H&E as universal inference input: Despite training on expensive multimodal data, Noetik's models run inference using only standard H&E pathology slides at deployment. Because H&E is collected for virtually every cancer patient globally, this allows retrospective analysis of existing trial cohorts — splitting responders from non-responders using images already on file, without requiring new data collection from past participants.
• •Autoregressive scaling for spatial biology: Noetik's Tario model applies next-token autoregressive training — the same objective scaling LLMs — to spatial transcriptomics data. Larger models only outperform smaller ones at longer context lengths, meaning the model must observe larger tissue regions simultaneously to capture nonlinear spatial patterns. This mirrors LLM scaling behavior and suggests tissue context length is a key variable for biological foundation model performance.
• •In vivo perturbation validation via barcoded mouse tumors: To validate human model predictions without relying on cell lines, Noetik uses a multiplexed CRISPR knockout platform injecting ~100 barcoded cancer cell variants into single mice, producing hundreds of genetically distinct tumors per animal. Human-trained models are then inferenced directly on mouse H&E, and predictions about immune infiltration and tumor phenotype are validated against known pathway biology across multiple gene knockouts simultaneously.
• •Data generation must precede model development: Noetik spent roughly 18 months generating data before training any functional model. The lesson for AI biotech startups: design datasets around the specific ML problem first, control for batch effects by distributing each patient sample across multiple slides and arrays, and reach a critical data threshold before expecting meaningful model signal. Dropping to 10-40% of training data causes substantial generalization failure, particularly across cancer types not seen during training.