UNLOCK 2026 Unlocking Clinical Response Prediction With Multimodal World Models Ron Alfa, NOETIK

Author: Ron Alfa, NOETIK Collection: UNLOCK-2026-Sessions Source: Unlocking Clinical Response Prediction With Multimodal World Models

Summary

Ron Alfa, representing NOETIK, presents the company's approach to building foundation models of human biology with the explicit goal of predicting clinical drug response. As an MD by training, Alfa opens with a deliberately contrarian position: while he uses language models extensively, he does not believe language models alone will solve the core problem of discovering new biology — but he does believe they will fully replace doctors. The fundamental problem NOETIK addresses is that the pharmaceutical industry spends approximately $300 billion per year on drug development (with $200 billion focused specifically on the question of "which drugs will work in patients"), and 90–95% of drugs fail — an outcome the industry has normalized but that NOETIK considers unnecessary.

NOETIK's thesis is that this failure rate stems from a fundamental disconnect between animal/cell model data generated early in drug discovery and actual human biology. The solution is to train multimodal foundation models on real human patient data from three paired modalities: H&E (hematoxylin and eosin) histopathology images (tissue), protein expression imaging (cell-level resolution), and spatial transcriptomics (RNA expression with spatial coordinates for every cell). NOETIK's South San Francisco lab has been running one of the world's largest spatial transcriptomics labs for approximately three years, generating paired multimodal data from 4,000 patients across approximately 12 cancer types, scaling to 16,000–20,000 patients by end of 2026 across all solid tumor oncology.

The talk's headline result is their Target 2 model: an autoregressive model trained on paired H&E and spatial transcriptomics data that can infer a full 19,000-gene transcriptomic panel from an H&E image alone. In a zero-shot validation on a failed phase 2 clinical trial (8% overall response rate, ORR), NOETIK's model identified responder patients whose biology aligned with the training data — achieving a 56% ORR in the NOETIK-defined patient subpopulation without any training on trial data. In January 2026, NOETIK announced a partnership with GSK focused on their virtual cell model. Scaling laws hold for biology: 200M-to-4B parameter models show continued improvement in prediction accuracy as both model size and context length increase, with no sign of plateau.