From Small Clinical Trials to a Single Patient, Genialis Demonstrates the Power of Biology-First AI in Cancer Research
New ESMO data from three trial-sized cohorts, and a 25-month analysis of GitLab co-founder Sid Sijbrandij's osteosarcoma highlight the potential to extract meaningful insights from limited patient data
MADRID, October 23, 2026--Genialis, the therapeutic intelligence company, will present data at the European Society for Medical Oncology (ESMO) Congress 2026 showing that its biology-first models predict treatment response even with limited available patient data. Across three distinct cancer types and treatment regimens, the models demonstrated superior predictive performance in cohorts of fewer than 70 patients, addressing one of the fundamental challenges in developing predictive biomarkers for precision oncology.
Clinical trials are small data. Early-stage studies often enroll relatively small patient populations, creating challenges for conventional modeling approaches attempting to extract meaningful predictive signals. Yet drug developers must make critical decisions about pipeline advancement and patient selection criteria, often before independent validation data are available.
A tumor RNA-seq profile measures roughly 20,000 genes; with 50 patients, a conventional deep learning model risks identifying statistical noise rather than meaningful biological signals. Overcoming this "curse of dimensionality" has been central to Genialis' approach to predictive modeling since its founding.
The Genialis™ Supermodel is a large molecular model of cancer biology trained on hundreds of thousands of transcriptomic samples. The Supermodel transforms tens of thousands of individual gene-expression measurements from RNA sequencing (RNA-seq) into biologically meaningful features called "biomodules," each representing a fundamental aspect of cancer biology, including oncogenic signaling, immune activation, and DNA damage response. By reducing the complexity of molecular data into interpretable biological features before model training, Genialis can develop predictive biomarkers from relatively small patient populations while providing insight into the biological mechanisms driving treatment sensitivity and resistance.
The ESMO poster (abstract #8664) evaluates this approach across datasets from three different cancer types in a typical size of early clinical studies: pretreatment RNA-seq from anti-PD-1-treated melanoma (n=41), pembrolizumab-treated gastric cancer (n=45) and KRAS G12C-mutant non-small cell lung cancer (NSCLC) treated with sotorasib (n=66). Under strict held-out testing, biomodule models outperformed matched gene-expression models in every cohort. In NSCLC, biomodules predicted sotorasib response with an AUROC of 0.8, against 0.6 for the best gene-expression model, and predicted responders stayed on treatment a median of 337 days versus 178. Each prediction came with a biological rationale, with predicted NSCLC responders showing MAPK-addicted biology, while non-responders exhibited PI3K bypass programs.