MADRID, Spain — October 23, 2026
Genialis, a therapeutic intelligence company developing artificial intelligence models for precision oncology, announced findings demonstrating the potential of its biology-first AI approach to predict cancer treatment response using limited patient data. The results, scheduled for presentation at the European Society for Medical Oncology (ESMO) Congress 2026, span three cancer cohorts involving fewer than 70 patients each and a retrospective analysis of tumor samples from GitLab co-founder Sid Sijbrandij, who has osteosarcoma. Genialis reported that its models outperformed matched gene-expression models across all three clinical datasets under held-out testing. The findings highlight an approach designed to address a key challenge in precision oncology: identifying reliable predictive biomarkers from small clinical studies, where conventional modeling can struggle to distinguish meaningful biological signals from statistical noise. The company’s Genialis Supermodel transforms RNA sequencing (RNA-seq) measurements into interpretable biological features, enabling researchers to investigate treatment sensitivity and resistance while developing hypotheses for patient selection. Together, the cohort analyses and single-patient case study illustrate how biology-first modeling may support research across small clinical trials and individualized cancer investigations.
Genialis Supermodel Translates Gene Expression Into Biological Features
The Genialis Supermodel is a molecular model of cancer biology trained on hundreds of thousands of transcriptomic samples. It converts measurements from tens of thousands of genes into interpretable features called biomodules, which represent biological processes such as oncogenic signaling, immune activation and DNA damage response. This approach is intended to reduce the complexity of high-dimensional molecular data before predictive model training. Genialis noted that conventional deep-learning models may struggle when analyzing approximately 20,000 genes across only 50 patients, increasing the risk of learning statistical noise rather than meaningful patterns. The company’s ESMO presentation, abstract #8664, evaluates biomodule-based models using pretreatment RNA-seq data from three cohorts: 41 patients with melanoma treated with anti-PD-1 therapy, 45 patients with gastric cancer treated with pembrolizumab, and 66 patients with KRAS G12C-mutant non-small cell lung cancer (NSCLC) treated with sotorasib. Under held-out testing, the biomodule models outperformed their matched gene-expression models in all three cohorts. In the NSCLC cohort, the model achieved an area under the receiver operating characteristic curve (AUROC) of 0.8, compared with 0.6 for the best gene-expression model. Predicted responders also remained on treatment for a median of 337 days, compared with 178 days among the comparison group reported by the company. The model’s biological interpretations associated predicted responders with MAPK-dependent biology, while predicted non-responders showed PI3K bypass programs. These findings provide an initial demonstration of predictive performance in small datasets, rather than proof of clinical utility across broader patient populations.
Single-Patient Analysis Tracks Osteosarcoma Biology Over 25 Months
Genialis also retrospectively analyzed tumor RNA-seq data from Sid Sijbrandij, who was diagnosed with osteosarcoma in November 2022. After his cancer returned and standard treatment options were exhausted, Sijbrandij assembled a team to investigate personalized experimental treatment approaches, combining extensive diagnostics with the collection and public sharing of clinical and molecular data through osteosarc.com. The dataset included more than 30 terabytes of information. Genialis examined four tumor RNA-seq libraries generated over 25 months by three clinical laboratories using different specimens. According to the company, the Supermodel enabled consistent biological comparisons across the four samples, including in circumstances where raw gene-expression measurements were not directly comparable. By connecting each sample’s biological features with therapeutic hypotheses, the analysis illustrated how researchers could revisit those hypotheses as additional samples became available. Genialis said this approach may become increasingly relevant as precision oncology focuses on smaller, more precisely defined patient groups and potentially individual patients. Rafael Rosengarten, Ph.D., Genialis chief executive officer and co-founder, said the ability to compare tumor biology across samples collected over time could create new opportunities to investigate disease and develop therapeutic hypotheses. The analysis is retrospective and demonstrates a research application; it does not establish that the model selected an effective treatment or improved Sijbrandij’s clinical outcome.
Advertisement
ESMO Presentation Highlights AI Applications in Drug Development
Genialis will present the cohort findings in a poster titled “Biology-first modeling enables interpretable patient-selection strategies from small oncology cohorts” at ESMO Congress 2026 in Madrid, Spain. The poster, abstract #8664 and presentation number 152P, is scheduled for the AI for clinical research and drug development category on Sunday, October 25, from 12:00 to 12:45 p.m. Central European Time. The congress runs from October 23–27, 2026. Genialis develops and validates AI models intended to predict patient responses to targeted inhibitors, immunotherapies and other emerging therapeutic classes. The company works with pharmaceutical and diagnostics partners and is headquartered in Houston, Boston and Ljubljana. Its biology-first approach seeks to combine predictive performance with interpretable biological insights, helping researchers investigate why tumors may respond or resist particular treatments. The latest findings suggest that this strategy can be applied to both small clinical cohorts and longitudinal molecular data from an individual patient. Further validation will be needed to determine how reliably the models generalize to independent populations and whether their predictions can improve treatment selection or clinical outcomes in prospective settings.
Source: Genialis release
