At the SophIA Summit, Stéphane Benzekry, head of the Inria–Inserm COMPO team, outlined the group’s work on mechanistic learning and artificial intelligence in immuno-oncology.
COMPO (COMPutational Pharmacology & Clinical Oncology) is a cross-disciplinary research team of mathematicians, pharmacologists and oncologists affiliated with Inria and Inserm and based at the Centre de Recherche sur le Cancer de Marseille (Inserm U1068). The team develops mechanistic simulations of biological and pharmacological processes alongside statistical and machine-learning tools to support therapeutic decision-making, personalize care and optimize clinical-trial design.
Linked to early-phase clinical trials through the Marseille CLIP2 center, COMPO has access to routine-care data and integrates mathematical modeling and AI to predict and guide cancer patient management. The team collaborates on most CLIP2-APHM research projects and contributes model-derived insights to the design of early-phase trials.
Benzekry launched the COMPO project in 2021 with three primary aims: deliver clinical decision tools to improve therapeutic management, better inform and design clinical trials, and test biological hypotheses using data from clinical oncology.
The team’s structure intentionally embeds mathematicians, pharmacists, oncologists and computer scientists to keep modeling grounded in clinical realities. Benzekry emphasized that clinical constraints—managing toxicity, treating systemic metastatic disease and addressing patient complexity—must shape modeling priorities. He gave examples where theoretical optimization of tumor shrinkage misses practical issues such as multiple metastatic sites and toxicity-driven dosing decisions, illustrating how clinician input reshapes model objectives.
COMPO is a key participant in France’s national digital pharmacological twin effort (PEPR/France 2030), which seeks to build multi-scale, causal, patient-specific simulators. These “digital twins” or DIGPHATs aim to simulate drug effects, dosing strategies and treatment scenarios but depend on richly annotated multimodal datasets—genomics, imaging and longitudinal immune or circulating biomarkers—that are difficult to obtain and harmonize.
Work currently focuses on early prototypes for specific diseases, notably advanced non-small-cell lung cancer, with the first functional mechanistic digital twins expected within three to five years for validation in real-world settings. COMPO is also experimenting with AI-agent systems in which large language models propose mechanistic equations, test them offline and iteratively refine them through agent-to-agent exchanges to automate parts of the modeling process.
COMPO leads biostatistics, machine learning and modeling for the PIONeeR RHU dataset, one of the largest longitudinal immunotherapy biomarker collections in advanced lung cancer. After multi-year negotiations and extensive preprocessing to handle missing values and multimodal measures, the team analyzed roughly 450 patients’ data. Initial efforts to predict immunotherapy response shifted toward predicting resistance because first-line treatment decisions will not change without phase III evidence.
The team identified an 18-marker resistance signature driven mainly by routine blood biomarkers, simpler than complex tissue-based signatures but not yet powerful enough to alter frontline care. This outcome underscores the importance of longitudinal modeling—tracking biomarker trajectories over time—to identify patients who might benefit from alternative therapies later in treatment and to provide the causal framework required for digital twin development.
Leave a Reply