During the SophIA Summit, we caught the initiator behind COMPutational Pharmacology & Clinical Oncology group for a discussion on mechanistic learning and AI in immuno-oncology.
COMPO (COMPutational Pharmacology & Clinical Oncology) is a cross-disciplinary team combining mathematicians, pharmacologists, and oncologists. It is jointly affiliated with Inria and Inserm, based at the Centre de Recherche sur le Cancer de Marseille (Inserm U1068). The team develops computational tools and models using mechanistic methods—simulations of biological and pharmacological processes—and statistical or machine-learning methods. The results support therapeutic decision-making, personalize medicine, and optimize clinical trial design.
Through its link to early-phase clinical trials, notably via the Marseille CLIP2 center, COMPO has direct access to routine-care data. The team uses artificial intelligence and mathematical modeling to predict and guide the care of cancer patients. Collaborating on most CLIP2-APHM research projects, the COMPO team participates in the development of early-phase trials based on models derived from their research work.
During the SophIA Summit, we discussed mechanistic learning as a method of predicting response and survival, digital pharmacological twins (DIGPHAT), and AI in immuno-oncology with Stéphane Benzekry. As Head of the Inria-Inserm team at COMPO, Benzekry explained how he launched the project in 2021 to model data arising from clinical oncology with three aims: to bring clinical decision tools to enhance therapeutic management, to better inform and design clinical trials, and to test biological hypotheses.
COMPO’s Multidisciplinary Model & the Reality of Clinical Integration
COMPO was intentionally designed to unite oncologists, pharmacists, mathematicians, and computer scientists, allowing mechanistic modeling to remain tightly linked to clinical reality. For Benzekry, computational oncology must be grounded in the constraints and priorities of day-to-day oncology practice—namely toxicity management, systemic disease, and the complexity of real patients.
He illustrated this with concrete examples drawn from COMPO’s collaborations: when mathematicians imagine “optimizing” a treatment schedule, they often focus on shrinking a theoretical tumor. Clinicians, however, remind them that metastatic patients rarely have just one tumor, and that the primary challenge is systemic disease across multiple sites. Similarly, modelers initially aim to maximize efficacy, but clinicians emphasize that toxicity—not tumor volume—is what governs dosing decisions, treatment interruptions, and patient survival in routine practice. These examples underscore why embedding mathematicians within clinical environments fundamentally changes the modeling priorities.
“It’s important to me because I wanted to have something as close as possible to bedside, to a clinic, to a clinical practice,” Benzekry said.
He also emphasized the difficulty of building a shared technical and clinical language. A small coding mistake, such as a variable labeled “basic mass index” instead of “body mass index,” can slip through without domain supervision, reinforcing the necessity of integrated, cross-trained teams.
Mechanistic Learning, Digital Twins & Emerging AI Approaches
COMPO has become an essential element of France’s national digital pharmacological twin effort (PEPR/France 2030), which brings together pharmacology and computational teams to build multi-scale, causal, patient-specific simulators.
These “digital twins” are envisioned as virtual patients capable of simulating drug effects, dosing strategies, and treatment scenarios. But Benzekry notes that such systems require richly annotated multimodal datasets, including genomics, imaging, and longitudinal immune or circulating biomarkers. This kind of data is difficult to obtain and even harder to harmonize.
Current efforts focus on building early prototypes for specific diseases—particularly advanced non-small-cell lung cancer (NSCLC)—over the next three to five years. This timeline refers to the first functional versions of mechanistic digital twins tailored to a particular cancer type, informed by longitudinal biomarker data and ready for validation in real-world settings.
COMPO is also exploring the use of AI-agent systems, where large language models generate and iteratively refine mechanistic equations. In this framework, one agent proposes equations based on a language description of the biological problem; the equations are tested offline, and a second agent recommends modifications, leading to iterative improvement.
“We are implementing algorithms where we have AI agents talking to each other to automate the modeling process and, at the end of the day, we have something where we can see through the reasoning loops… this is completely emerging.”
Biomarker Discovery, Longitudinal Data, and Challenges in Translating Predictive Models to the Clinic
COMPO leads the biostatistics, machine learning, and modeling of the PIONeeR RHU dataset—one of the largest longitudinal immunotherapy biomarker datasets in advanced lung cancer. Negotiating data access took years, followed by extensive preprocessing to reconcile missing values, multimodal measurements, and the inherently limited sample size (approximately 450 patients).
Initially, the team sought to predict response to immunotherapy. Clinicians, however, clarified that first-line treatment decisions in lung cancer will not change without phase III evidence, regardless of predictive algorithms. As a result, COMPO shifted its focus to predicting resistance, which aligns more closely with second- and third-line clinical decision-making.
The team identified an 18-marker resistance signature, largely driven by routine blood biomarkers—far simpler than high-dimensional tissue or multiplex immunohistochemistry features. While encouraging, the predictive power remains insufficient to modify frontline care.
“It’s not sufficient to convince the physicians to actually switch treatment.”
This places renewed emphasis on longitudinal modeling: tracking biomarkers over time to predict which patients might benefit from alternative therapies later in their treatment course. Such longitudinal trajectories also provide the causal backbone needed for building digital twins.
COMPO’s collaboration with Adelis on cell-free DNA fragmentomics has yielded promising early results, now under review. These datasets will serve as key components of the developing digital twins. Regulatory, licensing, and data-access hurdles remain major barriers. Even COMPO’s long-standing dosing-adaptation software—developed decades earlier—is currently usable only in Marseille because of legacy licensing restrictions and modern CE-marking requirements. However, they remain hopeful about navigating the necessary administrative processes, so other clinics can benefit from the software.
