Mots-Clés
Agentic AI
Pharmacogenomics
Clinical Trials
GraphRAG
Ontology
Description
Offer: Master 2 / Engineering Internship - Agentic AI for pharmacogenomics: building and benchmarking an ontology-guided knowledge graph from clinical trial literature
Context: In oncology, a biomarker or a molecular signature is frequently reported as informative in one study and of no value in another. This lack of reproducibility stems primarily from the context in which each result was obtained: tumor type and stage, cohort composition, treatment received, clinical endpoint, and direction of the observed effect. This context remains buried in the text of publications, with no structured form that would allow it to be compared across studies.
The internship contributes to a processing pipeline that automatically transforms scientific publications into an ontology-guided knowledge graph. Each result is represented as a single relation that brings together the gene, the therapy, the cohort and the observed effect, rather than as a series of isolated facts whose connections would be lost. The resulting graph can then be queried by autonomous AI agents.
Positioning: The project lies at the state of the art of agentic AI applied to biomedicine, at the convergence of three recent advances: information extraction with large language models, agentic architectures (planning, tool use, reasoning loops), and ontology-guided GraphRAG, which grounds a model’s answers in a structured graph rather than in simple document retrieval. The internship applies this approach to a real clinical use case and evaluates it quantitatively, an aspect still rarely addressed in the literature.
Objectives:
- Optimize the existing processing pipeline: conversion and extraction of textual data from the literature, embedding models tailored to biomedical vocabulary, and knowledge augmentation using genomic annotation databases
- Continue building an evaluation benchmark from a corpus of pharmacogenomic publications reporting drug clinical trials in patients with kidney cancer
- Evaluate the complete pipeline on this benchmark by comparing the performance of several large language models.
Deliverables: Updated processing pipeline; completed and documented benchmark; quantitative model comparison report.
**References: **
- Saboo S. Always-On Memory Agent. Google Cloud Platform, generative-ai repository; 2026. https://github.com/GoogleCloudPlatform/generative-ai/tree/main/gemini/agents/always-on-memory-agent
- Wang J, Huang H, Ge X, Su J, Liu W, Lian S. OMD-GraphRAG: Enhancing GraphRAG with ontology-guided extraction, multi-dimensional clustering and dual-channel fusion. arXiv:2603.25152 2026
- Grabowski P, Alameen M, Bretones J, Cardell S, Carmona M, Edwards G, et al. Research Assistant: AstraZeneca’s agentic system for R&D. arXiv:2608.12395 2026
Candidate profile: Engineering degree or Master 2 in AI applied to life sciences, bioinformatics or data science. Python programming. Interest in agentic architectures and large language models; knowledge of ontologies is a plus; enthusiasm for biomedical data and a taste for rigorous benchmarking work.
Host laboratory: The internship will take place within the Genetics & Chemogenomics team of the Interdisciplinary Research Institute of Grenoble (CEA). He/she will be supervised by Guido UGUZZONI and Yoann CURE, experts in AI, and Christophe BATTAIL, expert in pharmacogenomics data, and will evolve in a multidisciplinary research environment with bioinformaticians and biologists.
Contract details: Six-month internship starting in February 2027
How to apply: Please send your CV and a cover letter to guido.uguzzoni@cea.fr and christophe.battail@cea.fr, using the subject line “M2 Internship – Agentic AI”.