SAVE THE DATE
Artificial Intelligence
in Biomedicine

Randomized controlled trials (RCTs) are the gold standard for evidence-based medicine, but their restrictive eligibility criteria can limit the generalizability of their findings to patients treated in routine clinical practice. This talk presents modern causal methods to bridge this gap and improve the translation of trial evidence into real-world care.
We first discuss transportability and generalization, focusing on how treatment effects estimated in trials can be extrapolated to target populations that differ in their characteristics. We then consider how these approaches can be extended to combine evidence from multiple trials and observational data sources. In this context, federated causal learning offers a promising alternative to traditional data pooling and meta-analysis, allowing treatment effects to be estimated across decentralized datasets without sharing individual-level data.
Together, causal generalization and federated learning provide a scalable framework for combining evidence while preserving causal interpretability and data privacy. These methods could contribute to more relevant clinical evidence, better treatment decisions, and ultimately more personalized care.

1. Federated Causal Inference: Multi-Centric ATE Estimation beyond Meta-Analysis. AISTAT 2025
2. A Unified Framework for the Transportability of Population-Level Causal Measures. Neurips 2025
3. Causal inference methods for combining randomized trials and observational studies: a review. Statistical Science. 2024.

Julie Josse is a senior researcher at Inria , where she leads the PreMeDICaL team in collaboration with Inserm. The team specializes in advancing precision medicine through causal learning and federated approaches that ensure medical data confidentiality, aiming to accelerate targeted therapy development and deploy decision-support algorithms with quantified prediction uncertainty.
Julie specializes in handling missing data, causal inference, and applying machine learning to multi-source, multi-modal health data, improving decision-making in areas such as respiratory disease, oncology, and fertility. She also led the Traumatrix project, creating AI-decision-support tools for ambulance trauma care optimization. A strong advocate for reproducible research, she actively contributes to open-source R software.
Before joining Inria in 2020, Julie was a professor at École Polytechnique, where she directed the Data Science for Business Master’s program in collaboration with HEC. She has held visiting positions at Stanford University and Google Brain Paris and has been recognized with honors such as the Inria–French Academy of Sciences Young Researchers Prize and a Marie Curie mobility grant.

The breakthroughs recognized by the 2024 Nobel Prize in Chemistry demonstrated that the biophysics of protein structure is encoded in evolutionary signals extracted from millions of sequences. This unlocked large-scale mapping of the protein universe, connecting poorly characterized proteins to known structures and functions at unprecedented scale, and revealed something more fundamental: that evolutionary pressure is itself a compression of physical constraints, rich enough that structural signal can be extracted in discrete form from sequence alone and used to design proteins with no counterpart in nature. But protein structure is only the beginning. Does the same logic extend to how proteins bind small molecules, interact with each other, respond to mutations, and can be engineered for new functions?

1. Durairaj, Janani, et al. "Uncovering new families and folds in the natural protein universe." Nature 622.7983 (2023): 646-653.
2. Škrinjar, Peter, et al. "Evaluating generalization in protein–ligand cofolding methods." Nature Structural & Molecular Biology (2026): 1-13.
3. Pantolini, Lorenzo, et al. "Rewriting protein alphabets with language models." bioRxiv (2025): 2025-11.
   
   

Jay is an Assistant Professor at the University of Lausanne, working on context-aware deep-learning methods for protein structure, molecular interactions, and protein design. Her group focuses on representation learning and generalizable modelling of biomolecular complexes, with the goal of developing more reliable methods for understanding and designing proteins.