Description du projet
Des modèles de fondation pour des prévisions fiables du cancer
Le diagnostic moléculaire joue un rôle crucial dans la médecine personnalisée. Les modèles d’IA actuels ont toutefois besoin d’assistance pour apprendre à partir des profils moléculaires des patients et faire des prévisions, en raison de la nature complexe de la biologie moléculaire des maladies et du nombre limité de données d’apprentissage. Le projet FoundationDX, financé parle CER, comble cette lacune en développant des modèles de fondation à partir des données biomoléculaires disponibles pour les tissus sains et malades. Grâce à l’apprentissage auto-supervisé, un moteur essentiel de l’IA, le projet se propose de créer une représentation complète de la biologie cellulaire sans nécessiter de données détaillées sur les patients. Cette approche permettra de prévoir de manière fiable les sous-types de cancer et le pronostic. Le projet s’efforce de proposer des solutions puissantes d’apprentissage automatique à des problèmes de diagnostic moléculaire jusqu’ici difficiles à résoudre.
Objectif
Molecular diagnostics is crucial in fulfilling the promise of personalized medicine. While we are amidst an AI revolution, current machine learning models (ML) struggle to effectively learn from molecular (‘omics’) patient profiles and fail to make robust predictions. Perhaps this is not a surprise. After all, molecular disease biology is immensely complex, and we ask ML models to predict such complicated things as patient prognosis, without them ‘knowing’ anything about molecular biology and based on limited training data.
To address this, I will create foundation models on top of the vast troves of available biomolecular data, such as multi-omics profiles in healthy and diseased tissues, high-resolution single-cell data and biological knowledge graphs. This unique approach is driven by self-supervised learning (SSL), an important driver of AI, which offers the opportunity to learn a comprehensive representation of the multimodal biology of the cell – without the need for well-annotated patient data.
Starting from this strong basis, the FoundationDX model can then reliably predict cancer subtype or prognosis as it no longer needs to start from scratch on too high-dimensional, too low sample-size datasets. Effectively, we give our systems biological ‘common sense’, foregoing the need for millions of labeled training samples. This uniquely enables us to address one of the most clinically relevant questions: which treatment is best for the patient?
The FoundationDX research program is designed to deliver key insights into how the SSL revolution can be used to drive progress in the field of molecular diagnostics. It contains a ‘clinical-grade’ benchmarking module and solves three urgent diagnostic challenges, including noninvasive subtyping of pediatric brain cancer. The time for powerful, robust and generalizable, knowledge-aware machine learning solutions to previously intractable molecular diagnostics problems has come. FoundationDX aims to deliver this.
Champ scientifique
- natural sciencescomputer and information sciencesknowledge engineering
- social sciencespolitical sciencespolitical transitionsrevolutions
- medical and health scienceshealth sciencespersonalized medicine
- natural sciencescomputer and information sciencesartificial intelligencemachine learning
- natural sciencesbiological sciencesmolecular biology
Mots‑clés
Programme(s)
- HORIZON.1.1 - European Research Council (ERC) Main Programme
Régime de financement
HORIZON-ERC - HORIZON ERC GrantsInstitution d’accueil
3584 CX Utrecht
Pays-Bas