L'IA peut-elle prédire l'issue d'un essai clinique de médicament à partir de la seule structure moléculaire ?

Les progrès en chimie générative et en simulation permettent aux modèles de prévoir l'efficacité et les effets secondaires des médicaments à partir des données de composés. Tester cette capacité remet en question les délais traditionnels de découverte de médicaments et la dépendance aux essais cliniques humains, offrant un potentiel de réduction des coûts et d'accélération du développement des médicaments.

Background

Current artificial intelligence systems can analyze molecular structures to predict various properties and potential biological activities of compounds, which can be useful in the early stages of drug development. However, predicting the outcome of a clinical drug trial based on molecular structure alone remains a complex and unsolved task. Multiple factors influence trial outcomes, including pharmacokinetics, pharmacodynamics, and patient-specific variables such as genetics, comorbidities and concomitant medications. AI models, particularly those based on machine learning and deep learning algorithms, have shown promise in predicting certain aspects of drug behavior — such as efficacy and toxicity — from molecular structure when trained on large datasets of known drugs and their properties. These systems can identify patterns and suggest new compounds with desirable characteristics, but their accuracy depends heavily on the quality and breadth of training data. Despite progress, models that attempt to forecast full clinical trial outcomes using only molecular structure — without supplementary experimental data such as in vitro assay results, pharmacokinetic profiles, or early human safety data — have not yet achieved reliable performance. The primary obstacle is the complexity of human biology and the high inter-patient variability in drug response, which are difficult to capture from chemical structure alone. Ongoing research focuses on integrating multi-omics data, real-world clinical records, and mechanistic modeling to improve predictive accuracy. As of May 13, 2026, the National Institutes of Health reports that while AI is increasingly embedded in drug discovery workflows, its ability to predict the outcome of a clinical drug trial based solely on molecular structure remains unproven and is an active area of methodological development (Source: National Institutes of Health).