[ENACT] Conception par IA prenant en compte la dynamique des modulateurs allostériques

Keywords

Deep learning, Protein design

Subject details

Allostery is a fundamental regulatory mechanism in biological systems, whereby a local event such as ligand binding induces functional effects at long distances within a macromolecule. It underlies many essential biological processes and, when dysregulated, contributes to numerous diseases. As a result, the design of allosteric modulators represents a major challenge in drug discovery, offering novel therapeutic strategies. However, the rational identification and design of allosteric modulators remain limited by the intrinsically dynamic, heterogeneous nature of macromolecular complexes, which is poorly captured by static structural approaches. The PhD project aims to address these limitations by developing a Deep learning architecture for the design of allosteric modulators, explicitly integrating conformational dynamics into molecular representation learning and generative modeling. The project combines topological representations of molecular systems, deep learning architectures, and large-scale molecular dynamics (MD) simulations to model how structural variability and long-range communication constrain ligand binding and modulation. The central principle of the project is to design allosteric modulators targeting a given binding site while accounting for its conformational variability. The first stage of the project focuses on data construction and curation. The work will rely on DynaRepo [1] , a large-scale database of macromolecular conformational dynamics developed within the host team, which provides direct access to high-quality MD trajectories of diverse macromolecular complexes. These data will be complemented with existing resources describing experimentally characterized allosteric sites, enabling annotation and supervised learning. The second stage will focus on learning meaningful representations of allosteric sites. Invariant and topology-aware representations [2] will be developed to capture both structural organization and conformational variability across ensembles of conformations. Recent dynamics-aware deep learning approaches developed by the team (DynamicGT [3]) will serve as a methodological foundation to model long-range communication and dynamic constraints directly from MD data. In the third stage, these learned representations will be integrated into a generative framework for small-molecule design. Our team has substantial knowledge to support this stage, highlighted in our recent work SURFACE-bind [4]. Novel deep learning models, including graph neural networks and topology-informed convolutional and attention mechanisms [5], will be developed to generate candidate allosteric modulators compatible with both structural and dynamic constraints of the target site. The models will be designed to be interpretable, robust, and transferable across different biological systems. Beyond its application to allosteric drug design, the project will deliver general-purpose, dynamics-aware data-driven methodologies applicable to other complex systems characterized by structural heterogeneity and long-range interactions. The developed representations and learning architectures are expected to be transferable to a wide range of problems in computational biology, chemistry, and physics, reinforcing ENACT's mission to advance foundational AI methods for scientific discovery. This PhD project is embedded in a strong interdisciplinary and collaborative environment at Inria/LORIA, at the interface of AI, computational chemistry, and structural biology. The project builds on established collaborations with LPCT (computational chemistry and molecular dynamics), CRM2 (structural biology and crystallography, enabling potential experimental validation), the European MDDB consortium (large-scale MD data infrastructures), EPFL (protein design), and Lund University (biological and immunological applications). These collaborations support both the scientific ambition of the project and the training of the PhD candidate.