Delta Research Team at Inria

Dynamic-aware deep learning models

Our team aims to develop data-driven approaches capable of systematically learning the conformational heterogeneity of biological systems to solve fundamental challenges stem from static-based approaches. Such frameworks typically consists of multiple (task-specific) steps, including data generation, feature extraction, representation and encoding, and the architecture design. The task-in-hand has a direct impact on the method of choice for each step. Our recent efforts in this direction led to introducing DynaRepo and DynamicGT.

We established DynaRepo, the first official MDDB node in France (second largest in Europe), hosting >5.5 billion frames for >700 macromolecular complexes (>1 ms of simulation time), including protein-protein and protein-nucleic acid assemblies. Each complex was simulated in triplicate for 500 ns, with pre-computed analyses (e.g., RMSD, SASA, correlations) to facilitate AI training. Leveraging DynaRepo, we developed DynamicGT, the first dynamic-aware architecture for the prediction of protein binding sites. It is a novel GT architecture that captures the conformational heterogeneity of proteins for residue classification of binding sites. Inspired from the principles of cooperative graph neural networks, DynamicGT enables dynamic message passing between surface and core atoms, where a geometric transformer imposes strict message-passing constraints to predict binding site probabilities.

Enhanced data-driven protein design

We are actively contributing to developing innovative frameworks for synthetic biology and protein design. One of the aims of our team is to develop a framework for large-scale protein binder design with high specificity and affinity toward the target. Our effort in this direction so far led to the introduction of SURFACE-Bind database, an interactive web server that provides researchers with access to detailed data of the human surfaceome, facilitating the design of next-generation therapeutics. We introduced a large-scale computational resource coupled with an enhanced experimental framework for de novo protein binder design. Our approach targeted the entire human surfaceome (2,886 protein targets), a critical class of therapeutic targets, paving the way for new protein-based therapeutics. Leveraging a surface-based geometric convolutional neural network, we systematically evaluated the entire human surfaceome, identifying ~4,500 binding sites and targeted them by a library of 640,000 small fragments (seeds), equivalent to performing 3 billion protein-protein docking calculations. We provided detailed insights into the targetability of each predicted binding interface and suggested a high-quality set of binder seeds to support future target-specific protein design approaches.

Allosteric Communications in macromolecular complexes

Allostery relies on long-range communication networks that connect distant functional regions within macromolecular assemblies. While extensively studied in proteins, the role of nucleic acids in these communication pathways remains poorly understood. We developed ComPASS, a large-scale, dynamics-informed computational framework that integrates molecular dynamics simulations to construct weighted inter-residue communication networks spanning proteins and nucleic acids. By combining dynamical correlations, interactions, and spatial relationships, ComPASS reveals how signals propagate across complex assemblies. Our analyses uncover diverse mechanisms of allosteric signal transmission in protein–protein and protein–nucleic acid systems, providing new insights into the dynamic principles governing molecular function.

Machine learning-guided integrative structural biology & Host-pathogene interactions

The aim of the Delta team in this part is to perform research that can be applied to real, ongoing biological problems. Accordingly, we actively collaborate with national and international groups (mostly experimental biologists) to further evaluate our scientific developments in practice. The application part of our research follows three main strategies: First, antimicrobial resistance (AMR), toward elucidating the molecular complexes that are necessary for viral or bacterial survival and infection. Second, designing novel protein binders (peptides or proteins) with therapeutic potentials, and third, elucidating the dynamic behavior of large macromolecular complexes. More specifically, we study the mechanisms of survival and infection in Streptococcus pyogenes, The dynamic behavior of type IV pili across various bacterial species, and the interactions of Shigella IpaA with human focal adhesion proteins.