Mots-Clés
Deep learning
Graph Neural Networks
Protein-Protein interactions
Description
Project Description
This internship is funded for six months by the grant from Programme Inria Quadrant (PIQ). The main goal is to develop a graph neural network architecture for the prediction of binding affinity within protein-protein complexes. The internship student will be supervised by Yasaman Karami (Chargee de recherche, Inria) with expertise in proteins conformational dynamics and allostery, and will be hosted in the Delta team within the Inria center at the Universite de Lorraine. Our team consists of two permanent researchers with several PhD and postdoc members, and is expected to grow by hiring new members. The team provides a multidisciplinary and international environment, and benefits from experts in structural bioinformatics, as well as in computer science and deep learning. Our main goal is to develop deep learning models, to study, and predict protein structure, interactions, function and to further design synthetic molecules. The team has access to computational resources, including efficient GPUs and CPUs, from different cluster centers including Grid5000, Jean Zay, etc.
Scientific context
Protein–protein interactions govern essential biological processes, from transcription and translation to viral replication. The function of macromolecular complexes is directly linked to the arrangement of their atoms in 3D and dynamics. Therefore, characterizing the structure, dynamics and conformational changes of biomolecules can help understand the molecular mechanisms of underlying diseases. Yet, AI models for different prediction tasks are limited by the scarcity of high-quality dynamic data, as nearly all current datasets provide only static structures. Recently, we teamed up with the European initiative of MDDB and developed the first MDDB node is France to host large scale molecular dynamics (MD) trajectories for ~700 macromolecular complexes, DynaRepo [1]. Taking advantage of DynaRepo data, we took major steps in learning conformational dynamics by proposing DynamicGT, a novel architecture that combines cooperative graph neural networks with a graph transformer, to predict binding sites [2]. We showed that through learning dynamic features, we could improve the accuracy of predictions. We have also developed ComPASS, a large-scale computational method designed to study communication networks in macromolecular complexes [3]. ComPASS has been applied to different biological systems, facilitating the interpretation of the conformational dynamics. The main goal of this internship is to take advantage of these recent developments in the team (DynaRepo and DynamicGT), and propose a method for predicting binding affinity values. The main idea is to take another major step, taking advantage of the recent developments of AI and propose a novel approach to uncover distinct mechanisms in macromolecular systems. The internship student will directly work with the PhD and Postdocs of the team working on relatively similar projects.
[1] Mokhtari O, Bignong E, Khakzad H, Karami Y. DynaRepo: The repository of macromolecular conformational dynamics. bioRxiv, 2025.
[2] Mokhtari O, Grudinin S, Karami Y, Khakzad H. DynamicGT: a dynamic-aware geometric transformer model to predict protein binding interfaces in flexible and disordered regions. bioRxiv, 2025.
[3] Bheemireddy S, Gonzalez-Aleman R, Bignon E, Karami Y. Communication pathway analysis within proteinnucleic acid complexes. J. Chem. Theory Comput, 2025, 21, 17, 8255–8266.
Candidate Profile
- Training in Computer Science, or Bioinformatics
- Proficiency in Python and good coding practices is mandatory
- Experience in deep learning (PyTorch) is mandatory*
- Knowledge in protein biochemistry
- Ability to work in a team
- Excellent oral and written English skills
*Applications with no computer science/deep learning background will not be considered.