Mots-Clés
Deep learning
computational biology
Protein Representation Learning
Proteins
Description
From pairs to networks: Learning protein-protein interaction graphs
Project description
Drug discovery faces significant challenges, largely due to a limited understanding of cellular mechanisms, which are primarily governed by protein-protein interactions. This knowledge gap contributes to a discovery process that is costly, time-consuming, and frequently unsuccessful.
Recent advances in deep learning and computational biology, such as Boltz2, AlphaFold2/3, RoseTTAFold2/3, and protein language models, have demonstrated remarkable potential in learning biologically meaningful representations from protein sequences and structures, enabling high-quality modeling of protein structures, complexes, and biomolecular interactions. Collectively, these advances have laid a strong foundation for protein representation learning, leading to numerous tools that achieve promising results in predicting whether and how proteins interact. However, current representation learning almost exclusively focuses on isolated pairwise protein-protein interactions (PPI), whereas in reality, biology operates through complex, interconnected PPI networks.
This project aims to move beyond pairwise prediction toward network-level mapping of protein interactions. It will computationally predict multi-protein complexes and signaling pathways by learning protein representations informed by cellular context, biological and structural constraints, and biophysical priors.
As a computational case study, the proposed approach will be applied to drug repurposing and drug safety assessment by predicting the therapeutic and adverse effects of drugs on protein interaction networks. The goal is an in-depth identification of novel therapeutic indications for existing drugs and potential off-target interactions early on, saving significant time and research costs.
Research objectives
- Collect and preprocess publicly available protein interaction datasets, including functional complexes and signaling pathways.
- Evaluate existing deep learning models in terms of predictive performance.
- Develop a framework for learning representations of multi-protein interactions.
- Develop, train, and validate a deep learning model for reconstruction of protein interaction networks, optimizing both topological and biologically relevant metrics.
- Validate model predictions for multiple applications, including case studies of drug repurposing and prediction of off-target interactions.
This project will be supervised by Yasser Mohseni and hosted by the Dynamics of Structures and Interactions of Macromolecules in Biology (DSIMB) team within the Integrated Biology of the Red Blood Cell (BIGR) lab, located at Hôpital Necker, 75015 Paris, France.
Benefits
- 6-month paid internship (stipend provided).
- Access to high-performance computing resources.
- Opportunity to collaborate with an interdisciplinary team of experts in computational biology and deep learning.
- Subsidized meals.
Candidate profile
- Applicants must be enrolled in a Master’s program in Computer Science, Bioinformatics, or a related field.
- Strong programming skills in Python and solid coding practices.
- Practical experience with deep learning frameworks (e.g., PyTorch or TensorFlow).
- Good knowledge of computational biology.
- Ability to work effectively in a collaborative, interdisciplinary environment.
- Strong written and oral communication skills in English.
Application
Please send your CV and a motivation letter to
yasser.mohseni-behbahani@u-paris.fr
yasser.mohseni-behbahani@inserm.fr