Avancées dans l’apprentissage de la représentation graphique et leurs applications en biologie computationnelle

Graphs are a natural and flexible framework for modeling relationships and interactions across entities in various domains, including social networks, chemistry, and biology. This has fueled significant interest in graph representation learning, which focuses on learning informative representations and making predictions for graph-structured data. However, the inherently non-Euclidean nature of graph data poses substantial challenges for traditional machine learning models. Graph Neural Networks (GNNs) address these challenges by leveraging the message-passing scheme to learn node and graph representations. Despite their success, existing GNNs face significant difficulties in fully exploiting the information in graph data. In this thesis, we aim to overcome several limitations of GNNs, including their inability to capture synergistic interactions among nodes, difficulties in leveraging unlabeled graph data effectively, and other issues, such as the lack of interpretability. Furthermore, we develop new graph representation learning techniques to solve critical problems in computational biology, including cancer gene prediction and protein function prediction.In the first part of the thesis, we show that GNNs compute messages from each neighboring node independently, which limits their ability to capture complex, synergistic interactions among the nodes.To overcome this limitation, we propose a new GNN layer, which uses a self-attention mechanism and a recurrent neural network as an aggregator, thus considering the dependencies between the nodes.Secondly, we improve the performance of Graph Attention Networks (GATs) by introducing a novel loss function designed to guide the model in learning higher attention scores between nodes belonging to the same class.While the above methods have demonstrated strong performance in supervised settings, their reliance on labeled data limits their effectiveness in domains where annotations are scarce or expensive to obtain. One such domain is 3D protein graphs, where vast amounts of structural data exist, but only a small fraction is labeled with functional annotations.To this end, we propose a new geometric self-supervised pretraining framework for 3D protein graphs that predicts the subgraph distances relative to the overall geometric centroid of the protein graph.Moreover, we analyze the structural properties encoded by GNNs and reveal a strong connection between node embeddings and the number of random walks originating from nodes.Our findings connect to the over-squashing phenomenon, in which distant node information is compressed and lost, as well as to the stability of GNNs.In the second half of the thesis, we focus on applications of graph representation learning to computational biology.Cancer gene prediction is a critical task in computational biology, as the identification of cancer-associated genes can reveal insights into the underlying mechanisms of cancer and help in identifying new therapeutic targets.To this end, we propose a new framework based on GNNs to identify cancer genes, which leverages multiomics data and multiple gene and protein interaction networks.Our approach achieves state-of-the-art performance in predicting cancer driver genes while being interpretable.Furthermore, we tackle the problem of protein function prediction with a novel approach that combines GNNs and Large Language Models. Our framework integrates protein sequence and structure information to predict protein function in a text format, in contrast with previous classification models.