MOTIVATION: Graph neural network (GNN) models have emerged in many fields and notably for biological networks constituted by genes or proteins and their interactions. The majority of enrichment study methods apply over-representation analysis and gene/protein set scores according to the existing overlap between pathways. Such methods neglect knowledges coming from the interactions between the gene/protein sets. Here, we introduce a novel GNN-based enrichment analysis method called GNNenrich. GNNenrich, through multiple levels of embedding that integrate protein sequence properties and interactions network, establishes functional relationship to support biological interpretation. RESULTS: GNNenrich have been tested and compared to over-representation analysis technique (g:Profiler) and graph-based method (EnrichNet). It demonstrates the capacity to reproduce results provided by others approaches and offers new perspectives for interpretation, returning relevant results supported by protein-protein interactions (PPIs). AVAILABILITY AND IMPLEMENTATION: Source code is available at https://gitlab.com/cnrgh/gnn-enrich/gnn-enrich-article-demo.

Machine Learning has been subject to numerous healthcare applications, aiming to enhance early diagnosis and treatment strategies. Due to privacy regulations, restricting personal information sharing across institutions, the challenge of gathering diverse large-scale datasets using patient information is significant. Federated Learning (FL) provides an innovative solution by sharing models between remote learners and a central server instead of exchanging sensitive data.In this paper, we compare three Deep Learning models trained in FL for breast cancer subtyping, using gene expression. We innovate by studying models using either raw, ordered, or 2D-converted expression arrays. Three recent architectures are implemented depending on the input type, including a Transformer, a 1D CNN (NeXt-TDNN) and a 2D CNN (ConvNeXt) respectively. Each model’s performance is evaluated in regards to client count, data distribution, and aggregation strategy in order to measure the impact of both client configurations and server parameters.The results suggest that CNNs share comparable performance to Transformers despite carrying a lower parameter count. Transformers offer more consistent accuracy, while CNNs are more stable to train. In FL, aggregator choice and server learning rate both impact performance. Compared with a centralized setup, FL tends to yield a higher accuracy. This study highlights the impact of training setup on performance, convergence and stability in federated learning, showcasing the challenges and benefits in applying collaborative methods to clinical research.

International audience

The lack of interpretability, explainability and transparency makes deep learning models untrusted to perform reliably for making critical decisions. Despite their evaluation against disturbances including geometric transformations, occlusion and convolutional noises in the case of DNN-based image classifiers, the evaluation of contour classifiers has only been studied against rigid displacements (rotation and translation). In this paper, we introduce ContourCertif: a new system to certify deep contour classifiers against convolutional attacks. We use the abstract interpretation theory in order to formulate the Lower and Upper Bounds with abstract intervals to support other classes of advanced attacks including filtering.