Contributed Session 4: Omic Data & Health Informatics

Spatially resolved transcriptomics (SRT) is a recent technology that al- lows studying the spatial structure of gene expression. Several methods have been proposed to identify, from SRT data, genes with mRNA abundance that varies across the tissue, named spatially variable genes. However, at present, no method allows comparing spatial patterns between experimental conditions. Here we present an approach, based on a negative binomial model of gene-expression data and a likelihood ratio test, to identify genes that exhibit changes in the spatial expression structure. We designed simulation studies, where our framework exhibits strong performance, in terms of sensitivity, specificity and runtime. The method will be freely distributed within DESpace Bioconductor R package.

The development of Group Equivariant Non-Expansive Operators (GENEOs) has led to innovative applications in Artificial Intelligence (AI). A growing body of research suggests that these mathematical tools can be seamlessly integrated into eXplainable Artificial Intelligence (XAI), leveraging their inherent interpretability. This study seeks to validate this assertion by testing the robustness of GENEOnet, a GENEO-based network specifically designed for computational biochemistry. Through a statistical analysis, we show that GENEOnet exhibits remarkable resilience against perturbations arising from Molecular Dynamics (MD) simulations. Our findings thus substantiate the value of employing GENEOs in the pursuit of Trustworthy AI, where reliability and robustness are paramount.

The analysis of gene expression data has become increasingly prevalent in medical research, as it can provide valuable insights for disease discrimination and early diagnosis. A key challenge in this area is identifying the most informative genes among the large number of potential candidates often available. In this work, we apply a novel approach for gene selection that addresses redundancy within gene expression data with the ultimate goal to perform tumours classification. Specifically, starting from the mutual information measure, the k-medoids algorithm is applied to cluster genes with similar patterns of pairwise dependence. The principal advantage of this algorithm is the possibility to obtain observed data point, the so-called medoids, as centre of cluster. Therefore, the resulting medoids can be used as the representative features in subsequent statistical methods for classification.