Deploying multi-modal learning in the analysis of single-cell omics to facilitate perturbation modeling

Understanding the complexity of cellular and tissue heterogeneity and its impact on tissue and organ function is a fundamental challenge in life science research. Spatial transcriptomics, which combines imaging-based technologies with high-throughput single-cell sequencing, has revolutionised the understanding of cell and tissue organisation. This study aims to explore the feasibility of building a computational pipeline by investigating the application of machine learning (ML) and deep learning (DL) methods in spatial multi-omics, focusing on clustering cell states after perturbation modelling. In Part A, we optimised spatial clustering by benchmarking key parameters and identifying optimal settings for combining scaled spatial connectivity with gene expression data. This approach significantly enhanced clustering accuracy, achieving results that closely matched the ground truth. Building on these findings, Part B involved hyperparameter tuning of the GenKI perturbation tool on a processed DBiT-seq mouse embryo dataset. Our results demonstrate that specific parameter settings effectively capture comprehensive spatial patterns and gene expression profiles, underscoring the value of perturbation modelling in spatial multi-omics for identifying key regulatory genes and assessing their functional impact. The use of well-characterised mouse models and genetic profiles provided a robust framework for investigating spatial molecular patterns and cellular interactions, with implications for disease research and drug discovery. In summary, this study demonstrates the feasibility of constructing a computational pipeline for perturbation modelling in spatial multi-omics by integrating ML/DL algorithms with biological insights. This integrated approach facilitates the exploration of gene functions and interaction networks in a spatially resolved manner and holds significant potential for advancing precision medicine and AI-driven drug discovery.