Development of Artificial Intelligence Techniques in Spatial Transcriptomics

Vivienne Zhang

Spatial transcriptomics, an emerging technology that integrates imaging, spatial coordinates, and transcriptomic expression matrices, finds natural synergy with artificial intelligence (AI) due to AI’s inherent capabilities in image analysis and matrix data processing. However, as AI algorithms advance rapidly with increasing diversity, their application to the unique data characteristics and biological questions of spatial transcriptomics necessitates systematic evaluation. This thesis presents a comparative analysis of machine learning architectures to identify optimal solutions, subsequently developing enhanced algorithms through the strategic integration of existing methodologies. Specifically, this thesis focuses on two critical challenges in spatial transcriptomics: cell segmentation and spatial domain identification. Spatial transcriptomics analysis begins with the integration of spatial coordinates and gene expression matrices, requiring advanced computational frameworks for cell segmentation. This study systematically evaluates segmentation methodologies spanning image-centric approaches to expression-enhanced algorithms, benchmarking seven methods across the Adult Mouse Hemi-Brain Stereo-seq and Seq-Scope Mouse Liver datasets. Analytical comparisons between overlap-correlation metrics and conventional cross-validation approaches revealed metric-dependent performance variations, with traditional validation demonstrating superior robustness and consistently identifying U-Net-based architectures (Cellpose and DeepCell) as optimal performers, while exposing significant limitations in correlation-based evaluation through its weak association with precision/recall metrics (|r|<0.75) versus strong cellsize dependency (r>0.75). Identifying spatial domains is crucial in spatial transcriptomics, yet effectively integrating gene expression, spatial location, and histology remains challenging. This thesis presents STESH, a Spatial Transcriptomics clustering method that combines Expression, Spatial information and Histology. STESH extracts histological features using an improved convolutional neural network and generates expression, histology, spatial, and collaborative convolution modules for a multi-view graph convolutional network with a decoder and attention mechanism. STESH was evaluated across multiple tissue types and technology platforms. STESH consistently outperformed ten state-of-the-art methods and achieved the highest scores in adjusted Rand index, normalised mutual information, and Fowlkes-Mallows index.

Development of Artificial Intelligence Techniques in Spatial Transcriptomics

Files and links (1)

Abstract

Details

Metrics