Advances in Explainable Models for Single-cell Transcriptomics and Spatial Transcriptomics Data Analysis

Tissue heterogeneity is supported by single-cell sequencing technology, and spatial transcriptomics provides an important technological platform to study the problem, continuous changes in cell state, and microenvironment interactions. Yet the high dimensionality and sparsity of data, batch effects, and spatial point mixing may confuse model inferences. In this review, we discuss recent progress of interpretable model and method developments in single-cell and spatial omics analysis with biological applications, especially for the deep generative model single-cell Variational Inference (scVI) as well as its multimodal extensions total Variational Inference (totalVI) and Multimodal Variational Inference (MultiVI); the spatial mapping method Tangram; and the intercellular communication inference frameworks NicheNet and CellChat. We additionally discuss how pre-trained base models like single-cell Bidirectional Encoder Representations from Transformers (scBERT) and single-cell Generative Pre-trained Transformer (scGPT) can be used for cross-dataset transfer, and the problems this poses to interpretability. This work highlights the construction of an evidential chain centered around “gene/pathway - spatial domain - communication network”, and indicates that spatial colocalization, multimodal consistency, and uncertainty assessment can enhance the confidence and repeatability of mechanistic inference, hence providing a testable hypothesis-generating platform for tumor immunology studies.