Spatial transcriptomics quantifies spatial molecular profiles alongside histology, enabling computational prediction of spatial gene expression distribution directly from whole slide images. Inspired by image-to-text alignment and generation, we introduce Magic, a self-training contrastive learning model designed for histology-to-gene expression prediction. Magic (i) employs contrastive learning to derive shared embeddings for histology and gene expression while utilizing a momentum-based module to generate pseudo-targets to reduce the impact of noise; and (ii) leverages a transformer-based decoder to predict the expression of 300 genes based on histological features. Trained on 75 760 spots from 56 breast cancer slices and validated on 11 026 spots from five independent slices, Magic outperforms existing methods in aligning and generating histology-gene expression data, achieving a 10% improvement over the second-best approach. Furthermore, Magic demonstrates robust generalization, effectively predicting gene expression in colorectal cancer samples and The Cancer Genome Atlas (TCGA) datasets through zero-shot learning. Notably, Magic's predicted gene expression captures interpatient differences, highlighting its strong potential for clinical applications.

Simultaneous profiling transcriptomic and chromatin accessibility information in the same individual cells offers an unprecedented resolution to understand cell states. However, computationally effective methods for the integration of these inherent sparse and heterogeneous data are lacking. Here, we present a single-cell multimodal variational autoencoder model, which combines three types of joint-learning strategies with a probabilistic Gaussian Mixture Model to learn the joint latent features that accurately represent these multilayer profiles. Studies on both simulated datasets and real datasets demonstrate that it has more preferable capability (i) dissecting cellular heterogeneity in the joint-learning space, (ii) denoising and imputing data and (iii) constructing the association between multilayer omics data, which can be used for understanding transcriptional regulatory mechanisms.

MOTIVATION: Joint profiling of single-cell transcriptomics and epigenomics data enables us to characterize cell states and transcriptomics regulatory programs related to cellular heterogeneity. However, the highly different features on sparsity, heterogeneity and dimensionality between multi-omics data have severely hindered its integrative analysis. RESULTS: We proposed deep cross-omics cycle attention (DCCA) model, a computational tool for joint analysis of single-cell multi-omics data, by combining variational autoencoders (VAEs) and attention-transfer. Specifically, we show that DCCA can leverage one omics data to fine-tune the network trained for another omics data, given a dataset of parallel multi-omics data within the same cell. Studies on both simulated and real datasets from various platforms, DCCA demonstrates its superior capability: (i) dissecting cellular heterogeneity; (ii) denoising and aggregating data and (iii) constructing the link between multi-omics data, which is used to infer new transcriptional regulatory relations. In our applications, DCCA was demonstrated to have a superior power to generate missing stages or omics in a biologically meaningful manner, which provides a new way to analyze and also understand complicated biological processes. AVAILABILITY AND IMPLEMENTATION: DCCA source code is available at https://github.com/cmzuo11/DCCA, and has been deposited in archived format at https://doi.org/10.5281/zenodo.4762065. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.