Building a pre-trained model for single-cell multi-omics analysis beyond highly variable feature selection

Abstract Single-cell multi-omics technologies enable simultaneous measurement of transcriptomic, epigenomic, and proteomic modalities in individual cells, offering a unique opportunity to dissect cellular identity and regulation. The heterogeneity of multi-omics data, together with the wide range of downstream analytical tasks, underscores the need for a versatile computational framework for integrative analysis. However, current multi-omics analysis approaches typically rely on task-specific training and highly variable feature selection, which can hinder effective knowledge transfer across datasets and discard biologically informative signals. Here, we present scPMMI, a pre-trained multimodal model for integrative analysis of single-cell multi-omics data. By grouping features into patches according to genomic coordinates, scPMMI processes tens of thousands of features per cell with reduced reliance on prior feature selection, thereby preserving more complete biological information. Our results demonstrate that scPMMI learns biologically coherent cell representations by jointly modeling multi-omics data. Through further transfer learning, the model can be readily adapted to achieve superior performance across diverse downstream tasks, including multi-omics integration, cell type annotation, protein abundance prediction, transcriptomic enhancement, and trajectory reconstruction.