Simultaneously extracts and integrates cell morphology and cell surface protein expression for multimodal single-cell analysis.

Single-cell multimodal analysis based on machine learning integrates data from multiple modalities such as cell morphological features, gene expression, proteomics, and metabolomics and observes single cells from multiple dimensions. This multi-dimensional features method can accurately capture the differences between cells and enable us to understand cell heterogeneity deeply. However, the acquisition and processing of single-cell multimodal data are complex, with problems such as data noise, missing values, and batch effects, which restricts the development of machine learning models that highly depend on data quality and preprocessing for single-cell heterogeneity research. Herein, we propose an integrated single-cell multimodal coupling analysis strategy that simultaneously extracts and integrates the morphological features of cells and the surface protein features of cells. The features of cell surface proteins and cell morphology were integrated and obtained through the fluorescence imaging technology using DNA aptamer labeling. As a proof-of-concept, we select six morphological features of seven cells and four cell surface protein features as the bimodal information coupling input of the model for single-cell classification and prediction. The results show that the accuracy of multimodal coupling analysis is much higher than that of cell morphological features, up to 94.46 %. Moreover, the stability of multimodal coupling analysis is higher than that of cell surface protein features. It reaches stability after only 10 training processes, and the fluctuation range is much smaller than that of cell morphological features and cell morphological features training, which indicates that our multimodal coupling model expects to overcome the risk of overfitting.