CIRI-Deep is a deep learning model designed to predict differentially spliced circular RNAs (circRNAs) across various types of RNA-seq data, including single-cell and spatial transcriptomic datasets. The model addresses the limitations of traditional circRNA analysis methods, which often struggle with low sequencing depths and poly(A) selection biases. CIRI-Deep was trained on a comprehensive dataset of 25 million high-confidence circRNA regulation events from 397 human tissue samples. It demonstrates high accuracy in predicting differential splicing events, even in low-depth or poly(A) selected data. The model incorporates both cis and trans features, including RBP binding motifs and expression levels, to accurately predict circRNA regulation. Adapted versions of CIRI-Deep, such as CIRI-deepA, are also developed to handle poly(A) selected data. These models enable the analysis of circRNA heterogeneity at the single-cell and spatial levels, providing valuable insights into tissue-specific and cell-type-specific circRNA expression patterns. The study highlights the potential of deep learning in advancing circRNA research by broadening the applicability of circRNA analysis to a wider range of datasets.CIRI-Deep is a deep learning model designed to predict differentially spliced circular RNAs (circRNAs) across various types of RNA-seq data, including single-cell and spatial transcriptomic datasets. The model addresses the limitations of traditional circRNA analysis methods, which often struggle with low sequencing depths and poly(A) selection biases. CIRI-Deep was trained on a comprehensive dataset of 25 million high-confidence circRNA regulation events from 397 human tissue samples. It demonstrates high accuracy in predicting differential splicing events, even in low-depth or poly(A) selected data. The model incorporates both cis and trans features, including RBP binding motifs and expression levels, to accurately predict circRNA regulation. Adapted versions of CIRI-Deep, such as CIRI-deepA, are also developed to handle poly(A) selected data. These models enable the analysis of circRNA heterogeneity at the single-cell and spatial levels, providing valuable insights into tissue-specific and cell-type-specific circRNA expression patterns. The study highlights the potential of deep learning in advancing circRNA research by broadening the applicability of circRNA analysis to a wider range of datasets.