IST: an ontology-guided attention-based autoencoder for interpretable analysis of single-cell transcriptomic data.

Extracting biologically interpretable insights from single-cell RNA sequencing (scRNA-seq) data remains a major challenge. While deep learning approaches have demonstrated strong predictive performance for dimensionality reduction and representation learning, their latent representations often lack clear biological meaning, and prior biological knowledge is frequently incorporated only weakly or inconsistently. As a result, learned features may not faithfully reflect known gene-function relationships, limiting biological interpretability and downstream insight. We present IST (Interpretation of Single-cell Transcriptomic data), a novel attention-based framework designed to reliably integrate biological prior knowledge into the analysis of single-cell transcriptomic data. IST incorporates gene ontology structure directly into the model architecture and employs attention mechanisms to capture context-dependent gene activity. A tailored correlation-based loss enforces alignment between learned gene activities and known gene-function relationships, ensuring that biological priors are faithfully reflected in the learned representations rather than acting as superficial regularization. We evaluate IST on three scRNA-seq datasets. Across all datasets, IST identifies condition-specific gene activity patterns, uncovers novel relationships among biological processes, and infers previously unannotated gene functions supported by existing literature. Compared with state-of-the-art methods, IST provides more interpretable and biologically grounded representations of gene functions and activities, while maintaining competitive predictive performance. The datasets and tool are available at https://doi.org/10.6084/m9.figshare.30811334.