Aims: Alternative splicing serves as a primary mechanism for diversifying the proteome, making the prediction of distinct isoform functions critical for understanding complex disease mechanisms. However, determining the specific functional roles of isoforms remains hindered by high sequence homology among variants and the sparsity of isoform-level annotations.

Methods: In this study, we propose SpliceEM, a deep learning framework for isoform function prediction at single-cell resolution. SpliceEM utilizes a splicing event-aware encoder with cross-modal attention to separate functional signals from global protein sequences. A Heterogeneous Graph Transformer captures the dependencies among isoforms, genes, and Gene Ontology terms. To bridge the annotation gap, we incorporate a self-distillation framework guided by an Exponential Moving Average teacher model and Multi-Instance Learning, optimized by an Asymmetric Loss and hierarchical constraints.

Results: Benchmarking on human datasets demonstrates that SpliceEM outperforms existing methods in isoform function prediction, particularly in identifying rare functional terms under data-sparse conditions. Furthermore, splicing-function analysis reveals that specific splicing events, such as skipped exons and alternative first exons, act as prominent drivers in oncogenic signaling cascades and context-specific functional switching.

Conclusion: SpliceEM provides a computational foundation for exploring transcriptomic functional diversity. By shifting the focus from global sequences to localized splicing events and utilizing hierarchical biological priors, it offers high-resolution insights into cell-type-specific molecular mechanisms and potential therapeutic targets.