Integrating intra- and inter-cell gene-gene interactions into deep omics data analysis for enhanced single-cell cancer biology

Single-cell RNA sequencing (scRNA-seq) is widely used in cancer research to dissect tumor heterogeneity, characterize malignant and immune cell states, and identify dysregulated regulatory programs. However, accurately accounting the gene-gene interactions in cellular samples remains challenging due to the high dimensionality and nonstructured sequences of single-cell data. Existing Transformer-based approaches often rely on oversimplified gene-embedding strategies-such as ordering, coarse value binning, or direct value projection-that reduce biological resolution and overlook key regulatory dependencies. Moreover, most current methods emphasize intra-cell interactions while neglecting population-level patterns that are essential for understanding tumor microenvironmental regulation. To address these limitations, we propose a novel dual-branch Transformer framework that explicitly integrates intra-cell and inter-cell gene-gene interactions. The method comprises two complementary branches: (1) an intra-cell interaction branch that uses a Transformer augmented with interaction-aware embeddings to capture fine-grained regulatory relationships within individual cells based on graph-derived gene representations; and (2) an inter-cell interaction branch that applies a Vision Transformer (ViT) to image-based representations of single-cell profiles, spatially organized to reflect global gene-gene interaction structures across cell populations. A cross-attention module links the two branches, enabling coordinated learning between intracellular and intercellular regulatory signals. Extensive evaluation across diverse downstream tasks-such as cell-type classification, including cancer cell recognition and characterization; gene regulatory network inference; and protein abundance prediction-demonstrates that this interaction-aware architecture consistently outperforms state-of-the-art approaches. Notably, the framework achieves approximately a 30% average improvement in protein abundance prediction, a 4% improvement in cell-type classification accuracy, and a 4% improvement in gene regulatory network inference performance. By jointly capturing cell-intrinsic regulatory signals and population-level interaction patterns, the proposed framework offers a powerful computational strategy for dissecting tumor heterogeneity and characterizing regulatory interactions between malignant and microenvironmental cell populations.