Explainable Artificial Intelligence in Single-Cell Transcriptomics for Cancer Research: From Cellular Heterogeneity to Biomarker and Therapeutic Target Discovery

Single-cell transcriptomics has transformed cancer research by resolving malignant, immune, stromal, and vascular populations that are obscured in bulk tissue measurements. At the same time, the dimensionality, sparsity, technical variability, and hierarchical organization of single-cell RNA sequencing data have made artificial intelligence (AI) increasingly important for representation learning, cell annotation, state discovery, trajectory inference, perturbation modeling, and biomarker prioritization. Yet high predictive performance is not equivalent to biological understanding. In oncology, an AI-derived feature may reflect tumor-intrinsic signaling, stromal abundance, immune infiltration, technical batch structure, or a cohort-specific shortcut. Explainable artificial intelligence (XAI) is therefore most useful when feature attribution is linked to cellular localization, cross-cohort stability, and biological evidence rather than treated as a post hoc visualization. This critical review synthesizes major computational approaches for AI-enabled single-cell analysis in cancer, including deep generative models, transfer learning, foundation models, cell-cell communication frameworks, trajectory methods, and multimodal integration with spatial and molecular data. We examine representative cancer applications in malignant-state mapping, tumor microenvironment characterization, biomarker discovery, and therapeutic target prioritization. We further propose a cell-aware explainability framework in which predictive validity, attribution stability, cell-type localization, mechanistic plausibility, and translational relevance are evaluated as distinct evidence layers. Emerging single-cell foundation models such as scGPT, scFoundation, Geneformer, and scBERT expand the possibility of reusable biological representations, but they also intensify concerns about data provenance, benchmark leakage, model scale, interpretability, and population transferability. Future progress will depend less on increasingly complex architectures and more on rigorous external validation, reproducible preprocessing, patient-level inference, spatial or perturbational confirmation, and alignment with a clearly defined clinical or therapeutic use case.

کلیدواژه ها:

single-cell RNA sequencing ، explainable artificial intelligence ، cancer ، transcriptomics