A knowledge graph-guided framework for multitask single-cell analysis

Mathew, B., Maity, A., Ganguly, R., Halder, A., Sabu, J. T., Gupta, N., Peela, S. C. M., Samanta, S., Kumari, S., Bhattacharjee, N., et al.

10.64898/2026.04.25.720773 · was preprinted

benchmarked

Surfaced because: benchmarked against baselines.

relevance 0.41 openness 0.25 novelty 0.40

Abstract

Connecting cell identity to biological function is central to understanding how tumors change during treatment. Here, we introduce scKNIFE, a knowledge-guided nonnegative matrix factorization framework that integrates gene expression with a heterogeneous graph of functional annotations, metabolic reactions and metabolites, and signaling interactions. scKNIFE jointly learns cellular rep resentations and biological-entity loadings, linking the factors that organize cells to their constituent genes and connected functions. Across breast-cancer cohorts, the learned representations retained broad cellular organization without using per-cell identity labels as factorization targets. scKNIFE identified baseline tumor programs associated with subsequent clinical response or T-cell clonotype expansion and resolved treatment-associated functional variation within immune and stromal popula tions. In a seven-patient triple-negative breast cancer atlas, analysis alongside tissue metabolomics identified distinct trajectories of inferred fatty acid activation and carnitine-associated activities, in cluding higher myeloid L-octanoylcarnitine scores at biopsy in all seven patients. Communication analysis further resolved an endothelial-perivascular PDGF connection with lower inferred scores at surgery in six patients. scKNIFE brings cellular identity, metabolic programs and candidate intercellular communication into one representation, enabling functional interpretation of tumor heterogeneity across patients and treatment states.

Lifecycle

biorxiv v2 2026-10-24 source ↗

Discussion

No qualifying discussion yet.