Novel Parameter-Free and Interpretable Integration of CITE-seq RNA and ADT Profiles via Tensor Decomposition-Based Unsupervised Feature Extraction

Abstract

CITE-seq jointly profiles cellular transcripts and surface proteins, but RNA and ADT modalities differ markedly in dimensionality, sparsity, and noise characteristics. We applied tensor-decomposition-based unsupervised feature extraction to paired CITE-seq data by constructing a gene x cell x protein tensor and performing HOSVD. The proposed workflow does not require explicit RNA/ADT modality-weight tuning or prior HVG-based gene filtering, and it provides cell-mode singular vectors together with post hoc unsupervised gene selection. Across ImmGen T-cell CITE-seq datasets, TD-derived cell representations preserved cell-type-related local structure and showed competitive kNN-based consistency compared with scMoMaT, a related factorization-based reference. However, ADT-only embeddings and Seurat WNN graphs often showed higher cell-type neighborhood consistency, indicating that TD-based UFE should not be interpreted as a replacement for marker-based or graph-based cell-type analysis. Enrichment analysis of TD-selected genes supported their biological plausibility but was interpreted as an exploratory check rather than proof of complete marker recovery. These results position TD-based UFE as a lightweight tensor-based unsupervised feature extraction framework for paired RNA/ADT data, rather than as a universally superior CITE-seq integration method.