Deep learning models for cell cycle phase prediction from single-cell RNA sequencing data

Accurate prediction of cell cycle phases is essential for mitigating confounding effects in single-cell RNA sequencing (scRNA-seq) analysis and for understanding how certain diseases (e.g. cancer) develop and respond to treatment. We evaluated traditional machine learning (AdaBoost, Random Forest, LightGBM) and deep learning approaches (DNN3, DNN5, CNN, hybrid models, ensembles) for cell cycle phase prediction from scRNA-seq data. Models were trained using consensus phase labels derived from four established tools (CellCycleScore, ccAFv2, Revelio, and Tricycle) applied to single-cell gene expression datasets representing diverse biological contexts, including a human leukemia cell line (REH), peripheral blood mononuclear cells (10× PBMC), human pluripotent stem cells and their differentiated derivatives (GSE75748), 10 k embryonic mouse brain cells, and mouse hematopoietic stem and progenitor cells (GSE81682). Performance was evaluated on three independent datasets: human embryonic stem cells (GSE64016; 247 cells), human osteosarcoma cells (GSE146773; 1152 cells), and mouse embryonic stem cells (E-MTAB-2805; 288 cells), all with experimentally verified labels. REH-trained models achieved strong cross-dataset performance, reaching 74.35% accuracy on GSE146773 (Top 3 decision fusion (DF), and score fusion (SF)) and 72.9% on GSE64016, while attaining 56.9% on the cross-species Buettner mESC benchmark. SHapley Additive exPlanations (SHAP) analysis confirmed enrichment in cell-cycle-related gene sets for features. Notably, models trained on samples with lower recovery rates of cell cycle markers exhibited lower performance. These results highlight the strength of deep learning for robust, scalable cell cycle phase prediction from scRNA-seq data.