Single-cell RNA sequencing has given researchers an unprecedented view of how individual cells behave. One important piece of information scientists often want to know is where each cell is in the cell cycle, the series of stages that cells pass through as they grow and divide. Accurately identifying these stages is important because the cell cycle strongly influences gene expression. If researchers do not account for these differences, cell cycle activity can interfere with the interpretation of RNA sequencing data and make it more difficult to identify meaningful biological changes.

Researchers from the Lane Department of Computer Science and Electrical Engineering at West Virginia University evaluated whether deep learning could improve cell cycle phase prediction from single-cell RNA sequencing data.

Several computational tools already estimate cell cycle phase, but they often produce different results. To address this challenge, the researchers combined the predictions from four established methods to generate consensus cell cycle labels. These consensus labels were then used to train a variety of artificial intelligence models.

The team compared several traditional machine learning approaches, including AdaBoost, Random Forest, and LightGBM, with multiple deep learning models. These included deep neural networks, convolutional neural networks, hybrid architectures, and ensemble methods designed to combine predictions from multiple models.

To test how well these approaches generalized, the researchers trained the models using diverse single-cell RNA sequencing datasets representing leukemia cells, immune cells, human stem cells, mouse brain cells, and hematopoietic stem cells. They then evaluated performance using three completely independent datasets with experimentally verified cell cycle labels.

Phase-specific precision and recall heatmaps for REH-trained models evaluated on the GSE146773, GSE64016, and Buettner mESC benchmarks

The deep learning models consistently performed well across these independent datasets. Some ensemble approaches correctly predicted cell cycle phases in more than 74% of cells, demonstrating that artificial intelligence can accurately identify cell cycle status even when analyzing data generated from different biological systems.

The researchers also investigated which genes contributed most strongly to the models’ predictions using a technique called SHAP analysis. This approach helps explain how artificial intelligence models make decisions by identifying the genes that have the greatest influence on each prediction. The analysis confirmed that the models relied heavily on genes already known to regulate the cell cycle, increasing confidence that the predictions reflected genuine biology rather than statistical artifacts.

Another important observation was that prediction accuracy depended on data quality. Datasets containing fewer detectable cell cycle marker genes produced lower model performance, highlighting the importance of generating high quality RNA sequencing data for reliable downstream analysis.

As single-cell RNA sequencing datasets continue to grow in both size and complexity, computational methods capable of automatically identifying cell cycle phases will become increasingly valuable. Improved prediction models can help researchers reduce confounding effects during data analysis while providing deeper insight into cell growth, development, cancer biology, and responses to treatment.

Availability – Complete source code for preprocessing, consensus label generation, model training, evaluation, and SHAP analysis is available at: https://github.com/Mituvinci/cell-cycle-prediction_using_ML