Transcriptomic forecasting with neural ordinary differential equations.

Area of Science:

• Computational Biology
• Genomics
• Systems Biology

Background:

• Single-cell transcriptomics reveals molecular changes underlying cellular phenotypes.
• Understanding dynamic cellular processes requires estimating temporal gene expression changes, not just inferring trajectories from static snapshots.

Purpose of the Study:

• To develop a novel computational method, RNAForecaster, for predicting future gene expression states in single cells.
• To enable short-term estimation of future cellular states from high-throughput transcriptomic data with temporal information.

Main Methods:

• Developed RNAForecaster, a neural ordinary differential equation-based model.
• The method operates in an embedding-independent manner, predicting gene expression for multiple future time steps.
• Validated using simulated single-cell transcriptomic data with cellular tracking and real metabolic labeling scRNA-seq data.

Main Results:

• RNAForecaster accurately predicted future expression states in simulated single-cell transcriptomic data.
• The method successfully recapitulated expected gene expression changes during cell cycle progression over a 3-day period using metabolic labeling scRNA-seq data.
• Demonstrated the capability of RNAForecaster for predicting temporal gene expression dynamics.

Conclusions:

• RNAForecaster is a powerful tool for predicting short-term future gene expression states in single cells.
• The method enhances the analysis of dynamic cellular processes using temporal single-cell transcriptomic data.
• RNAForecaster facilitates a deeper understanding of cellular dynamics from high-throughput datasets.