scNODE : generative model for temporal single cell transcriptomic data prediction.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Single-cell gene expression analysis is crucial for understanding cell development.
• Current methods face limitations due to sparse, discrete timepoint sampling in experiments.
• Missing temporal data hinders comprehensive analysis of cell developmental trajectories.

Purpose of the Study:

• To develop a deep learning model, scNODE, for predicting single-cell gene expression at unobserved timepoints.
• To address the challenge of missing temporal information in single-cell RNA sequencing (scRNA-seq) data.
• To improve downstream analyses such as cell trajectory inference and gene perturbation studies.

Main Methods:

• scNODE integrates a variational autoencoder with neural ordinary differential equations.
• The model utilizes a continuous and nonlinear latent space for gene expression prediction.
• A dynamic regularization term is incorporated for robustness against distribution shifts.

Main Results:

• scNODE demonstrated superior predictive performance compared to state-of-the-art methods on three real-world scRNA-seq datasets.
• Predictions from scNODE improved cell trajectory inference in scenarios with missing timepoints.
• The learned latent space proved valuable for in silico perturbation analysis.

Conclusions:

• scNODE effectively predicts gene expression at unobserved timepoints, filling critical data gaps.
• The model enhances the analysis of cell development by enabling continuous temporal modeling.
• scNODE offers a powerful tool for advancing research in developmental biology and computational genomics.