NEURALGENE: INFERRING GENE REGULATION AND CELL-FATE DYNAMICS FROM NEURAL ODES

Abstract

In biology, cell-fate decisions are controlled by complex gene regulation. Although gene expression data may be collected at multiple time points, it remains difficult to construct the continuous dynamics from the data. In this work, we developed a data-driven approach, NeuralGene, a model based on neural ordinary differential equations (ODEs), to reconstruct continuous dynamical systems governing gene regulation from temporal gene expression data. In addition, NeuralGene has the flexibility of incorporating partial prior biological information in the model to further improve its accuracy. For a given cell at a static time point, the NeuralGene model can impute its continuous gene expression dynamics and predict its cell fate. We applied NeuralGene to a simulation toggle-switch model to verify its utility in modeling and reconstructing temporal dynamics. In addition, NeuralGene was applied to experimental single-cell qPCR data to show its ability for gene expression imputation and cell-fate prediction.