Data-driven model discovery and model selection for noisy biological systems.

Area of Science:

• Computational Biology
• Systems Biology
• Dynamical Systems Theory

Background:

• Ordinary differential equation models are crucial for representing biological dynamics but traditionally require extensive prior knowledge.
• Existing data-driven methods like sparse identification of nonlinear dynamics (SINDy) face challenges with biological noise and incorporating prior knowledge.

Purpose of the Study:

• To develop a robust data-driven framework for biological model discovery and selection that handles noisy and sparse data.
• To improve upon existing methods by incorporating prior knowledge and mitigating the effects of biological noise.

Main Methods:

• Utilizing hybrid dynamical systems, where neural networks approximate unknown system dynamics and denoise data.
• Employing sparse regression on neural network simulations to infer differential equation models.
• Implementing model selection to compare the proposed framework against alternative approaches.

Main Results:

• The proposed framework successfully infers biological models from data with high levels of various biological noise types.
• Hybrid dynamical systems approach demonstrates superior performance in model discovery compared to traditional methods.
• Accurate model inference is achieved even with sparse and noisy single-cell transcriptomics data.

Conclusions:

• This data-driven framework offers a practical solution for biological model discovery, especially when dealing with noisy, sparse data and incomplete mechanistic understanding.
• The approach enhances the ability to learn latent dynamics and incorporate prior knowledge, advancing systems biology research.