Abstract
Discovering interpretable mathematical models from experimental data is a cornerstone of scientific inquiry, yet purely data-driven approaches often struggle with high noise and physical inconsistency. We present a framework that integrates physical "priors"—such as conservation laws and symmetries—into the model discovery process using group-sparse regression. By employing a novel group Iterative Hard Thresholding (gIHT) algorithm combined with stability selection, we demonstrate the ability to recover accurate Ordinary and Partial Differential Equations (ODEs/PDEs) from noisy biological datasets. Our results show that enforcing structural constraints significantly improves robustness to noise and data efficiency compared to standard sparse regression methods.
Introduction
In systems biology and related fields, the abundance of high-throughput data has not always translated into a deeper understanding of underlying mechanisms. While "black box" machine learning models excel at prediction, they lack the interpretability required for scientific insight. Sparse regression techniques, such as SINDy, attempt to bridge this gap by identifying the simplest set of differential equation terms that fit the data. However, these methods are often sensitive to noise and can produce models that violate fundamental physical principles, such as mass conservation. This paper addresses these challenges by proposing a method that bakes physical consistency directly into the optimization process.
Methods
The central innovation of this work is the application of group sparsity to enforce physical constraints. Rather than penalizing individual terms in a candidate library, the framework groups terms that are physically linked. For instance, in a reaction network, terms representing the consumption and production of chemical species are grouped to ensure mass conservation.
To solve the resulting optimization problem, we developed the group Iterative Hard Thresholding (gIHT) algorithm. Unlike convex relaxations like Group LASSO, gIHT directly addresses the $L_0$ norm, providing superior performance when library terms are correlated. Furthermore, we utilize stability selection—training the model on multiple random sub-samples—to identify terms that consistently appear, thereby filtering out noise-driven artifacts without manual parameter tuning.
Results
The framework was validated on several benchmarks, including the JAK-STAT signaling pathway, protein transport, and reaction-diffusion systems. In the JAK-STAT case, the method successfully identified the correct model structure even at 10% noise levels, a regime where standard sparse regression fails.
For spatiotemporal systems, such as 1D protein transport, the inclusion of "model equivalence" priors allowed for 100% success in model recovery with minimal data points (200-400), even under 15% noise. In 2D reaction-diffusion systems, enforcing rotational and species symmetry via group sparsity ensured that the discovered PDEs remained physically plausible, whereas unconstrained methods yielded inconsistent results.
Discussion and Conclusion
By treating physical laws as structural constraints rather than just additional data, this framework provides a robust path toward automated scientific discovery. The emphasis on stability over mere data fit allows for the extraction of "true" underlying dynamics from messy biological observations. While current limitations include the assumption of non-overlapping groups, future work will focus on integrating these sparse regression techniques with Physics-Informed Neural Networks (PINNs) to improve derivative estimation in extremely high-noise environments. In conclusion, the integration of group sparsity and stability selection offers a powerful tool for learning physically consistent models that adhere to the fundamental laws of the natural world.
Summary updated