Learning stable and predictive structures in kinetic systems.

Area of Science:

• * Computational science
• * Systems biology
• * Chemical kinetics

Background:

• * Learning kinetic systems from data is crucial for many scientific fields.
• * Stable models are essential for reliable data-driven inference and generalization.
• * Reproducible research relies on identifying invariant underlying models.

Purpose of the Study:

• * Introduce CausalKinetiX, an efficient framework for identifying kinetic system structures.
• * Leverage a causal perspective to enhance the learning of kinetic models.
• * Enable concise descriptions of dynamics and predictions in unseen experiments.

Main Methods:

• * Developed a computationally efficient framework, CausalKinetiX.
• * Identified structure from discrete-time, noisy observations from heterogeneous experiments.
• * Assumed the existence of an underlying, invariant kinetic model.

Main Results:

• * Demonstrated improved structure identification in kinetic systems using a causal approach.
• * Achieved concise descriptions of dynamics across multiple experimental settings.
• * Showcased enhanced prediction capabilities in unseen experiments.
• * Observed significant improvements in out-of-sample generalization compared to established methods.

Conclusions:

• * A causal perspective significantly benefits the learning of kinetic system structures.
• * CausalKinetiX provides a robust and efficient method for kinetic modeling.
• * The framework enhances model generalization and predictive performance.