Work

Benchmarking temporal causal discovery for fully and partially observed biochemical kinetic models

Year: 2026
Type: preprint
Abstract: Accurately reconstructing networks of biochemical reactions is a crucial yet unsolved problem for leveraging advances in experimental techniques that track intracellular dynamics in a time-resolved fashion. This task is further challenged by the fact that we typically observe only part of the system, while many variables shaping the dynamics remain unobserved. Causal discovery methods can infer causal relations from purely observational data and account for unobserved variables. They therefore hold promise for yielding faithful, mechanistically interpretable representations of underlying biochemical interactions, but their potential in this context remains underexplored. In this work, we provide a systematic benchmark of state-of-the-art temporal causal discovery methods (tMIIC, PCMCI+/LPCMCI and SVAR-FCI) on time-series data from simulations of biochemical kinetics of mass action and Michaelis-Menten type modelled as systems of nonlinear stochastic differential equations, including an experimentally calibrated model of the epidermal growth factor receptor signalling network. By selectively omitting data, we consider the problem of reconstructing these networks in the presence of both latent confounders and unobserved species participating in reactions. We find that tMIIC and PCMCI+/LPCMCI achieve the highest overall prediction quality, with tMIIC able to handle larger samples within the same runtime budget. We identify, as the main condition for optimal performance of causal discovery, the consistency between data-sampling regimes (sampling intervals, trajectory length) and the typical timescales of the system. We establish that causal discovery can detect latent confounders via bidirected edges and unobserved species via time-delayed edges, locating hidden effects and encoding information on their typical timescales, provided that the data sampling guarantees adequate temporal resolution. As such, our results can guide the choice of hyperparameters in causal discovery algorithms and of data-sampling regimes to optimize network-reconstruction performance. Author summary Systems of intracellular biochemical reactions are highly complex. Inferring, from observational data, network representations of these systems, with nodes representing biochemical species and edges representing their reactions, is of fundamental importance for quantitatively characterizing their function and the effects of dysregulation, yet it remains challenging. A further difficulty is that biochemical systems often involve components that cannot be directly measured, although we would like to know if and how these ‘hidden nodes’ influence what we observe. Causal discovery methods, i.e. methods that can learn networks of causal relationships, have the potential to bring substantial improvements, one of their strengths being their mechanisms for detecting hidden-node effects from data. In this work, we deploy and benchmark causal discovery to learn causal networks from time-series data generated by models of fully or partially observed stochastic biochemical kinetics, comparing state-of-the-art methods and identifying the sampling-regime characteristics that lead to optimal performance relative to the typical timescales of the kinetics. Our results demonstrate the ability of these methods to reconstruct the original biochemical networks from which the data were generated and to detect hidden nodes, showcasing the feasibility and usefulness of causal discovery as part of data-driven mathematical-modelling pipelines for biochemical reaction systems.
Language: English
Open Access status: green
OpenAlex ID: W7234373689
Cites: 65
Cited by:
Related to: 10