TL;DR
This paper introduces Reactmine, a novel algorithm for inferring chemical reaction networks (CRNs) from time-series data. Reactmine overcomes limitations of existing methods like SINDy, particularly when dealing with "wild-type" data lacking experimental perturbations, by employing a bounded-depth search tree to identify reactions sequentially and enforce sparsity.
What Problem Does This Paper Address?
Understanding complex biological systems often requires building mechanistic models, such as gene regulatory networks (GRNs) or chemical reaction networks (CRNs). Traditionally, this involves extensive manual literature review. While automated methods exist, inferring CRNs from time-series concentration data is particularly challenging. Existing methods, like SINDy, often rely on sparse regression and perform poorly when experimental data lacks diverse initial conditions (e.g., gene knockouts or zero concentrations), which are crucial for de-correlating reaction effects. This paper addresses the need for a CRN inference method that can effectively learn from limited, "wild-type" time-series data.
Methods
Reactmine is a statistical search algorithm that infers CRNs in a sequential, bounded-depth tree search. At each step, it identifies potential reaction candidates based on observed concentration changes. These candidates are then ranked by the variance of their inferred kinetics on their supporting time points. The top candidates are added as successor nodes in the search tree. After exploring the tree, a global re-optimization of kinetic parameters for the most promising CRN candidates is performed on the entire time series. This approach enforces sparsity by limiting the search depth and prioritizing reactions that best explain observed dynamics. Reactmine can handle mass action law kinetics and is extensible to Michaelis-Menten and Hill kinetics.
Datasets
The authors evaluated Reactmine on:
- Simulated data: From hidden CRNs, including a chain CRN, a loop CRN, a reactant-parallel CRN, a product-parallel CRN, and the MAPK signaling network.
- Real-world biological data:
- Fluorescence videomicroscopy data of cell cycle and circadian clock markers in mammalian fibroblasts.
- Biomedical measurements of systemic circadian biomarkers related to clock gene expression in peripheral organs.
Key Insights and Ideas
- Sparsity from sequential search: Reactmine's tree-search approach naturally enforces sparsity by limiting the depth of inferred reactions, avoiding the pitfalls of direct sparse regression on wild-type data.
- Variance as a ranking criterion: Ranking reaction candidates by the variance of their inferred kinetics on supporting transitions is an effective heuristic for identifying relevant reactions, especially when direct correlation is ambiguous.
- Wild-type data limitations: The paper highlights that methods relying on sparse regression (like SINDy) struggle with wild-type time-series data because the lack of experimental perturbations (e.g., zero initial conditions) leads to high correlations between reaction effects, making it difficult to disentangle them.
- Global re-optimization is crucial: While the search tree identifies candidate reactions, a final global re-optimization step refines kinetic parameters and selects the best overall CRN.
Key Results
- Superior performance on simulated data: Reactmine successfully recovered hidden CRNs from simulated data where SINDy failed, particularly in scenarios with limited initial conditions. For example, on the chain CRN, Reactmine inferred the correct structure and kinetics, while SINDy produced sparse ODE systems that did not fit the data.
- Successful application to real biological data: Reactmine was applied to two real datasets:
- Cell cycle and circadian clock: Identified key reactions like G2 → REV, aligning with previous biological findings.
- Systemic circadian regulators: Identified Food Intake and Temperature as major influencing factors on clock gene transcription, consistent with prior analyses.
- Robustness to hyperparameters: Sensitivity analyses showed that while Reactmine's performance depends on hyperparameters like $\alpha$ (reaction acceptance threshold) and $\beta$ (number of candidates), reasonable ranges exist for good performance.
Limitations and Future Work
- 0/1 Stoichiometry Restriction: The current implementation is primarily restricted to reactions with 0/1 stoichiometry. The authors suggest dropping this restriction for a limited number of reaction schemas as a future direction.
- Latent Species: Reactmine does not currently handle latent (unobserved) species. The authors propose generalizing the approach to infer models with latent species, which would address the challenge of inconsistent transitions.
- Hyperparameter Tuning: While grid search is used for hyperparameter tuning, the authors acknowledge the need for better understanding of their sensitivity and potentially more automated or data-driven methods for their selection.
- Computational Cost: The grid search for hyperparameters can be computationally intensive, as noted in the supplementary material.
Summary updated