Abstract
In this paper, we propose a new method to identify biochemical reaction networks (i.e. both reactions and kinetic parameters) from heterogeneous datasets. Such datasets can contain (a) data from several replicates of an experiment performed on a biological system; (b) data measured from a biochemical network subjected to different experimental conditions, for example, changes/perturbations in biological inductions, temperature, gene knock-out, gene over-expression, etc. Simultaneous integration of various datasets to perform system identification has the potential to avoid non-identifiability issues typically arising when only single datasets are used.
Introduction
Reverse engineering biological systems is a fundamental challenge in systems and synthetic biology. While time-series data of molecular concentrations (e.g., mRNA or proteins) are increasingly available, determining the underlying "wiring diagram" of reactions and their associated kinetic parameters remains difficult. A primary obstacle is non-identifiability, where multiple distinct models can explain the same single-experiment dataset. To overcome this, researchers utilize heterogeneous datasets—replicates or experiments under varying conditions like temperature shifts or genetic perturbations. However, integrating these diverse data sources into a robust mathematical framework is computationally intensive and often leads to information loss through simple averaging.
Methodology: Sparse Bayesian Learning
The proposed approach utilizes a unified optimization framework grounded in Sparse Bayesian Learning (SBL). The core of the method involves a "dictionary" of potential biological interactions, including mass-action kinetics and Hill functions.
The algorithm exploits the principle of block-sparsity. It assumes that while the dictionary of possible interactions is large, the actual biological network is sparse. By organizing data from multiple experiments into a block structure, the method identifies interactions that are consistently active across all datasets.
The implementation uses the Alternating Direction Method of Multipliers (ADMM) to ensure scalability. The iterative process alternates between a reweighted Group Lasso step for reaction selection and a reweighted inverse covariance estimation step to account for data noise and uncertainty.
Experimental Results
The method was validated using a simulated eight-species generalized repressilator, a synthetic gene circuit with a ring topology of mutual repression. Heterogeneity was introduced by randomizing initial conditions and varying kinetic parameters by up to 20% to mimic natural biological variation.
The proposed algorithm demonstrated significant improvements over traditional Group Lasso techniques. In noiseless scenarios, the algorithm achieved a Root Normalized Mean Square Error (RNMSE) approaching zero as the number of experiments ($C$) and the length of time series ($M$) increased.
In contrast, the standard Group Lasso approach plateaued at an RNMSE of approximately 0.75, failing to reconstruct the network accurately even with additional data. The SBL-based method successfully mapped external perturbations to internal dynamics, identifying which specific experimental changes influenced particular parameters.
Discussion and Future Directions
A key strength of this framework is its ability to extract information from variance across experiments without requiring a priori knowledge of specific kinetic forms (e.g., whether a protein acts as an activator or repressor). By treating noise covariance as an unknown variable, the model maintains robustness against experimental "messiness."
Future work will address current limitations, such as sensitivity to the sampling rate of time-series data and high levels of dynamic noise. The authors aim to define the minimal sampling requirements for successful identification and transition from simulated models to real-world "wet lab" datasets from synthetic biology and natural metabolic pathways.
Summary updated