Reconstructing Kinetic Models for Dynamical Studies of Metabolism using Generative Adversarial Networks

Abstract

Kinetic models of metabolism relate metabolic fluxes, metabolite concentrations and enzyme levels through mechanistic relations, rendering them essential for understanding, predicting and optimizing the behaviour of living organisms. However, due to the lack of kinetic data, traditional kinetic modelling often yields only a few or no kinetic models with desirable dynamical properties, making the analysis unreliable and computationally inefficient. We present REKINDLE (Reconstruction of Kinetic Models using Deep Learning), a deep-learning-based framework for efficiently generating kinetic models with dynamic properties matching the ones observed in cells. We showcase REKINDLE’s capabilities to navigate through the physiological states of metabolism using small numbers of data with significantly lower computational requirements. The results show that data-driven neural networks assimilate implicit kinetic knowledge and structure of metabolic networks and generate kinetic models with tailored properties and statistical diversity. We anticipate that our framework will advance our understanding of metabolism and accelerate future research in biotechnology and health.

Generative machine learning produces kinetic models that accurately characterize intracellular metabolic states

“…However, acquiring the parameters of these models with traditional kinetic modeling approaches is computationally expensive and arduous 15,32 . To improve the efficiency of generating kinetic models, we have recently proposed REKINDLE 32 . This unsupervised deep-learning method uses generative adversarial networks (GANs) 52 for this task.…”

“…A parameterized kinetic model exhibits highly nonlinear but deterministic responses that depend on the intracellular state determined by the network topology and the integrated data. To capture this nonlinear behavior and determine kinetic parameters, we require function approximators with similar complexity, such as neural networks 32 . In RENAISSANCE, we iteratively optimize the weights of feed-forward neural networks (generators) using NES (Figure 1a) to obtain kinetic parameters leading to biologically relevant kinetic models, meaning that the metabolic responses obtained from these models have experimentally observed dynamics (Methods).…”

“…Recent efforts employing new tailor-made parametrization 27 and machine learning [32][33][34] improved the efficiency of constructing near-genome-scale kinetic models. Nevertheless, challenges remain regarding extensive computational time 27 and the need for training data from traditional kinetic modeling approaches [32][33][34] .…”

Rational strain design with minimal phenotype perturbation

“…These kinetic models consist of a system of nonlinear ordinary differential equations (ODEs) characterized by a set of kinetic parameters. To generate such models, we can use traditional kinetic modeling approaches such as MASSPy [17], Ensemble Modeling [19], ORACLE [15], [18], [26], [27], and machine learning empowered methods such as REKINDLE [21] and iSCHRUNK [20] [22].…”

How artificial intelligence enables modeling and simulation of biological networks to accelerate drug discovery

“…Notably, systems biologyinformed AI can be used to select among different reaction kinetic schemes (e.g., mass-balance, S-systems, log-linear/linlog, Michaelis-Menten, generalized Hill, convenience, modular) according to the available mechanistic knowledge, experimental datasets, and desired level of details (Du et al, 2016;Kim et al, 2018). Several approaches combining deep learning with GNNs or GANNs have been recently used to fully parametrize ODE-based genome-scale metabolic models in terms of Michaelis-Menten affinity constants (Choudhury et al, 2022) or enzyme turnover numbers (Li et al, 2021), respectively. Bayesian meta-modeling is another systems biology-based approach that uses different mathematical representations, scales, and levels of granularity from prior models in order to simulate cell activity (Raveh et al, 2021).…”