Machine Learning Enables Computationally Efficient and Stable Integration of Genome-Scale Metabolic Networks with Reactive Transport Models to Predict Microbial Metabolic Switching

Authors

Hyun-Seob Song1* ([email protected]), Firnaaz Ahamed1, Joon-Yong Lee2, Christopher Henry3, Janaka Edirisinghe3, William Nelson2, Xingyuan Chen2, David Moulton4, Tim Scheibe2

Institutions

1University of Nebraska–Lincoln, Lincoln, NE; 2Pacific Northwest National Laboratory, Richland, WA; 3Argonne National Laboratory, Lemont, IL; 4Los Alamos National Laboratory, Los Alamos, NM

URLs

Abstract

Genome-scale metabolic networks (GEMs) provide a detailed view of microbial metabolism and its interactions with the environment. In contrast with coarse-grained biogeochemical models, the high-resolution description of microbial processes in GEMs can help establish a molecular-level understanding of microbially driven biogeochemical cycles by incorporating high-throughput omics data. However, integrating GEMs with reactive transport models (RTMs) presents a challenge due to the significant computational burden caused by iterative implementation of linear programming (LP) to obtain flux balance analysis (FBA) solutions in each time and spatial grid during simulations.

To overcome this challenge, the research team developed a novel machine learning-based method that efficiently integrates FBA with RTM. The main breakthrough is to train artificial neural networks (ANNs) as a surrogate FBA model and use the resulting reduced-order FBA models as source and sink terms in RTM. The team’s case study of the Shewanella oneidensis MR-1 strain demonstrates the effectiveness of the proposed method. The simulation of this organism’s growth presents an additional challenge due to the intricate dynamics of metabolic switches among multiple substrates. During the aerobic growth on lactate, S. oneidensis produces metabolic byproducts (e.g., pyruvate and acetate), which are subsequently consumed as alternative carbon sources when the preferred ones are depleted.

No computational methods that allow for the simulation of such intricate dynamics by combining FBA (or its surrogate models) with RTM are currently available. The research team addressed this issue by adopting a cybernetic approach that predicts metabolic switches as the product of dynamic competition among multiple growth options. In both zero-dimensional batch and one-dimensional column configurations, ANN-based reduced-order models achieved a significant reduction in computational time, up to several orders of magnitude, compared to the original LP-based FBA models. Additionally, the ANN models generated robust solutions without the need for special measures to prevent numerical instability.

Due to such promising properties, the ANN-based FBA model developed in this work is currently integrated as a key component of CompLab, a recently developed Lattice Boltzmann-based modeling tool that simulates fluid flow and solute transport in porous media. This new capability will also be integrated into the MAGIEC (Metagenome Integration with Ecosystem Models) pipeline, which is being developed in collaboration with the KBase and IDEAS–Watersheds teams. Altogether, this work significantly improves ability to link molecular-level data and models with large-scale ecosystem modeling with enhanced computational efficiency.