Deep Learning Framework Scales Enzyme-Constrained Models Across Diverse Organisms

Researchers Rana A. Barghout, Lya Chiñas Serrano, Benjamin Sanchez-Lengeling, and Radhakrishnan Mahadevan have developed an integrated computational framework combining CPI-Pred and kinGEMs to generate enzyme-constrained genome-scale models. In systems biology, the construction of accurate enzyme-constrained genome-scale models, commonly designated as ecGEMs, has remained a critical challenge. This modeling difficulty stems primarily from sparse kinetic data alongside the persistent need for biologically meaningful representations.

To resolve these longstanding limitations, the research team created an integrated approach linking two computational components. First, CPI-Pred operates as a deep learning model that predicts kinetic parameters, specifically turnover numbers denoted as kcat, directly from biological sequence and compound embeddings. Next, the kinGEMs pipeline incorporates these uncertainty-aware parameters into ecGEMs for metabolic optimization. By systematically leveraging representations at multiple scales, the framework captures sequence, structure, and kinetic data to enhance overall model generalizability and accuracy.

Rigorous benchmarking demonstrates the framework’s capability to predict cellular growth rates and metabolic fluxes that remain consistent with experimental observations and measured proteome fractions. Crucially, the system accounts explicitly for the uncertainty inherent in machine learning-predicted kinetic parameters.

The scalability of kinGEMs was confirmed when the pipeline successfully generated ecGEMs for ninety-three models spanning a phylogenetically diverse set of target organisms. This broad biological scope encompassed Gram-negative bacteria, Gram-positive bacteria, mycobacteria, parasitic protists, fungi, and mammalian cell lines, illustrating that the uncertainty-aware modeling pipeline scales reliably across diverse taxonomic groups.

These methodological innovations breakdown the long-standing barrier to the practical application of ecGEMs to non-model organisms.

Subject of Research: Machine learning integration for enzyme-constrained genome-scale metabolic models

Article Title: kinGEMs: A robust and scalable framework for resource-constraint models through stochastic tuning of deep learning-predicted kinetic parameters

Article References: A. Barghout, R., Chiñas Serrano, L., Sanchez-Lengeling, B., & Mahadevan, R. (2026). kinGEMs: A robust and scalable framework for resource-constraint models through stochastic tuning of deep learning-predicted kinetic parameters. PLOS Computational Biology, 22(11), e1014881. https://doi.org/10.1371/journal.pcbi.1014881

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014881

Keywords: systems biology, metabolic modeling, enzyme-constrained models, deep learning, kinetic parameters, synthetic biology, metabolic engineering, computational biology, kinGEMs, robust, scalable, framework