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

kinGEMs generates enzyme-constrained genome-scale models for 93 phylogenetically diverse organisms by stochastically tuning deep learning–predicted kcat parameters to better match growth and flux data.

• Why it matters: Accurate ecGEMs are constrained by sparse kinetic measurements and the difficulty of representing enzyme kinetics in a biologically meaningful way. Uncertainty from ML-predicted parameters also needs to be handled explicitly to avoid misleading metabolic optimization.
• What they did: The framework combines CPI-Pred, which predicts kcat from sequence and compound embeddings, with kinGEMs, which incorporates uncertainty-aware parameters into ecGEMs for metabolic optimization. It uses multi-scale representations and benchmarks predicted growth rates, fluxes, and proteome fractions.
• The result: Benchmarking shows predicted growth rates and fluxes consistent with experimental observations and measured proteome fractions, while reducing median flux variability by several folds. This enables better-constrained metabolic models and supports ecGEM construction beyond model organisms.

AI summary of the abstract · may contain errors · not medical advice · read the paper.