Accelerated design ofEscherichia coligenomes with reduced size using a whole-cell model and machine learning surrogate

Abstract

AbstractWhole-cell models (WCMs) are multi-scale computational models that aim to accurately simulate the function of all genes and biological processes within a cell. While WCMs offer deeper insights into why cells behave and respond in specific ways, they also require significant computational resources to run, making their development and use challenging. To address this limitation it is possible to use simpler machine learning (ML) surrogates that can learn to approximate specific behaviours of larger and more complex models, while requiring only a fraction of the computational resources. Here, we show how ML surrogates can be trained on WCM outputs to accurately predict whether cells divide successfully when a subset of genes are removed (knocked out). We used these surrogates and a genome-design algorithm to generate a reducedE. colicellin silico, where 39% of the genes included in the WCM were removed, a task made possible by our ML surrogate that performs simulations up to 16 times faster than the original WCM. These results demonstrate the value of adopting WCMs and ML surrogates for enabling genome-wide engineering of living cells, offering promising new routes for biologists to understand better how cellular phenotypes emerge from genotypes.