Primary source bioRxiv: ML4SD: Leveraging Machine Learning and High-Throughput Search Algorithms for an Iterative Growth-Coupled Design Innovation
A preprint posted on bioRxiv on 16 September 2026 by Alvaro Garantilla Becerra and Juan Nogales (Systems Biotechnology Group, Centro Nacional de Biotecnologia, CSIC, Madrid) describes ML4SD, a machine-learning loop for designing gene-knockout strains. This note covers the abstract only.
The work has not been peer reviewed, and all results are computational (in silico). The abstract reports no experiments in a bioreactor.
The problem it targets
In growth-coupled (GC) production, making the target product is required for the cell to grow. The abstract says that building knockout strains that deliver GC production from a list of candidate genes is labor and time demanding. As a result few designs are tested in vivo, which hampers standard machine-learning methods that try to learn GC patterns for a specific bioprocess.
What the authors built
ML4SD is an active-learning Design-Build-Test-Learn cycle. It trains ensembles of models on scores from genome-scale metabolic models (GEMs) for knockout designs. It then samples the next designs based on the predicted model performance and error.
The abstract states that the cycle generalizes only if the initial library is large and diverse, including suboptimal and non-viable designs. Libraries limited to minimal designs or Pareto-optimal knockouts led to overfitting models. To build such libraries the authors developed a strain design algorithm, gcSwarms, and tested it on a diverse set of bioprocesses.
The case study
ML4SD was tested in silico on converting lignin-derived 4-hydroxybenzoate to 6-caprolactam, the nylon-6 monomer. The abstract reports improvements in carbon yield (the share of substrate carbon ending up in product). The abstract gives no baseline for the carbon yield gain, so this note does not quote the figure.
On efficiency, ML4SD reached its result with 2.5- to 7.1-fold fewer designs than a search using gcSwarms alone. The authors present this as data efficiency. Code is linked on GitHub (extrevaro/ML4SD).
