Machine Learning in Genomic Engineering

Abstract

Genomic engineering has come a long way since the early days of restriction enzymes and random mutagenesis, but the design space is still vast enough that brute-force experimental screening cannot cover it efficiently. Machine learning offers a way to navigate that space more intelligently, predicting which guide RNA sequences will cut cleanly, which promoter variants will drive the right expression level, and which protein mutations will produce a desired functional change, all before anyone touches a pipette. The catch is that genomic data is noisy, biologically context-dependent, and often scarce for the exact application a researcher cares about. We surveyed 230 machine learning programmes applied to genomic engineering tasks, drawn from laboratories in Italy and Estonia between 2019 and 2023, spanning five application areas: CRISPR guide design, gene expression prediction, protein function optimisation, regulatory element engineering, and synthetic genome design. A Machine Learning Genomic Engineering Quality Index (MLGEQI) was built from five sub-scores -- prediction accuracy, biological interpretability, wet-lab validation rate, generalisability across cell types, and computational accessibility -- with weights from regression against translational uptake. MLGEQI predicted adoption into active genomic engineering pipelines with r = +0.84 and AUC = 0.884. CRISPR guide design programmes scored highest (mean MLGEQI 0.823), while synthetic genome design trailed at 0.604. Only 36.5 percent of programmes exceeded 0.75.