Machine Learning-Driven Prediction of Antimicrobial Resistance Phenotypes from Whole-Genome Sequence Data of Klebsiella pneumoniae Isolated Across Nigeria

Whole-genome sequencing (WGS)-based antimicrobial resistance (AMR) prediction offers transformative potential to replace time-intensive phenotypic susceptibility testing, but predictive models must be trained on local genomic diversity to perform reliably. This study developed and validated machine learning models for AMR phenotype prediction in Klebsiella pneumoniae using whole-genome sequence data from isolates collected across all six geopolitical zones of Nigeria. Three hundred K. pneumoniae isolates from clinical, community, and environmental sources were sequenced on Illumina and Nanopore platforms. Phenotypic susceptibility was determined by broth microdilution for 14 antibiotics. Genome-derived feature sets including resistance gene presence, SNP profiles, and k-mer vectors were used as model inputs. Five machine learning algorithms were compared: random forest, gradient boosting, support vector machine, deep neural network, and logistic regression. Model performance was evaluated by ten-fold cross-validation and tested on a held-out external validation set. The gradient boosting model achieved the highest overall predictive accuracy at 94.7 percent, with sensitivity of 96.2 percent and specificity of 93.8 percent for third-generation cephalosporin resistance. Carbapenem resistance prediction was the most challenging, with accuracy of 88.3 percent. Resistance gene-based features outperformed k-mer features for most antibiotic classes. Model performance was stable across geopolitical zones after training on pooled Nigerian data. This work establishes the first Nigeria-specific WGS-based AMR prediction model and provides a deployable framework for rapid genomic AMR reporting in national reference laboratories. Keywords: machine learning, WGS, Klebsiella pneumoniae, AMR prediction, Nigeria.