Viral genome surveillance via modifiable microarray sequencing and a supervised stack ensemble neural network model: SARS-CoV-2 as a case study

Introduction Continued SARS-CoV-2 evolution underscores the need for scalable genomic surveillance methods that are faster and less resource-intensive than conventional sequencing alone. This study evaluated a viral genome surveillance strategy that combines modifiable microarray-based resequencing with supervised ensemble machine learning to improve base-calling accuracy. Methods Building on a previously reported full-genome tiling microarray platform, we extracted expanded image-derived features from resequencing arrays. Features included probe intensity means and standard deviations across sense and antisense strands and multiple exposure times, yielding 48 input features per genomic position. Using 35 clinical SARS-CoV-2 samples, we created training, validation, and held-out test datasets and trained more than 12,000 neural network models under multiple preprocessing and hyperparameter configurations. Top-performing base models were combined using stacked ensemble learning with logistic regression and neural network meta-models. Results In held-out testing on clinical samples with cycle threshold ≤24, the stacked ensemble models achieved high per-position base-calling accuracy (>99%). Because accuracy was computed across thousands of correlated genomic positions per specimen, these values should be interpreted as per-position performance rather than as an estimate of performance across independent clinical specimens. Discussion In this proof-of-concept evaluation using a limited set of archived clinical SARS-CoV-2 samples, microarray sequencing coupled with stacked ensemble learning produced high per-position base-calling accuracy in high-viral-load specimens and enabled recovery of many Illumina-confirmed variants. The approach may be adaptable to other viral targets through microarray redesign and model retraining. However, broader validation across larger, more genetically diverse and contemporary sample sets—including more mutation-rich lineages and lower viral load specimens—will be needed to establish robustness beyond this early-isolate, Ct-qualified evaluation.