A retrospective methodological evaluation demonstrates that coupling modifiable tiling microarray resequencing with supervised stacked ensemble neural networks achieves over ninety-nine percent per-position base-calling accuracy in high-viral-load clinical specimens of severe acute respiratory syndrome coronavirus 2. The investigation, conducted by researchers affiliated with the University of New Mexico and Centrillion Technologies, examined an alternative genome surveillance framework designed to bypass some of the operational bottlenecks of standard sequencing platforms. While next-generation sequencing remains the primary benchmark for viral characterization, its elevated costs, infrastructure footprints, and protracted turnaround schedules can restrict distributed monitoring during prolonged outbreaks. Resequencing microarrays leverage established hybridization and optical imaging equipment to deliver rapid, multiplexed assays across viral targets, yet historical implementations frequently suffered from ambiguous hybridization signals, cross-hybridization artifacts, and signal saturation across varied exposure windows. By extracting high-dimensional image feature representations and applying multi-stage ensemble machine learning architectures, the authors sought to establish whether systematic base-calling accuracy and single-nucleotide variant identification could be substantially improved in authentic clinical diagnostic specimens.
The study evaluated archived clinical SARS-CoV-2 diagnostic specimens originally collected by the Wyoming Public Health Laboratory. Thirty-five de-identified positive samples were processed on a custom full-genome tiling microarray platform that mapped twenty-nine thousand eight hundred forty-six genomic positions across the viral genome. For every targeted position, the physical array incorporated eight distinct oligonucleotide probes covering all four canonical nucleotides across both sense and antisense strands. Hybridization signals were optically scanned at three independent exposure intervals of zero point two five seconds, one point zero second, and four point zero seconds. Rather than relying on simple summary intensity thresholds, the analytical pipeline extracted the mean pixel intensity and the standard deviation of pixel intensities across both strands and all three exposure periods. This configuration produced forty-eight continuous numerical features for every single genomic position, establishing a rich multidimensional input space designed to capture subtle binding dynamics, saturation thresholds, and probe-level technical variability across the viral genome.
To train predictive classifiers while preserving clean benchmarks, the investigative team partitioned the thirty-five clinical specimens into twenty training samples, seven internal validation samples, and eight candidate held-out test samples possessing matched Illumina whole-genome sequencing comparators. Because clinical specimens exhibit higher signal-to-noise ratios when viral loads are elevated, candidate testing was strictly restricted to specimens displaying diagnostic cycle threshold values less than or equal to twenty-four. Four clinical specimens fulfilled this inclusion requirement: specimen WYOM thirty-six with a cycle threshold of twenty-three point nine, specimen WYOM forty-one with a cycle threshold of twenty-two point seven, specimen WYOM fifty-nine with a cycle threshold of twenty-one point seven, and specimen WYOM sixty-four with a cycle threshold of twenty-two point seven. The primary independent biological units across all experimental splits remained the individual clinical samples, although individual specimens yielded substantial numbers of per-position training and validation examples.
Supervised learning requires reliable ground-truth target designations, which the investigators addressed by establishing two primary preprocessing parameters designated as mid-select and consensus type. The mid-select parameter evaluated the empirical exclusion of terminal genomic positions using boundary trimming values of zero, two hundred, one thousand, four thousand, and ten thousand bases to mitigate known array edge effects and unstable hybridization near sequence termini. To address inherent sequencing errors, seven distinct consensus-type preprocessing protocols were implemented across reference, hybrid, and non-hybrid frameworks. Consensus majority agreement tallied how many of the six strand-exposure calls supported the dominant nucleotide at each site. Thresholds requiring agreement across at least four, five, or six calls were paired with hybrid schemes that modified reference targets only when stringent empirical evidence supported an alternate variant call, or non-hybrid schemes that excluded positions lacking unanimous or near-unanimous agreement.
Model exploration involved constructing and training more than twelve thousand individual feed-forward neural network base architectures using Keras with the Adam optimizer and categorical cross-entropy loss functions. Network architectures incorporated systematic sweeps across depth from three to eighteen layers, first-layer widths from five to eleven neurons, minibatch sizes from one hundred to five thousand examples, and activation functions encompassing rectified linear units, scaled exponential linear units, and hyperbolic tangent functions. Input features underwent L2-normalization prior to model execution. Each network mapped the forty-eight array features into a five-element SoftMax output probability distribution representing adenine, cytosine, guanine, thymine, and an explicit ambiguous no-call state designated as N. This no-call category functioned to prevent forced classifications at positions displaying low posterior confidence or ambiguous hybridization characteristics.
Ensemble construction relied on stacked generalization applied to the highest-performing base classifiers. The researchers ranked all candidate base models by standardized validation accuracy, retained the top one thousand candidates, and filtered them via coefficient of variation metrics to verify training reproducibility across replicate runs. This screening retained five hundred eleven consistent models, which were subsequently trimmed to two hundred seventy-eight base models that each exhibited validation accuracies of at least ninety-nine percent. Probability outputs from these two hundred seventy-eight base classifiers were concatenated at each genomic locus to construct a meta-level feature vector containing one thousand three hundred ninety input attributes. Two distinct meta-model architectures were subsequently trained on sample-level validation partitions using reference targets: a multinomial logistic regression classifier with inverse regularization strength set to zero point twelve, and an independent feed-forward neural network classifier trained across two hundred fifty epochs with validation-loss checkpointing.
Evaluation across the four independent held-out clinical test specimens demonstrated that both stacked meta-models markedly surpassed individual base models in classification performance. In aggregated one-versus-all confusion matrix analyses spanning the four test samples, individual base models yielded four hundred seventy-five thousand two hundred eighty-six true negatives and one hundred seventeen thousand one hundred thirty-four true positives. In comparison, the logistic regression meta-model yielded four hundred seventy-six thousand six hundred fifty-nine true negatives and one hundred eighteen thousand five hundred seven true positives, while the neural network meta-model yielded four hundred seventy-six thousand six hundred fifty-five true negatives and one hundred eighteen thousand five hundred three true positives. Across the four specimens, macro-averaged sensitivity ranged from ninety-seven point ninety-five percent to ninety-nine point seventy percent, specificity ranged from ninety-nine point thirty-two percent to ninety-nine point ninety percent, and overall per-position accuracy spanned ninety-eight point ninety-five percent to ninety-nine point eighty-six percent.
To evaluate utility for actionable epidemiological tracking, the investigators examined model performance in recovering single-nucleotide variants verified by matched Illumina sequencing. Across the four cycle-threshold-qualified test samples, Illumina sequencing identified thirty-two single-nucleotide variant positions across the twenty-nine thousand eight hundred forty-six array sites. The neural network meta-model correctly detected both the genomic location and the specific mutated nucleotide for twenty-seven of these thirty-two variants, achieving an overall variant recovery accuracy of eighty-four point thirty-eight percent.
Analysis of base-calling quality scores revealed an important operational trade-off between genomic coverage and sequence accuracy. The quality score, formulated from the maximum SoftMax class probability, demonstrated that stacked meta-models applied substantially stricter filtering than unstacked base models. At a quality score threshold of twenty, the logistic regression meta-model achieved accuracies between ninety-nine point eighty-six percent and ninety-nine point ninety-seven percent across the four test specimens, while maintaining genomic coverage between ninety-four point twenty-two percent and ninety-eight point thirty-six percent. In unstacked base models, coverage declined noticeably only above quality scores of thirty-five, whereas stacked ensemble coverage experienced steeper attrition above quality scores of twenty-five. The authors observed that this sharper decline reflects the greater discriminative stringency of the stacked architectures, which actively suppress uncertain classifications to elevate overall sequence fidelity.
The study authors emphasized several critical methodological limitations and interpretive cautions regarding these findings. Because genomic positions within an individual specimen share identical extraction, hybridization, and imaging conditions, measurements across adjacent loci are statistically correlated. Consequently, computing performance over hundreds of thousands of per-position instances creates pseudo-replication that inflates apparent sample sizes, meaning per-position accuracy figures cannot be interpreted as performance estimates across hundreds of thousands of independent clinical patients. Furthermore, evaluation was restricted to archived early isolates with cycle threshold values of twenty-four or lower, leaving operational performance in lower-viral-load samples, degraded specimens, or complex contemporary lineages with heavy mutation burdens to be determined in future multi-laboratory investigations.
Subject of Research: Microarray-based viral genome resequencing evaluated with stacked ensemble machine learning models for SARS-CoV-2 surveillance
Article Title: Viral genome surveillance via modifiable microarray sequencing and a supervised stack ensemble neural network model: SARS-CoV-2 as a case study
Article References: Plackowski, K., Ebinumoliseh, I., Ogbuewu, E., Hoff, K., & Edwards, J. S. (2026). Viral genome surveillance via modifiable microarray sequencing and a supervised stack ensemble neural network model: SARS-CoV-2 as a case study. Frontiers in Analytical Science, 6, 1834714. https://doi.org/10.3389/frans.2026.1834714
Image Credits: AI Generated
DOI: 10.3389/frans.2026.1834714
Keywords: SARS-CoV-2, Genomic Surveillance, Microarray Resequencing, Stacked Ensemble Learning, Neural Network Models, Base Calling, Variant Identification, Illumina Sequencing, Machine Learning, Hybridization Signal, Cycle Threshold, Quality Score, Supervised Learning, Multinomial Classification, Logistic Regression Meta-Model, Viral Genomics
News Source: Cassandra Pierce. (November 11, 2026). Microarray Resequencing with Stacked Neural Networks Yields Accurate SARS-CoV-2 Surveillance. Scienmag.



