The Extra Molecular Signal Has a Different Job

TL;DR for operators

Adding another molecular signal to a diagnostic AI system does not necessarily improve its classification accuracy. MOXAI demonstrates that combining DNA methylation and copy number information can reduce variability across validation folds even when methylation alone remains more accurate.

The study also separates two qualities often treated as interchangeable: predictive performance and explanation quality. DINOv2 achieves 81.30% accuracy on an independent glioblastoma cohort, compared with 67.48% for ResNet50, but ResNet50 produces more localized genomic attribution maps. DINOv2’s reported 94.25% accuracy applies only to predictions exceeding a 0.9 confidence threshold.

For diagnostic software developers and evaluators, the operational lesson is to assess four properties independently: full-cohort accuracy, prediction stability, confidence-dependent coverage, and explanation usefulness. These retrospective findings inform evaluation design, not clinical deployment readiness.

A diagnostic model has more than one performance requirement

Consider two AI systems being evaluated for molecular tumor classification. One predicts subtypes more accurately on an independent dataset. The other identifies more specific genomic regions associated with its predictions. Adding another molecular measurement improves consistency across training partitions but does not reliably improve accuracy.

Selecting a system using its highest reported classification score would overlook these differences. The decision depends on whether the intended workflow prioritizes reliable classification, interpretable molecular evidence, or the ability to identify cases requiring further assessment.

MOXAI, developed by Alom and coauthors, investigates precisely these competing properties in adult-type diffuse glioma classification.¹ Its contribution is not simply another neural architecture. It provides a representation that allows multiple molecular signals from one assay to be modeled together, then tests whether those signals improve prediction and biological interpretability.

One molecular assay, two sources of information

DNA methylation (DNAM) describes chemical modifications associated with DNA regulation. Copy number variation (CNV) reflects gains and losses of genomic material.

Both can be derived from a methylation array, making their integration attractive for molecular diagnostic software. A combined classifier could exploit complementary information without necessarily requiring a second molecular assay.

MOXAI represents these measurements as genome-ordered matrices. The authors retain 408,768 CpG probes from Illumina HumanMethylation450 arrays and arrange their measurements into zero-padded 640-by-640 matrices. DNAM and CNV can enter the model separately or as two channels of the same representation.

This transformation makes high-dimensional molecular data compatible with image-based architectures, including ResNet50 and DINOv2. A separate graph attention network (GAT) organizes mapped probes hierarchically through genes and chromosomes.

The representation is consequential because genomic position carries information. Copy number abnormalities frequently extend across neighboring genomic regions. Preserving approximate genomic adjacency gives image-based models access to patterns that arbitrary probe arrangements may obscure.

The matrix is nevertheless an engineered representation, not a literal image of genomic geometry. Row boundaries and zero padding introduce artificial spatial relationships.

Genomic ordering can matter more than adding another input

The study benchmarks 551 TCGA tumor samples across six molecular classes. It compares DNAM-only, CNV-only, and combined DNAM+CNV configurations, alongside model architecture and probe ordering.

A particularly informative representation comparison concerns ResNet50 trained on CNV alone.

Source: MOXAI, Section IV.A, Figure 2.

The chromosome-ordered representation improves reported mean accuracy by 11.02 percentage points under this specific architecture and modality.

This is evidence that representation design can materially influence performance. It does not establish an equivalent gain for every architecture or molecular input.

For AI developers working with structured scientific measurements, the implication is worth testing directly: preserving relationships already known in the underlying domain may change what an existing architecture can learn, even without adding features.

The second signal improves stability more reliably than accuracy

Across the reported benchmarks, DNAM generally distinguishes glioma molecular classes more effectively than CNV alone.

Combining the two sources sometimes helps, but the improvement is not uniform. In one chromosome-ordered ResNet50 comparison, multimodal mean accuracy is 84.91%, slightly below the 85.76% reported for CNV alone.

The more distinctive multimodal result concerns variation across cross-validation folds.

For the reported chromosome-ordered comparison, the standard deviation of classification accuracy is approximately 2.62 percentage points with combined inputs, versus 4.76 percentage points with DNAM alone.

These are measures of variation across evaluation folds, not confidence intervals for patient-level predictions. Lower variability suggests more consistent performance across the tested partitions; it does not establish reliability under future clinical distribution shifts.

The paper’s abstract also reports multimodal cross-validation accuracy reaching 92.98%, but that headline should not be treated as describing every architecture or ordering condition.

The operational interpretation is that supplementary molecular information can have a different role from the dominant predictive signal. CNV may contribute to consistency even when its addition does not maximize average accuracy.

For developers, that gives multimodal integration a testable justification beyond pursuing the highest benchmark score.

Independent validation changes the ranking—and confidence changes the population

The study evaluates trained models on an independent cohort of 492 glioblastoma samples. This cohort contains three glioblastoma molecular subtypes, rather than all six classes represented in the development dataset.

DINOv2 performs substantially better than ResNet50 on this independent evaluation.

Source: MOXAI, Section IV.F, Figure 6.

DINOv2’s full-cohort advantage is 13.82 percentage points. That is meaningful evidence for stronger generalization within this independent three-class evaluation.

The 94.25% result answers a narrower question: how accurate is DINOv2 when it makes predictions exceeding a particular confidence threshold?

Filtering out lower-confidence predictions changes which samples contribute to the accuracy calculation. Higher selective accuracy may therefore come with fewer retained predictions.

Figure 6 examines this relationship, but the source record does not establish the exact number of samples retained above the 0.9 threshold. Consequently, the 94.25% figure cannot characterize the accuracy available for every incoming case, nor can its operational coverage be determined precisely from the reported narrative.

For a diagnostic workflow that permits deferral of uncertain cases, confidence filtering could be a useful design option. Its effectiveness would depend on the proportion deferred, the characteristics of those cases, and the reliability of downstream assessment.

Better classification does not guarantee better genomic explanations

MOXAI also uses class activation mapping (CAM) and gradient-weighted CAM (Grad-CAM) to highlight genomic regions associated with model predictions.

These methods translate importance estimates from the model’s spatial representation back toward CpG sites and chromosomal regions.

ResNet50 produces more localized and biologically differentiated attribution maps than DINOv2, whose patch-based representation provides coarser localization.

The reported important regions correspond to several established glioma-associated characteristics, including:

• Chromosome 1 and the 1p/19q signature associated with oligodendroglioma.
• Chromosome 7 abnormalities associated with EGFR.
• Chromosome 10 abnormalities associated with PTEN loss.

The oligodendroglioma chromosome-arm analysis additionally reports a statistically distinct 19q attribution pattern at p < 0.05.

These analyses assess biological concordance: whether highlighted regions correspond to known molecular characteristics. They do not establish that the highlighted loci causally determine the prediction or that the attribution faithfully reconstructs the model’s decision process.

CAM and Grad-CAM can also produce different importance distributions for the same architecture and molecular input. That disagreement makes explanation-method choice part of evaluation.

The architecture tradeoff is therefore substantive. DINOv2 offers stronger independent predictive performance, while ResNet50 offers more localized genomic explanations. Neither result alone establishes which architecture best serves a particular clinical workflow.

Four evaluation criteria for diagnostic AI procurement

Cognaptus interprets these results as a reason to separate four questions when assessing molecular AI software.

These criteria need not favor the same model.

A software developer considering CNV integration might justify the additional processing through improved stability rather than greater mean accuracy. A diagnostic evaluator comparing DINOv2 with ResNet50 might prioritize predictive generalization while separately requiring genomic attribution at a defined resolution.

Neither decision should be made from the highest reported score alone.

What remains outside the evidence

MOXAI is a retrospective proof of concept. Its 551-sample development cohort is small relative to the molecular feature space, and the six classes are unevenly represented. Independent validation covers only three glioblastoma subtypes.

The reported experiments do not establish prospective diagnostic effectiveness, improvement in patient outcomes, or validated within-class prognostic risk assessment.

Attribution agreement with known biomarkers supports biological plausibility, but it is not causal validation. Additional highlighted regions, such as chromosome 18 in AIDH, remain hypotheses requiring independent biological investigation.

These boundaries matter because the intended clinical decision may involve populations, outcomes, and consequences not represented by the benchmark.

The value of the extra signal depends on what is measured

MOXAI’s most transferable contribution is its genome-ordered representation of molecular measurements from a single assay. The subsequent comparisons show why its value cannot be assessed through classification accuracy alone.

Additional molecular information may reduce evaluation variability without producing the highest accuracy. A model that generalizes better may provide coarser genomic explanations. Confidence thresholds may improve accuracy on retained predictions while leaving the disposition of excluded cases unresolved.

For clinical AI developers and diagnostic software evaluators, these findings support a more discriminating evaluation process: determine what the extra signal contributes, identify which cases a reported accuracy actually covers, and validate explanations independently of predictive success.

That is a sounder basis for further diagnostic research than treating the best benchmark result as evidence of clinical readiness.

Cognaptus: Automate the Present, Incubate the Future.