PaperlayerView exact revision on arXiv

Source-linked AI summary

Evaluating Single and Multi-Omics Based Explainable Artificial Intelligence (MOXAI) for Molecular Subclass Classification of Adult-Type Diffuse Gliomas

Md Zahangir Alom, Quynh T. Tran, Breuer Alexandar, Brent A. Orr

arXiv:2609.32190v1cs.CVcs.AI

TL;DR

Existing approaches do not systematically integrate methylation and CNV data or attribute predictions to specific genomic loci. MOXAI represents array-derived genomic data spatially, applies deep-learning classifiers and attribution methods, and reports improved multimodal stability and stronger DINOv2 generalization on independent samples.

  • Problem

    Existing models typically report methylation and CNV separately, lack within-class risk assessment, and lack established methods to attribute predictions to specific genomic loci.

    Source: §I · p. 1

  • Method

    MOXAI encodes methylation and CNV array data as spatial genomic matrices and applies ResNet50, DINOv2, and GAT with CAM and Grad-CAM attribution.

    Source: §I · pp. 1–2, A. Dataset · p. 2, B. Model Architecture · p. 2

  • Results

    94.25% accuracy for DINOv2 versus 85.94% for ResNet50 on independent test samples, while multimodal models showed more stable cross-validation performance than methylation-only models.

    Source: §VI · p. 7, §V · p. 7

  • Takeaways & Limitations

    The reported CAM and Grad-CAM explanations align with established molecular features of adult glioma subtypes.

    Source: §I · p. 2, B. XAI Analysis: CNV Model · p. 5

    The TCGA cohort is predominantly adult and may not capture the diversity of clinically encountered tumors; class imbalance may bias classification performance.

    Source: §V · p. 7

Abstract

from arXiv · show

DNA methylation (DNAM) profiling has emerged as a powerful diagnostic tool for classifying brain and solid tumors. However, existing computational models typically analyze methylation and copy number variation (CNV) data separately, failing to capture the complementary information their integration could provide. Moreover, current classification models lack mechanisms for within-class risk assessment analogous to traditional tumor grading, and no established explainability method can attribute classification decisions to specific genomic loci. In this paper, we present MOXAI (Multi-Omics Based Explainable AI), a deep learning framework that integrates DNA methylation and copy number data from methylation arrays to classify molecular subtypes of adult-type diffuse gliomas, alongside single-modality variants for comparison. Using a cohort from The Cancer Genome Atlas (TCGA), we trained ResNet50, DINOv2, and Graph Attention Network (GAT) models on methylation data alone, copy number data alone, and combined multimodal data. We further developed explainable AI (XAI) methods based on class activation maps (CAMs) and gradient-weighted CAM (Grad-CAM) to identify the specific CpG sites, genes, and chromosomal regions most relevant to each classification decision. The multimodal model achieved up to 92.98% cross-validation accuracy, outperforming models trained on CNV data alone. DINOv2 showed the strongest generalization, reaching 94.25% accuracy (confidence >0.9) on independent validation sets. XAI results aligned with established molecular features of adult-type diffuse glioma subtypes, confirming the biological interpretability of the framework.

I. INTRODUCTION

The introduction identifies gaps in integrating methylation and CNV data, assessing within-class risk, and explaining predictions at genomic-locus resolution. MOXAI addresses these gaps with a spatial genomic representation for deep learning and attribution.

I. INTRODUCTION

  • MOXAI transforms genomic features into a 2D representation to apply image-based deep learning and explainability methods while retaining many of the 408,768 methylation probes.The authors note that the scale of the data makes conventional sequential networks computationally prohibitive and complicates explainability.Source: §I · p. 1
  • The motivating gaps include absent systematic integration of array-derived CNV with methylation scores, within-class risk assessment, and locus-level attribution methods.Source: §I · p. 1
  • The framework jointly models methylation and copy number data through single- and multimodal systems.Source: §I · p. 1
  • ResNet50, DINOv2, and GAT are evaluated for diffuse glioma classification using individual and combined modalities.Source: §I · pp. 1–2
  • CAM and Grad-CAM localize diagnostically informative CpG sites and chromosomal regions for each tumor class.The study evaluates TCGA adult-type diffuse glioma subtypes and independent glioma datasets; its explanations align with established molecular features.Source: §I · p. 2

A. DNA Methylation-Based Brain Tumor Classification

Prior work established methylation-based brain tumor classification and the diagnostic relevance of CNV alterations. The section frames spatially ordered multi-omics integration as an un demonstrated approach for methylation-array data.

A. DNA Methylation-Based Brain Tumor Classification

  • The Heidelberg classifier uses methylation-array beta-values in a random forest to classify more than 91 methylation classes, but omits CNV from its score and lacks locus-level explainability.Source: A. DNA Methylation-Based Brain Tumor… · p. 2

B. Copy Number Variation in Brain Tumors(Diffuse gliomas)

C. Multi-Omics Integration with Deep Learning

D. Explainable AI for Genomics

The section describes spatially resolved attribution methods for genome-ordered probes and a dataset representation that preserves genomic proximity in image-like matrices.

D. Explainable AI for Genomics

  • CAM and Grad-CAM produce spatially resolved attribution maps that can identify genomic regions driving classification decisions when adapted to genome-ordered probe matrices.Source: D. Explainable AI for Genomics · p. 2

A. Dataset

  • The TCGA cohort contains 551 samples across six classes, represented by 408,768 retained probes arranged by genomic location in 640×640 matrices.DNAM beta-values and CNV segment values are encoded as single-channel inputs or as two-channel multimodal inputs; independent testing uses 492 glioblastoma samples.Source: A. Dataset · p. 2

B. Model Architecture

The study adapts ResNet50, DINOv2, and GAT for single- and dual-channel inputs. The GAT aggregates probes through gene and chromosome representations before classification.

B. Model Architecture

  • ResNet50 and DINOv2 are adapted to single- and dual-channel input, with weights randomly initialized and trained from scratch.Source: B. Model Architecture · p. 2
  • GAT excludes probes mapped to null genes, aggregates the remaining probes into gene-level and chromosome-level representations, then classifies pooled transformer outputs with an MLP.The resulting inputs represent 297,730 probes, 17,630 genes, and 22 chromosomes.Source: B. Model Architecture · p. 3
  • The dataset figure reports 551 total samples, selecting common samples based on availability of both CNV and DNAM data.Source: Figure 1 · p. 3

C. Experimental Design

The experimental design compares single- and two-modality models, while the explainability method maps class activation scores back to CpG sites.

C. Experimental Design

  • Three conditions compare CNV-only, DNAM-only, and combined CNV+DNAM models using an 80:20 stratified split and 10-fold cross-validation.Held-out test performance was assessed with accuracy, precision, recall, and F1 score.Source: C. Experimental Design · p. 3

D. Explainable AI Methods

  • CAM scores from bottleneck-layer features are upsampled to the 640×640 input space to produce CpG-site importance heatmaps.For ResNet50, the cited feature volume is 16×16×1024 before global average pooling.Source: D. Explainable AI Methods · p. 3
  • Fig. 2 compares ResNet50 and DINOv2 validation accuracies for randomly ordered and chromosome-ordered probes.The caption identifies the two probe-ordering strategies but does not report their accuracy values.Source: Figure 2 · p. 3

IV. RESULTS

Classification performance varies with modality and probe ordering, while CNV-model explanations show subtype-specific chromosomal importance patterns that broadly align with known glioma biology.

A. Classification Performance

  • 93.84% accuracy for DNAM-only differentially significant probes and 92.30% for DNAM-only chromosome-ordered probes; multimodal ResNet50 means were 88.42% and 84.91%, versus CNV-only means of 74.74% and 85.76%.For chromosome-ordered probes, multimodal ResNet50 had lower standard deviation than DNAM alone (±2.62 vs. ±4.76); GAT testing accuracy was 62.6 ± 2.27 for DNAM alone and 75.57 ± 1.27 for multimodal.Source: A. Classification Performance · pp. 3–4

B. XAI Analysis: CNV Model

  • Chromosome-specific CAMs show elevated chromosome 9 importance for AIDH and AIDH-HG but little for OIDH, alongside a distinct chromosome 1 peak for OIDH in ResNet50.The prose relates chromosome 9 to CDKN2A deletion in IDH-mutant astrocytomas and notes chromosome 1 importance is not clearly observed in DINOv2.Source: B. XAI Analysis: CNV Model · p. 4, Figure 3 · p. 4
  • ResNet50 CAMs show chromosome 7–11 peaks for astrocytoma subtypes and broader chromosome 3–14 elevation for GBM-RTKI, while DINOv2 assigns its highest GBM-RTKI importance to chromosome 8.The prose connects chromosome 10 with GBM- IDH-wildtype and chromosomes 4 and 9 with PDGFRA amplification and CDKN2A deletion in GBM-RTKI.Source: B. XAI Analysis: CNV Model · p. 4
  • For OIDH, Grad-CAM emphasizes chromosome 1 and detects the 1p/19q co-deletion signature at chromosome 19; CNV-model explanations also identify subtype-specific chromosome patterns.Fig. 4 compares CAM and Grad-CAM heatmaps for CNV-only, DNAM-only, and combined modalities; blue-to-red encodes low-to-high importance, with masks thresholded at 0.85.Source: B. XAI Analysis: CNV Model · p. 5, Figure 4 · p. 5

C. XAI Analysis: DNAM Model

The DNAM-based explanations assign importance to chromosome 7 across several tumor classes, while Grad-CAM profiles are more diffuse for some subtypes.

C. XAI Analysis: DNAM Model

  • Chromosome 7 is the dominant ResNet50-CAM region for GBM-MES, GBM-RTKI, GBM-RTKII, and OIDH; AIDH and AIDH-HG show highest activation on chromosomes 7 and 8.The prose relates the chromosome 7 pattern to EGFR-locus methylation and regulatory elements.Source: C. XAI Analysis: DNAM Model · p. 5
  • Grad-CAM agrees with CAM for some subtypes, but its AIDH-HG and GBM outputs are more diffuse than the restricted CAM profiles.The authors suggest this may reflect reliance on more distributed DNAM signal when CNV data are absent.Source: C. XAI Analysis: DNAM Model · p. 5
  • For OIDH, Fig. 5 compares binned average CAM values across chromosome arms 1p versus 1q and 19p versus 19q, with CpG IDs on the x-axis and CAM values on the y-axis.The caption identifies Mann–Whitney significance analysis for chromosomes 1 and 19.Source: Figure 5 · p. 5

D. XAI Analysis: Multimodal Model

The combined model produces class-specific importance maps that include signals from both modalities, including regions not reflected in the single-modality models.

D. XAI Analysis: Multimodal Model

  • The multimodal masks identify chromosomes 1, 16, 17, and 18 for AIDH; 10 and 11 for AIDH-HG; 1 for OIDH; 10 and 11 for GBM-MES and GBM-RTKI; and 7, 10, and 11 for GBM-RTKII.Chromosomes 10 and 11 are related in the text to PTEN loss (10q) and CDK6/CCND1 amplification (11q).Source: D. XAI Analysis: Multimodal Model · pp. 5–6
  • Fig. 6 compares ResNet50 and DINOv2 on independent datasets, showing confidence distributions, accuracy with and without a confidence >0.9 threshold, and per-class accuracy.The supplied passage identifies the figure’s measures but gives no results for this comparison.Source: Figure 6 · p. 6
  • Chromosome 1 importance for OIDH persists in the multimodal model, while chromosome 18 shows elevated importance for AIDH despite not being reflected in either single-modality model.The authors state chromosome 18 is not known to play a role in AIDH and may merit downstream interrogation.Source: D. XAI Analysis: Multimodal Model · p. 6

E. Chromosome-Arm Resolution for OIDH

The chromosome-arm analysis of OIDH found elevated ResNet50-CAM activation in 1p and statistically significant activation patterns across chromosome 19, including a distinct pattern on 19q.

E. Chromosome-Arm Resolution for OIDH

  • At chromosome-arm resolution, Mann-Whitney testing of ResNet50-CAM importance values found elevated activation in proximal 1p bins, consistent with the telomeric 1p deletion characteristic of oligodendrogliomas.Chromosome 1 was divided into 1p and 1q probe ranges for the analysis.Source: E. Chromosome-Arm Resolution for OIDH · p. 6
  • The analysis used ResNet50-CAM importance values and the Mann-Whitney method to assess significance at chromosome-arm resolution.The chromosome-arm results are shown in Fig. 5.Source: E. Chromosome-Arm Resolution for OIDH · p. 6
  • On chromosome 19, elevated activation extended across both arms, with a distinct statistically significant activation pattern on 19q (p-value < 0.05).The analysis divided the chromosome 19 probes into 19p and 19q ranges.Source: E. Chromosome-Arm Resolution for OIDH · p. 6

F. Independent Validation

On independent validation, DINOv2 achieved higher overall and high-confidence accuracy than ResNet50, while the discussion reports more stable multimodal performance and limitations related to cohort composition and class imbalance.

F. Independent Validation

  • DINOv2 achieved 81.30% overall accuracy and 94.25% high-confidence accuracy (confidence > 0.9), versus ResNet50's 67.48% and 85.94%, while GAT achieved 62% overall accuracy.The independent validation dataset comprised 492 glioblastoma molecular subtype samples.Source: F. Independent Validation · pp. 6–7

V. DISCUSSION

  • Multimodal performance had lower cross-validation standard deviation than methylation-only performance, although CNV did not consistently improve peak accuracy.The discussion also reports that ResNet50 achieved the highest overall accuracy, while DINOv2 generalized better on independent test samples.Source: §V · p. 7
  • ResNet50-based explanations aligned with known glioma molecular hallmarks, while DINOv2's 16×16 patches yielded coarser importance maps and lower interpretability.The discussion identifies EGFR gain/amplification on chr7, PTEN loss on chr10q, and 1p/19q co-deletion as aligned hallmarks.Source: §V · p. 7
  • The TCGA dataset consisted predominantly of adult samples and may not fully capture the diversity of tumors encountered clinically.The work is presented as a proof-of-concept focused on adult-type diffuse gliomas.Source: §V · p. 7
  • Class imbalance across glioma subtypes may introduce bias into classification performance.Source: §V · p. 7

VI. CONCLUSION

MOXAI integrates methylation and copy number data from methylation arrays to classify adult-type diffuse gliomas and compare single-modality models. The conclusion reports greater classification stability with multimodal integration, stronger independent-test generalization by DINOv2, and more interpretable explanations from ResNet50.

VI. CONCLUSION

  • Multimodal CNV and methylation integration improved classification stability over single-modality models, while methylation alone was more discriminative than CNV alone.Source: §VI · p. 7
  • DINOv2 achieved 94.25% accuracy on independent test samples, compared with 85.94% for ResNet50.Source: §VI · p. 7
  • ResNet50-based CAM and Grad-CAM outputs were more interpretable than DINOv2's and aligned more closely with established molecular features of glioma subtypes.Source: §VI · p. 7
  • Future work will explore GIN, SAGE, and GCN graph-network variants and evaluate Integrated Gradients and SHAP.Source: §VI · p. 7
Loading 2609.32190v1…