Tassnem Haitham Ajaj

Background:

Cardiomyopathies (CMPs) are a heterogeneous group of inherited myocardial disorders characterized by structural and functional abnormalities of the heart muscle. The three major clinical subtypes—Arrhythmogenic Right Ventricular Cardiomyopathy (ARVC), Dilated Cardiomyopathy (DCM), and Hypertrophic Cardiomyopathy (HCM)—exhibit substantial genetic and phenotypic overlap, complicating variant interpretation and subtype assignment. Although large-scale databases such as ClinVar and dbNSFP provide extensive variant annotations, current prediction tools often focus on binary pathogenicity classification rather than disease-specific subtype discrimination. Given the overlapping molecular mechanisms involving sarcomeric, desmosomal, nuclear envelope, and RNA-binding genes, there is a critical need for interpretable, disease-specific machine learning frameworks capable of distinguishing cardiomyopathy subtypes. This study aimed to develop an explainable multi-class classification model integrating evolutionary conservation, amino acid position, population allele frequencies, and functional prediction scores to accurately classify cardiomyopathy-associated variants into ARVC, DCM, or HCM.

Results:

A curated dataset of 4,510 clinically significant variants across 18 cardiomyopathy-associated genes was constructed from ClinVar (2024 release) and functionally annotated using dbNSFP v4.9a. After preprocessing, filtering, and feature engineering, a multi-class XGBoost classifier was trained and evaluated. The baseline model achieved a test accuracy of 83.8%. Following hyperparameter optimization using Grid Search with five-fold cross-validation, performance improved to a test accuracy of 84.6%, cross-validation accuracy of 85.1%, and a macro F1-score of 0.81. Class-specific performance demonstrated highest predictive accuracy for DCM (precision = 0.91, recall = 0.94), while ARVC and HCM showed comparatively lower recall, partly due to class imbalance and gene-level phenotypic overlap. Misclassification analysis revealed systematic confusion primarily between DCM and ARVC variants, reflecting known biological overlap between desmosomal and sarcomeric pathways. Feature-level comparison showed that misclassified variants were located significantly closer to the N-terminus of proteins and exhibited distinct in silico prediction profiles. SHAP interpretability analysis identified amino acid position (aapos) as the most influential global feature, followed by evolutionary conservation (phyloP) and functional prediction scores (SIFT and PolyPhen-2). Gene-level SHAP patterns aligned with established pathogenic mechanisms, particularly for TTN, RBM20, and BAG3 variants. High-confidence predictions demonstrated strong disease specificity, and 97.2% alignment was observed between model predictions and curated annotations in the test set. Disagreement analysis identified a small subset of ultra-rare variants with damaging in silico profiles that may warrant further functional investigation.

Conclusion:

This study presents an explainable XGBoost-based multi-class framework for cardiomyopathy subtype classification that integrates biologically meaningful features while maintaining model transparency. The model achieved robust performance in distinguishing ARVC, DCM, and HCM variants and demonstrated biologically plausible decision patterns through SHAP analysis. Although challenges remain—particularly regarding class imbalance, transcript-specific effects, and pleiotropic gene behavior—the framework provides a strong foundation for disease-specific variant interpretation. By combining predictive accuracy with interpretability, this approach supports translational cardiogenomics and represents a step toward precision medicine applications in inherited cardiomyopathies. Future work incorporating isoform-aware annotations, functional genomics data, ancestry-adjusted allele frequencies, and external validation cohorts is expected to further enhance predictive performance and clinical utility.