Explainable attention-based multi-omics fusion with protein language models for CML-versus-control classification and biomarker discovery

Abstract

Chronic Myeloid Leukemia (CML) is a well-defined hematological malignancy driven principally by the BCR::ABL1 fusion oncogene and aberrant tyrosine kinase signaling. Computational methods to predict leukemia suffer from the limitation of using single-modality data or black-box models, which cannot adequately incorporate complementary molecular evidence and deliver interpretable biomarker support. In this study, we introduce an explainable multi-omics fusion approach that combines protein language model embeddings, gene-expression data, mutation-level data, and pathway-informed representations to classify CML-positive and control samples and identify candidate biomarkers. The pre-trained protein language models encode context from the sequence, and dense encoders represent modalities based on transcriptomic, mutational, and pathway information. They are fused adaptively for classification using an attention module, and biomarkers are ranked by SHAP, Integrated Gradients, and attention attribution to support biological interpretation. The proposed framework achieved an accuracy of (98.14%), precision (98.0%), recall (98.4%), F1-score (98.14%), ROC–AUC (0.991), and PR–AUC (0.987), outperforming classical machine learning, deep learning, single-modality, and conventional fusion baselines. Reliability analysis yielded a Brier score of 0.031, and an Expected Calibration Error of 0.024, and stable performance on five-fold cross-validation (97.9 ± 0.4% accuracy and 0.988 ± 0.003 ROC–AUC). Biologically relevant biomarkers identified by explainability analysis included the BCR::ABL1 fusion gene, ABL1 kinase-domain mutations, BCL2, HSP90, RUNX1, ASXL1, PARP1 and RB1, and key pathways were identified: JAK–STAT, PI3K–AKT, RAS–MAPK, apoptosis and DNA repair. The findings indicate that the proposed model achieved improved predictive performance under the evaluated experimental conditions and provided interpretable identification of candidate biomarkers in CML and control samples.