Early detection of multiple cancer types remains a significant challenge in precision oncology because conventional diagnostic approaches frequently depend on isolated molecular biomarkers, fixed classification thresholds, and limited representations of tumour heterogeneity. Although circulating tumour DNA (ctDNA) methylation signatures, somatic mutations, and circulating protein biomarkers provide complementary diagnostic information, differences in molecular abundance, analytical sensitivity, assay reliability, and biological specificity complicate their effective integration. This study proposes OncoFusion, a novel reliability-aware, attention-guided multimodal learning algorithm designed to integrate ctDNA methylation profiles, somatic mutation features, and protein biomarker concentrations for multi-cancer early detection and cancer signal-of-origin classification. The proposed architecture combines modality-specific feature encoders, a cross-modal attention transformer, uncertainty-weighted biomarker fusion, and a multitask classification framework to distinguish cancer-positive from cancer-negative samples while identifying probable cancer origins. A reliability-adaptive fusion mechanism dynamically adjusts the contribution of each biomarker modality according to sequencing depth, molecular coverage, missingness, assay quality, and predictive uncertainty. To evaluate its effectiveness, OncoFusion is designed for comparison with conventional single-biomarker threshold classifiers, logistic regression, random forest, support vector machines, XGBoost, LightGBM, and conventional stacking ensemble models. The experimental framework incorporates stratified nested cross-validation, independent external validation, bootstrap confidence intervals, calibration analysis, and ablation experiments to establish the contribution of individual biomarkers and architectural components. Diagnostic performance will be assessed using the area under the receiver operating characteristic curve (AUROC), area under the precision–recall curve (AUPRC), sensitivity at 95% and 99% specificity, F1-score, balanced accuracy, calibration error, and cancer signal-of-origin accuracy. Comparative results will be presented through receiver operating characteristic curves, precision–recall curves, confusion matrices, calibration plots, and biomarker contribution analyses, with particular emphasis on early-stage disease and low-abundance ctDNA signals. The central hypothesis is that reliability-aware cross-modal integration can improve diagnostic discrimination, maintain high specificity, and enhance robustness under incomplete or noisy biomarker measurements compared with single-modality and conventional ensemble approaches. The proposed OncoFusion framework provides a technically testable approach to multi-cancer early detection, with potential applications in precision oncology screening and laboratory-supported diagnostic decision-making, subject to independent clinical validation.