MOCDT: multi-cancer detection and tissue-of-origin classification via cfDNA multi-modal integration

Tumor-derived circulating tumor DNA (ctDNA) fragments present in blood provide rich molecular signals for identifying cancer and mapping its tissue of origin. However, leveraging these heterogeneous signals requires robust computational integration methods. Existing multi-modal approaches often fail to capture both inter-modality structure and inter-patient relationships, limiting their utility for robust cancer detection (CD) and fine-grained tissue-of-origin (TOO) classification.

We propose MOCDT, a cell-free DNA (cfDNA) multi-omics framework that follows a clinically aligned two-stage pipeline: high-specificity CD followed by conditional TOO classification. MOCDT combines (i) a supervised multi-modal autoencoder incorporating adversarial modality alignment and supervised contrastive geometry shaping, with (ii) a latent space patient similarity network and (iii) a residual Graph Convolutional Network for relational learning. Applied to a cfDNA cohort including healthy controls and eight cancer types, MOCDT achieved 95.74% specificity and 96.22% sensitivity for CD at a high-specificity operating point, and 75.2% Top1 and 91.06% Top3 accuracy for TOO classification. Latent attribution analysis showed that the model learns tissue-dependent latent features rather than relying on a single universal biomarker axis. Together, these results demonstrate that MOCDT enables accurate and interpretable cfDNA-based multi-omics integration, supporting clinically relevant liquid biopsy applications.

Code and Dataset are available at https://github.com/Ewha-AI/MOCDT.