MOCDT: multi-cancer detection and tissue-of-origin classification via cfDNA multi-modal integration.
Summary
This study introduces MOCDT, an innovative bioinformatics framework for multi-cancer detection and tissue-of-origin classification using cell-free DNA (cfDNA). MOCDT employs a two-stage pipeline: high-specificity cancer detection followed by conditional tissue-of-origin classification. The model integrates a supervised multi-modal autoencoder with a patient similarity network and a Graph Convolutional Network for relational learning. Applied to a cohort of healthy controls and eight cancer types, MOCDT achieved 95.74% specificity and 96.22% sensitivity for cancer detection, alongside 75.2% Top1 and 91.06% Top3 accuracy for tissue-of-origin classification.
Analysis
🔴 CLINIQUE : This work presents a significant advancement for early cancer diagnosis and characterization via liquid biopsy. The high specificity (95.74%) and sensitivity (96.22%) for cancer detection, combined with good accuracy for tissue-of-origin, suggest potential for clinical applications. Such a tool could justify early-phase clinical trials to validate its utility in mass screening or diagnosing cancers of unknown primary. Clinical impact could be seen in the medium term (3-5 years) by enabling earlier detection and faster therapeutic guidance. 🟢 BIOMOL : The MOCDT method leverages the rich molecular signals from circulating tumor DNA (ctDNA) fragments present in cfDNA. Multi-omic integration of cfDNA is a key translational approach, allowing the capture of heterogeneous information. The finding that the model learns tissue-dependent latent features, rather than relying on a single universal biomarker axis, is crucial for understanding underlying mechanisms and diagnostic robustness. The use of liquid biopsy (cfDNA) as a sample type is a major advantage due to its non-invasive nature. 🔵 BIOINFO : MOCDT represents an advanced application of machine learning for multi-modal data integration. The architecture combines a supervised multi-modal autoencoder with adversarial modality alignment and supervised contrastive geometry shaping, a latent space patient similarity network, and a residual Graph Convolutional Network. This approach is novel in its ability to capture both inter-modality structure and inter-patient relationships. The open-source availability of code and dataset (GitHub) ensures reproducibility and facilitates community evaluation. If clinically deployed, MOCDT could automate and improve the accuracy of variant review and patient prioritization for multidisciplinary tumor boards in oncology.