| 33 |
2026 |
MOCDT: multi-cancer detection and tissue-of-origin classification via cfDNA multi-modal integration
MOCDT: multi-cancer detection and tissue-of-origin classification via cfDNA multi-modal integration
Abstract
Circulating tumor DNA (ctDNA) contains rich molecular information that can be leveraged for both cancer detection and tissue-of-origin (TOO) prediction. However, effectively integrating multiple cfDNA-derived modalities, including methylation, copy number variation (CNV), and fragment size ratio (FSR), remains challenging. To address this issue, the authors developed MOCDT (Multi-Omics Cancer Detection and Tissue-of-Origin), a deep learning framework that integrates multi-modal cfDNA features through a clinically relevant two-stage workflow consisting of cancer detection followed by tissue-of-origin classification.
Purpose
The study aimed to:
Develop a robust deep learning framework capable of integrating multiple cfDNA modalities, including methylation, CNV, and fragmentomics data.
Improve the accuracy of blood-based multi-cancer detection.
Accurately identify the tissue of origin for detected cancers.
Overcome limitations of existing multi-omics approaches, such as modality imbalance, insufficient exploitation of inter-patient relationships, and inadequate learning of tissue-specific biological features.
Results
Study Cohort
Total participants: 1,423 individuals
Included healthy controls and patients with eight cancer types:
Colorectal cancer
Gastric cancer
Liver cancer
Pancreatic cancer
Lung cancer
Breast cancer
Ovarian cancer
Prostate cancer
Cancer Detection Performance
MOCDT achieved strong performance in distinguishing cancer patients fr|om healthy individuals:
Metric
Performance
Specificity
95.74%
Sensitivity
96.22%
Accuracy
96.09%
The proposed framework outperformed several existing multi-omics integration methods, including MOGONET, MoGCN, and MO-GCAN.
Tissue-of-Origin Classification Performance
Metric
Performance
Top-1 Accuracy
75.20%
Top-2 Accuracy
86.79%
Top-3 Accuracy
91.06%
These results indicate that the correct cancer type was included among the top three predicted tissue origins in over 91% of cases.
Biological Interpretation
Latent-space analyses demonstrated that MOCDT successfully learned tissue-specific molecular patterns rather than relying solely on generic cancer-associated signals. This finding supports the biological interpretability of the model and its ability to distinguish among different cancer origins.
Conclusions
MOCDT is an effective multi-modal deep learning framework that integrates methylation, CNV, and fragmentomics features fr|om cfDNA for both cancer detection and tissue-of-origin classification. The model demonstrated:
High sensitivity and specificity for cancer detection.
Strong tissue-of-origin prediction performance.
Effective learning of tissue-specific biological characteristics.
A clinically applicable two-stage workflow suitable for multi-cancer early detection (MCED).
Overall, the study suggests that integrating multiple cfDNA-derived signals through advanced artificial intelligence approaches can significantly enhance the performance of liquid biopsy-based cancer screening and tissue-of-origin prediction, supporting the future development of non-invasive multi-cancer early detection strategies.
|
Bioinformatics |
リンクを見る
|
| 32 |
2026 |
Enhanced multicancer screening assay through whole-genome methylation sequencing-based multimodal cell-free DNA analysis
Abstract
This study developed a multimodal cfDNA-based multicancer detection assay combining whole-genome methylation sequencing with machine learning analysis.
The model integrated methylation and fragmentomic features and demonstrated high performance across eight cancer types, achieving 93.2% sensitivity and 95% specificity. It also showed strong early-stage detection capability, with over 92% sensitivity for stage I and II cancers.
The assay further achieved 85.7% accuracy in predicting tissue of origin, highlighting its potential for improving multicancer early detection and cancer screening.
Introduction
Cancer remains a leading cause of death worldwide, highlighting the urgent need for accurate and noninvasive early detection methods. Current screening approaches are often invasive, expensive and limited in detecting multiple cancer types, especially cancers without established screening guidelines.
Liquid biopsy using cell-free DNA (cfDNA) has emerged as a promising strategy for multicancer early detection (MCED). Recent advances in sequencing technologies have enabled analysis of methylation, copy number variation (CNV) and fragmentomic features fr|om circulating tumor DNA (ctDNA). However, existing MCED methods still face challenges in achieving high sensitivity for early-stage cancers due to the low abundance of ctDNA.
To overcome these limitations, this study developed a multimodal MCED framework integrating four cfDNA features: average methylation fraction (AMF), CNV, fragment size ratio (FSR) and fragment size distribution (FSD). Using an ensemble machine learning model, the study evaluated detection performance across eight major cancer types and assessed tissue-of-origin prediction accuracy. The findings demonstrate the potential of combining multiple cfDNA characteristics to improve noninvasive early cancer detection.
Materials and methods
This study analyzed plasma cfDNA samples fr|om patients with eight cancer types—colorectal, gastric, liver, pancreatic, lung, breast, ovarian and prostate cancer—as well as healthy controls. Blood samples were collected before treatment, and cfDNA was extracted fr|om plasma for analysis.
Whole-genome methylation sequencing was performed using the IMBdx AlphaLiquid screening platform, followed by extensive NGS preprocessing and quality control. The study evaluated four major cfDNA features: average methylation fraction (AMF), copy number variation (CNV), fragment size distribution (FSD) and fragment size ratio (FSR). Cancer-specific methylation markers and genomic alterations were identified using statistical filtering and machine learning-based optimization.
Single-feature models were first developed for each cfDNA characteristic using machine learning algorithms such as random forest, logistic regression and support vector machines. These models were then integrated into an ensemble framework combining methylation, genomic and fragmentomic signals along with demographic factors including age and sex. The final multimodal model was trained to detect cancer signals and predict tissue of origin while maintaining high specificity. Statistical analyses were performed using R and Python with rigorous validation procedures.
Results
The study analyzed 1,415 samples, including 1,034 cancer samples fr|om eight cancer types and 381 healthy controls. More than half of the cancer cases were stage I–II, enabling robust evaluation of early cancer detection performance.
Unsupervised clustering analyses using methylation, copy number variation (CNV), and fragmentomic features showed clear separation between healthy individuals and cancer patients, especially in advanced-stage disease. Distinct cancer-specific methylation signatures and fragmentomic patterns were identified, supporting the complementary value of integrating multiple cfDNA features.
Among single-feature models, average methylation fraction (AMF) showed the highest overall sensitivity (85.3%), particularly for early-stage cancers. CNV, fragment size ratio (FSR), and fragment size distribution (FSD) also contributed meaningful diagnostic information across different cancer types.
The multimodal ensemble model integrating AMF, CNV, FSR, and FSD achieved strong overall performance with 93.2% sensitivity and 95% specificity. Importantly, sensitivity remained high for early-stage cancers, reaching 92.3% for stage I and 92.2% for stage II disease. Sensitivity was especially high for colorectal, breast, and gastric cancers, while also demonstrating strong detection capability for difficult-to-screen cancers such as pancreatic, ovarian, and prostate cancer.
For tissue-of-origin (TOO) prediction, the model achieved 72.9% top-1 accuracy and 85.7% top-2 accuracy across the eight cancer types. The ensemble framework successfully leveraged complementary methylation, genomic, and fragmentomic signals to improve both cancer detection and tissue classification performance.
Discussion
This study demonstrated that a multimodal cfDNA analysis framework integrating methylation (AMF), copy number variation (CNV), and fragmentomic features (FSR and FSD) can significantly improve multicancer early detection (MCED). The ensemble model achieved high sensitivity (93.2%) and specificity (95%), outperforming several existing cfDNA-based cancer screening approaches.
The model showed particularly strong performance for early-stage cancers, highlighting the importance of combining complementary cfDNA features. AMF effectively detected early epigenetic alterations, while CNV and fragmentomic analyses contributed additional discriminatory power, especially in later-stage disease. The multimodal strategy also improved tissue-of-origin prediction accuracy.
Despite these promising results, challenges remain for cancers with low ctDNA abundance, such as ovarian and prostate cancer, and for certain stage III cancers affected by variable ctDNA shedding. The authors also acknowledged limitations related to single-cohort validation and emphasized the need for large-scale prospective external validation studies.
Overall, the findings support the potential clinical utility of multimodal cfDNA analysis as a scalable and noninvasive approach for improving early cancer detection and future cancer screening strategies.
Data availability
The raw data for this study were generated by IMBdx Inc. Data supporting the findings of this study are available fr|om the corresponding author upon reasonable request. Data access may be subject to institutional and ethical regulations.
|
Experimental & Molecular Medicine |
リンクを見る
|
| 31 |
2025 |
Enhanced Detection of Druggable Mutations in Non–Small Cell Lung Cancer Using Targeted Collection of Bronchial Washing Fluid Compared With Plasma and Tumor Tissue
Abstract
PurposeNext-generation sequencing (NGS) has become the gold standard for the molecular testing of patients with non–small cell lung cancer (NSCLC). This prospective study evaluated the performance of NGS using cell-free tumor DNA (ctDNA) extracted fr|om bronchial washing fluid (BWF) collected via a targeted washing technique to detect druggable mutations.
Materials and Methods All study participants simultaneously underwent NGS using three sample types: (1) BWF, (2) plasma, and (3) tumor tissue collected during bronchoscopy. The full patient set (FPS) included all enrolled patients, whereas the analysis intent group (AIG) included patients who underwent successful NGS across all specimen types (BWF, plasma, and tissue).
Results Sixty and 50 patients were included in the FPS and AIG groups, respectively. In FPS, the detection rate of druggable mutations in BWF using NGS was 65%, which was significantly higher than that of plasma (47%) and tissue samples (48%; P 5 .003 and P 5 .002, respectively). In the AIG, the concordance rate for detecting druggable mutations between BWF and tissue samples was 94%. In addition, the detection rate of co-occurring genetic alterations in BWF using NGS was significantly higher than that in plasma samples (92% v 64%, P 5 .001), whereas it was comparable with that in tissue samples (92% v 94%, P 5 1.000). No significant adverse events occurred during the BWF collection.
ConclusionsNGS using ctDNA fr|om BWF obtained through a targeted washing technique is a feasible and reliable method for genomic profiling of NSCLC, providing a promising approach for identifying druggable mutations.
|
JCO Precision Oncology |
リンクを見る
|
| 30 |
2025 |
An open-label, phase IB/II study of abemaciclib with paclitaxel for tumors with CDK4/6 pathway genomic alterations
Abstract
Background
Disruption of cyclin D-dependent kinases (CDKs), particularly CDK4/6, drives cancer cell proliferation via abnormal protein phosphorylation. This open-label, single-arm, phase Ib/II trial evaluated the efficacy of the CDK4/6 inhibitor, abemaciclib, combined with paclitaxel against CDK4/6-activated tumors.
Patients and methods
Patients with locally advanced or metastatic solid tumors with CDK4/6 pathway aberrations were included. Based on phase Ib, the recommended phase II doses were determined as abemaciclib 100 mg twice daily and paclitaxel 70 mg/m2 on days 1, 8, and 15, over 4-week-long cycles. The primary endpoint for phase II was the overall response rate (ORR). The secondary endpoints included the clinical benefit rate (CBR), progression-free survival (PFS), overall survival (OS), and safety. Tissue-based next-generation sequencing and exploratory circulating tumor DNA analyses were carried out.
Results
Between February 2021 and April 2022, 30 patients received abemaciclib/paclitaxel (median follow-up: 15.7 months), and 27 were included in the efficacy analysis. CDK4/6 amplification (50%) and CCND1/3 amplification (20%) were common activating mutations. The ORR was 7.4%, with two partial responses, and the CBR was 66.7% (18/27 patients). The median OS and PFS were 9.9 months [95% confidence interval (CI) 5.7-14.0 months] and 3.5 months (95% CI 2.6-4.3 months), respectively. Grade 3 adverse events (50%, 21 events) were mainly hematologic. Genetic analysis revealed a ‘poor genetic status’ subgroup characterized by mutations in key signaling pathways (RAS, Wnt, PI3K, and NOTCH) and/or CCNE amplification, correlating with poorer PFS.
Conclusion
Abemaciclib and paclitaxel showed moderate clinical benefits for CDK4/6-activated tumors. We identified a poor genetic group characterized by bypass signaling pathway activation and/or CCNE amplification, which negatively affected treatment response and survival. Future studies with homogeneous patient groups are required to validate these findings.
|
ESMO Open |
リンクを見る
|
| 29 |
2023 |
|
International Journal of Cancer |
リンクを見る
|
