Letter to the Editor: An ultra-sensitive assay using cell-free DNA fragmentomics for multi-cancer early detection

Early detection can benefit cancer patients with more effective treatments and better prognosis, but existing early screening tests are limited, especially for multi-cancer detection. This study investigated the most prevalent and lethal cancer types, including primary liver cancer (PLC), colorectal adenocarcinoma (CRC), and lung adenocarcinoma (LUAD). Leveraging the emerging cell-free DNA (cfDNA) fragmentomics, we developed a robust machine learning model for multi-cancer early detection. 1,214 participants, including 381 PLC, 298 CRC, 292 LUAD patients, and 243 healthy volunteers, were enrolled. The majority of patients (N = 971) were at early stages (stage 0, N = 34; stage I, N = 799). The participants were randomly divided into a training cohort and a test cohort in a 1:1 ratio while maintaining the ratio for the major histology subtypes. An ensemble stacked machine learning approach was developed using multiple plasma cfDNA fragmentomic features. The model was trained solely in the training cohort and then evaluated in the test cohort. Our model showed an Area Under the Curve (AUC) of 0.983 for differentiating cancer patients from healthy individuals. At 95.0% specificity, the sensitivity of detecting all cancer reached 95.5%, while 100%, 94.6%, and 90.4% for PLC, CRC, and LUAD, individually. The cancer origin model demonstrated an overall 93.1% accuracy for predicting cancer origin in the test cohort (97.4%, 94.3%, and 85.6% for PLC, CRC, and LUAD, respectively). Our model sensitivity is consistently high for early-stage and small-size tumors. Furthermore, its detection and origin classification power remained superior when reducing sequencing depth to 1× (cancer detection: ≥ 91.5% sensitivity at 95.0% specificity; cancer origin: ≥ 91.6% accuracy). In conclusion, we have incorporated plasma cfDNA fragmentomics into the ensemble stacked model and established an ultrasensitive assay for multi-cancer early detection, shedding light on developing cancer early screening in clinical practice.


Main text
The global cancer burden is increasing rapidly, and nearly 19.3 million new cases and 10.0 million cancer deaths were estimated in 2020 [1]. Over 60% of newly diagnosed cases and 70% of cancer mortality can be attributed to 10 common cancer types [1]. Among them, liver cancer, colorectal cancer, and lung cancer rank the top three causes and account for over one-third of cancer deaths [1]. Although cancer identified early is more likely to have a favorable prognosis [2], only limited early screening programs have been made available for specific cancer types [3]. Furthermore, detection limits, radiation exposure, fear of pain, monetary cost, etc., of existing screening programs are also obstacles in their implementation [4][5][6]. Therefore, exploring accurate and affordable biomarkers is needed for promoting early detection.
As a new class of biomarkers for cancer detection, cellfree DNA (cfDNA) in circulation is released from apoptosis and necrosis, and contains molecular signatures of its origin [7,8]. For instance, tumor somatic mutations can serve as a classifier to distinguish circulating tumor DNA (ctDNA) shed from tumor cells and nontumorous cfDNA [9]. Epigenetic modifications such as DNA methylation and fragmentomic signatures such as fragmentation patterns and end motifs have also been utilized for identifying cancer [10][11][12][13][14]. However, assays based on single cfDNA features often yield inadequate detection ability, especially for stage I cases of prevalent cancer types [12,[14][15][16]. As identification at stage I often provides a better chance for the cure than later stages, developing more robust methods is critical to promote cancer early detection.
More recently, multi-dimensional predictive models that combine multiple fragmentomic and genomic features and even incorporate clinical information have improved their detection power for specific cancer types [12,17]. Particularly, Ma et al. have leveraged the ensemble stacked strategy to integrate multiple fragmenomic features with machine learning algorithms and successfully built an ultrasensitive model for detecting stage 0/I colorectal adenocarcinoma [18]. Given the potential of the ensemble stacking strategy, we attempted to develop a multi-dimensional model using cfDNA fragmentomics from WGS data for multi-cancer detection and origin localization. Owing to their high prevalence and substantial impact, we built the model targeting liver, colorectal, and lung cancers in a cohort of the Chinese population. Our model demonstrated ultrasensitivity for cancer detection and accurately differentiated cancer origins, ideal for promoting cancer screening programs.

Participant characteristics and disposition
We included 1,214 participants with previously untreated diseases: 381 primary liver cancer (PLC), 298 colorectal cancer (CRC), 292 lung adenocarcinoma (LUAD), and 243 healthy volunteers without cancer (Fig. 1A). This study was approved by the Ethnic Committees and in accordance with the ethical standards as laid down in the 1964 Declaration of Helsinki and its later amendments. Written informed consents were provided by all participants. Details about enrollment information are in Supplementary Materials and Methods. The participants were subject to WGS and fragmentomic feature extraction and randomly split into the training and test datasets and 90.4% for PLC, CRC, and LUAD, individually. The cancer origin model demonstrated an overall 93.1% accuracy for predicting cancer origin in the test cohort (97.4%, 94.3%, and 85.6% for PLC, CRC, and LUAD, respectively). Our model sensitivity is consistently high for early-stage and small-size tumors. Furthermore, its detection and origin classification power remained superior when reducing sequencing depth to 1× (cancer detection: ≥ 91.5% sensitivity at 95.0% specificity; cancer origin: ≥ 91.6% accuracy). In conclusion, we have incorporated plasma cfDNA fragmentomics into the ensemble stacked model and established an ultrasensitive assay for multi-cancer early detection, shedding light on developing cancer early screening in clinical practice. in a 1:1 ratio. We took the whole training dataset to build the first-level cancer detection model and then the cancer samples in the training dataset to train the second-level cancer origin model. The workflow of model construction is described in Fig. 1B (Table  S2). We observed an upward trend from the early to later stages for the distribution of cancer scores in allcancer, PLC, and CRC classes (Fig. S1). A propensity score matching analysis balanced the age and sex factors between cancer and non-cancer groups in the test dataset. The resultant subset consisting of 113 PLC, 73 CRC, 85 LUAD, and 85 age and gender-matched healthy controls remained high performance in distinguishing cancer patients from non-cancer controls (AUC: 0.988, 95% CI: 0.980-0.996, Fig. S2A). We also performed 10-fold cross-validation during training to evaluate model overfitting. The 10-fold cross-validation AUCs for all-cancer and individual cancer types were equally high compared to the independent test dataset (Fig. S2B), reassuring that overfitting was not a major concern.
Our model exhibited ultrasensitivity in detecting cancers at various stages (Fig. 2D). The sensitivity is above 90% for stages 0 and I, and elevated to nearly 100% for stages II and III. Furthermore, we used patient demographics and clinical characteristics to categorize disease subgroups for evaluation (Table S3 and Figs. S3-S5). The model's detection sensitivity was consistently high even in the challenging categories, such as MIA and <1 cm tumors of LUAD. We assessed the model's robustness by gradually down-sampling the coverage to 1× (Fig. 2E  Table S4). Despite a slight dip, the model remained stable with over 91.5% sensitivity for all-cancer. Even for the least detectable class of LUAD, the sensitivity at 1× is still above 87%.
Furthermore, the cancer detection model was assessed in a preliminary at-risk patient cohort and showed an overall specificity of 92.4% (Table S5, details in Supplementary Results).

Locating cancer at its origin by the cancer origin model
All test dataset patients correctly identified as "Cancer" by the cancer detection model were subsequently analyzed in the cancer origin model. The model correctly identified the cancer origin for 431 patients (accuracy 0.931, 95% CI: 0.900-0.950) for the three cancer types (Fig. 2F and Table  S6). The sensitivities for individual cancer types were 97.4% (95% CI: 94.0-99.1%), 94.3% (95% CI: 89.1-97.5%), and 85.6% (95% CI: 78.4-91.1%) for PLC, CRC, and LUAD, respectively. We plotted the cancer origin scores of each type for all patients (Fig. 2G). Generally, the top scores matched the true cancer types. Such consistency is the most compelling for the PLC patients, followed by the CRC patients, while the LUAD group has more erroneous CRC predictions ( Fig. 2F and G). We further inspected the origin scores of the misinterpreted patients (Fig. S6). The score differences between the true origin and the misinterpreted class were minimal (≤ 0.05) for potential improvement. The cancer origin model is robust with lower coverage WGS data ( Fig. 2H and Table S7). The accuracies for PLC, CRC, and LUAD at 1× coverage are 97.7%, 92.4%, and 90.6%, respectively, whereas the predictions of each patient at different sequencing coverages were listed in Fig. 2I, H. Our study has several limitations. First, we performed the proof-of-concept study using liver cancer, colorectal cancer, and lung cancer for their high prevalence. Targeting a broader population and more cancer types, including the less prevalent ones, would be necessary to develop the assay and eliminate cancer treatment inequity. Second, we are expanding our current cohort size to enable independent validation and improve the estimation accuracy of relatively small-size subgroups (e.g., cHCC-ICC, MIA, stage IB LUAD).

Conclusions
By integrating multiple fragmentomic features from cfDNA WGS data, our ensemble stacked model exhibited superior detection and localization power for the prevalent cancer types of PLC, CRC, and LUAD even at stages 0 and I. The robustness of our model is consistently high using as low as 1× sequencing coverage depth, suitable for developing accurate and affordable early detection assays for clinical practice.