To the editor

Recently, researchers have focused on utilizing plasma cell-free DNA (cfDNA), including cfDNA fragmentomic profiles, to develop noninvasive approaches for detecting solid malignancies such as colorectal adenocarcinoma (CRC) [1,2,3,4,5,6]. But the limited sensitivities of these current detection methods, by the use of either single molecular feature or single algorithm, reduce their potential utilization in clinical practice, while ensembled stacked machine learning approach can improve robustness and accuracy [7, 8]. Herein, we constructed a multi-dimensional ensembled stacked machine learning approach, employing five different base models on five optimized fragmentation features, to provide an ultrasensitive and cost-effective model for detecting early-stage CRC and advanced adenoma (advCRA).

In this study, 149 early-stage colorectal adenocarcinoma (CRC) patients, 46 advCRA patients and 115 healthy volunteers were recruited in the training cohort from a single center, which was used to train the machine learning models (Figs. 1, 2A). To eliminate the potential impact on the predictive power by the different coverages and maximize affordability, WGS data were down-sampled to 4X coverage, unless otherwise noted. The test cohort (N = 311), which consisted of 149 early-stage CRC, 46 advCRA patients and 116 healthy controls, was used to evaluate model performances. ROC curves were constructed using five individual features including Fragment Size Ratio (FSR), Fragment Size Distribution (FSD), EnD Motif (EDM), BreakPoint Motif (BPM) and Copy Number Variation (CNV), as well as the DELFI fragment size pattern [1] and the 4-bp end-motif pattern by Jiang et al. [2], to demonstrate the advantage of using a multi-dimensional ensembled stacked machine learning model approach, as well as adapting existing fragmentation features [7]. Detailed methodology is described in supplementary methods section (Additional file 1).

Fig. 1
figure 1

Schematic illustration of study design. Plasma samples were collected from patients with advanced colorectal adenoma (advCRA) or early-stage (stage 0/I) adenocarcinoma (CRC), as well as healthy controls. The cfDNA was then extracted from the participant’s plasma sample and subject to whole-genome sequencing. Five different feature types, including Fragment Size Ratio (FSR), Fragment Size Distribution (FSD), EnD Motif (EDM), BreakPoint Motif (BPM) and Copy Number Variation (CNV), were calculated using mapped sequencing reads. For each feature type, a base model was constructed based on the ensemble learning of five algorithm, GLM, GBM, random forest, deep learning and Xgboost. The base model predictions were then ensembled into a large matrix, which was subsequently used by a GLM algorithm to train the final ensemble stack model

Full size image
Fig. 2
figure 2

Evaluation of ensemble stacked machine learning model. A Graphical representation of datasets composition. The training cohort (N = 310) included 149 early-stage CRC patients, 46 advCRA patients and 115 healthy controls and was used to train the stacked ensemble model. The test cohort (N = 311), which included 149 early-stage CRC patients, 46 advCRA patients and 115 healthy controls, was independently used to evaluate model performances. B ROC curves evaluating the overall performance of the predictive model, which was constructed using 4 X coverage WGS data, in distinguishing advCRA/early-stage CRC patients from healthy controls in the test cohort. C Table evaluating model performances in the test dataset. D Boxplots illustrating cancer score distribution in the healthy, advCRA and early-stage CRC groups in the test cohort based on the 4 X overage model. The 95% specificity cutoff for cancer score was 0.62 as shown by the dotted line

Full size image

The ensembled stacked model had a higher AUC (0.988) than base models using any individual feature (AUC range 0.881–0.981), validating the multi-dimensional ensembled stacked approach (Additional file 1: Fig. S1). A similar pattern was observed as the ensembled stacked model had the highest sensitivity for detecting advCRA/early-stage CRC (97.4%, 95% CI 94.1–99.2%) compared to all base models (sensitivity range 57.4–89.2%) at 94.8% specificity (95% CI 89.1–98.1%) (95% CI 89.1–98.1%) (Additional file 1: Fig. S1, Table S1). Additionally, our adaptation to the existing fragmentation features was justified by showing better performances than the original features: the adapted 6-bp EDM feature showed higher AUC (0.981, 95% CI 0.969–0.993) than the original 4-bp end-motif feature (0.969, 95% CI 0.953–0.985), while models using FSR or FSD both had higher AUC (0.881, 95% CI 0.843–0.919; 0.892, 95% CI 0.855–0.930) than the original DELFI fragment pattern (Additional file 1: Fig. S1).

The stacked model showed better AUC while differentiating early-stage CRC (0.990, 95% CI 0.981–0.998) than advCRA (0.983, 95% CI 0.968–0.999) (Fig. 2B). Similarly, the model showed excellent sensitivities for detecting both advCRA (95.7%, 95% CI 85.2–99.5%) and early-stage CRC (98.0%, 95% CI 94.2–99.6%) at the 94.8% specificity (95% CI 89.1–98.1%) (Fig. 2D). The advCRAs more closely resembled early-stage CRCs than healthy controls (Fig. 2C), which was further validated by two additional models (Additional file 1: Fig. S2A, S2B). The current gold standard colonoscopy can be used to histopathologically distinguish advCRA and early-stage CRC following our model’s predictions.

We then constructed an ensembled stacked model using the raw depth NGS data (4.7–24.04X, median 9.75X), still showing great performances an identical AUC of 0.988 (95% CI 0.979–0.997) (Additional file 1: Fig. S3, Table S2). A limit of detection analysis was performed by further down-sampling the 4X coverage WGS data to 3X, 2X, 1X and 0.5X. The down-sampled data was then used to evaluate the 4X model. The AUCs showed a gradual decrease during the down-sampling process (0.988, 0.987, 0.985, 0.982 and 0.977 for 4X, 3X, 2X, 1X and 0.5X data, respectively) (Additional file 1: Fig. S4A).

In summary, our multi-dimensional ensembled stacked model, which uses plasma cfDNA WGS data, showed great potential for accurate noninvasive colorectal cancer screening prior to the current gold standard colonoscopy in clinical practice by demonstrating an unparalleled high sensitivity in detecting early-stage CRC as well as advCRA. However, this study was limited by several factors, namely the relatively small cohort size. The small number of healthy controls within the test cohort can impact the model performance, likely resulting in an underestimation of sensitivity. A multicenter, large-scale prospective study is needed to validate the clinical value of our methods further.