Tri‐modal liquid biopsy: Combinational analysis of circulating tumor cells, exosomes, and cell‐free DNA using machine learning algorithm

Dear Editor, Analysis of tumor biomarkers in circulation, commonly known as liquid biopsy, has been highlighted as an effective real-time monitoring technique for the surveillance of therapeutic responses and tumor progression.1–3 However, existing liquid biopsy assays that utilize a single tumor biomarker lack the sensitivity and specificity required to obtain clinically reliable information.4 In this study, we established a multimodal liquid biopsy (MMLB) system that integrates the expression profiles of the three different tumor biomarkers, circulating tumor cells (CTCs), exosomes, and cell-free DNA (cfDNA), using amachine learning algorithm (Figure 1). CTCs, exosomes, and cfDNA were isolated using alginate beads functionalized with anti-epithelial cell adhesion molecule antibodies (aEpCAM), anti-CD63 antibodies (aCD63), and polydopamine-silica (PDA-SiO2), respectively. Prior to clinical application, the capture capability of the bead-based systems was validated using in vitro samples (Figure S1). The aEpCAM-functionalized beads achieved ∼73.1% capture efficiency of EpCAM SW480 cells, with ∼99.4% leukocyte removal, ∼93.4% cell retrieval, and ∼87.0% cell viability (Figure 2A). Meanwhile, aCD63-functionalized beads achieved slightly lower recovery of exosome nucleic acids (∼38% less) and PDASiO2 beads achieved 1.27-fold more capture of cfDNA compared with commonly-used commercial kits (Figures 2B and 2C). Furthermore, all three assays demonstrated high selectivity toward their target biomarkers (Figures 2D–2F). The clinical applicability of the beads was investigated using samples obtained from 72 colorectal cancer patients, 14 patients with benign colorectal tumors, and 14 healthy individuals (Table S1). All three systems were capable of differentiating cancer patients from non-cancer controls, with an AUC-ROC of 0.826 (CTCs; p < 0.001), 0.763 (exosomes; p < 0.001), and 0.820 (cDNA; p < 0.001),


L E T T E R T O E D I T O R
Tri-modal liquid biopsy: Combinational analysis of circulating tumor cells, exosomes, and cell-free DNA using machine learning algorithm Dear Editor, Analysis of tumor biomarkers in circulation, commonly known as liquid biopsy, has been highlighted as an effective real-time monitoring technique for the surveillance of therapeutic responses and tumor progression. [1][2][3] However, existing liquid biopsy assays that utilize a single tumor biomarker lack the sensitivity and specificity required to obtain clinically reliable information. 4 In this study, we established a multimodal liquid biopsy (MMLB) system that integrates the expression profiles of the three different tumor biomarkers, circulating tumor cells (CTCs), exosomes, and cell-free DNA (cfDNA), using a machine learning algorithm ( Figure 1).
The clinical applicability of the beads was investigated using samples obtained from 72 colorectal cancer patients, 14 patients with benign colorectal tumors, and 14 healthy individuals (Table S1). All three systems were capable of differentiating cancer patients from non-cancer controls, with an AUC-ROC of 0.826 (CTCs; p < 0.001), 0.763 (exosomes; p < 0.001), and 0.820 (cDNA; p < 0.001), This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. However, the diagnostic capabilities of the bead-based systems were still insufficient to be applied in clinical practice due to the inherent variability of each measurement. CTCs were only found in 65.3% of cancer patients, exosomes were elevated in both benign and malignant tumor patients, and cfDNA levels of early-stage patients were similar to those of the non-cancer cohorts.
To improve the diagnostic accuracy, k-means clustering was utilized to deduce patterns among the three assays ( Figures 3A and 3B). The optimal number of clusters (k) was determined to be five based on the elbow method (denoted as A1-A5; Figure S6). Between the clusters, there was a significant difference in the expression profiles of the three biomarkers ( Figures 3C-3E). Interestingly, 92.9% (13/14) of healthy individuals and 85.7% (12/14) of patients with benign tumors belonged to cluster A1, which exhibited the lowest expression of all three biomarkers (Table  S2). Principle component analysis (PCA) was then used to simplify CTC-exosome-cfDNA expressions into a 2D plot ( Figure 3F). The x-axis of the plot (PCA-X) demonstrated moderate-to-strong correlations with all three biomarkers. The PCA-X was thus defined as MMLB Score which was equivalent to the linear combination of CTCs, exosomes, and cfDNA, with standardized coefficients of 0.600, 0.592, and 0.184, respectively ( Figure S7). The MMLB Score was significantly higher among cancer patients (0.40 ± 1.11) compared to non-cancer cohorts (−1.03 ± 0.33; p < 0.001), exhibiting a greater AUC-ROC (0.894; p < 0.001) than any of the single biomarkers ( Figures 3G-3K).
We repeated the clustering and PCA on samples from patients with malignant tumors (MMLB Cancer ) to determine if the score was predictive of pathological status and survival outcomes (Table S3) MMLB Cancer outperformed all three biomarkers individually, as well as serum antigens that are routinely tested in clinical practice ( Figures 4F and S9). MMLB Cancer also exhibited a moderate correlation with patients' LVI status, presence of nodal metastasis, and prevalence of distant metastasis (Figures S10 and S11). Furthermore, a Kaplan-Meier survival analysis demonstrated that patients with low MMLB Cancer (≤median) exhibited a 3.36-fold (p = 0.009) longer mean disease-free survival (DFS) than those with high MMLB Cancer (Table S4), whereas no significant differences were found from the single tumor biomarkers (Figures 4G-4J). Likewise, MMLB Cancer outperformed all three biomarkers for predicting the overall survival (OS) of patients ( Figures 4K-4N). As a result, the hazard ratio of MMLB Cancer was 1.370 (p = 0.006) and 1.623 (p = 0.015) for predicting DFS and OS, respectively, demonstrating superior prediction capabilities in comparison to any of the single biomarkers ( Figure 4O).
The MMLB analysis was further applied to detect KRAS mutations (MMLB KRAS ) by combining CTC counts, miR-100 expression in exosomes, and the fraction of cfDNA KRAS mutant alleles, which have all been reported to be overexpressed in patients with KRAS mutation. 5,6 Clustering analysis revealed that the group which had the lowest expression for all three biomarkers (K1) showed significantly lower tissue KRAS mutant allele fraction (∼5.5%) than the other two groups (∼16.9%) (Figures 4P,  4Q, and S12-S15). Furthermore, PCA demonstrated that MMLB KRAS exhibited a strong correlation with tissue KRAS mutation status and detected the mutation with high accuracy (Figures 4R-4U and S16).
Our MMLB analysis demonstrated that the utilization of machine learning algorithms has great advantages, as our system was shown to have greater diagnostic/prognostic values than simply adding or multiplying individual tumor biomarkers (Table S5). It should also be noted that the MMLB analysis is not only limited to our bead-based systems but is also applicable to other liquid biopsy assays. In future studies, improvements will be validated in a larger patient cohort with various tumor types using highly sensitive liquid biopsy assays that our group has developed previously. [7][8][9][10] In conclusion, we have demonstrated that a machine learning-based approach that integrates CTC, exosome, and cfDNA liquid biopsy measurements into a single-score MMLB system was more predictive than each marker alone. This approach may overcome the limitations of existing liquid biopsies with limited sensitivity and specificity. MMLB analysis demonstrated significantly improved accuracy in diagnosing malignancies, determining the pathological status of patients, predicting survival outcomes, and detecting gene mutations. These findings suggest that our approach markedly enhances liquid biopsy assays to ultimately achieve personalized medicine and improve patient outcomes.

AVA I L A B I L I T Y O F D ATA A N D M AT E R I A L S
All data generated from this study are included in this published article and supporting information. Raw data are available from the corresponding author on reasonable request.

E T H I C S A P P R O VA L A N D C O N S E N T T O PA R T I C I PAT E
Informed consent was obtained in accordance with the Declaration of Helsinki. This study was approved by the Ethics Committee of Eulji University (study numbers: EU 2017-44 and EU2018-68)