Development of a urine-based metabolomics approach for multi-cancer screening and tumor origin prediction

Abstract

Background: Cancer remains a leading cause of mortality worldwide. A non-invasive screening solution was required for early diagnosis of cancer. Multi-cancer early detection (MCED) tests have been considered to address the challenge by simultaneously identifying multiple types of cancer within a single test using minimally invasive blood samples. However, a multi-cancer screening strategy utilizing urine-based metabolomics has not yet been developed.

Methods: We enrolled 911 cancer patients with 548 lung cancer (LC), 177 with gastric cancer (GC), and 186 with colorectal cancer (CRC), alongside 563 individuals with non-cancerous benign diseases and 229 healthy controls (HC) and investigated the metabolic profiles of urine samples. Participants were randomly allocated to discovery and validation cohorts. The discovery cohort was used for identifying multi-cancer and tissue-specific signatures to build the cancer screening and tumor origin prediction models, while the validation cohort was employed for assessing the performance of these models.

Results: We identified and annotated a total of 360 metabolites from the urine samples. Using the LASSO regression algorithm, 18 metabolites were characterized as urinary metabolic biomarkers and exhibited excellent discriminative performance between cancer patients and HC with AUC of 0.96 in the validation cohort. In comparison with the performance of traditional tumor markers CEA, the screening model performed higher sensitivity across the cancer stages, with a particularly increase in sensitivity among early-stage cancer patients. Moreover, the screening model also exhibited in high classification of cancers from non-cancerous group, comprising with HC and benign disease participants. Furthermore, two non-overlapping metabolic panels were selected to differentiate LC from Non-LC and GC from CRC with the AUC values of 0.87 and 0.83 in validation cohorts, respectively. Additionally, the model accurately predicted the origin of three lethal cancers: lung, gastric, and colorectal, with an overall accuracy of 0.75. The AUC values for LC, GC, and CRC were 0.88, 0.88, and 0.80, respectively.

Discussion: Our study demonstrates the potential of urine-based metabolomics for multi-cancer early detection. The approach offers non-invasive cancer screening, promising widespread implementation in population-based programs for early detection and improved outcomes. Further validation and expansion are needed for broader clinical applicability.

Keywords: machine learning; multi-cancer screening; pathway; tumor origin prediction; urinary metabolomics.

Figure 1

The workflow diagram of study design.

Figure 2

The performance of metabolic panel-based model for discriminating cancer patients from healthy controls. Kyoto Encyclopedia of Genes and Genomes (KEGG) metabolic pathways enriched by least absolute shrinkage and selection operator (LASSO) selected features (A). The receiver operating characteristic (ROC) curve for the diagnosis of cancer patients vs. healthy controls (B). Detection rates of metabolic panel-based model and carcinoembryonic antigen (CEA) at > 99% specificity in the validation cohort (C).

Figure 3

The performance of metabolic panel-based model for discriminating cancer patients from non-cancer group. The receiver operating characteristic (ROC) curve for the diagnosis of cancer patients (A). Detection rates of metabolic panel-based model at 95% specificity in the validation cohort (B). Heatmap of 18 least absolute shrinkage and selection operator (LASSO)-selected metabolites in non-cancer group and the cancer group at different stages (C).

Figure 4

Classification of cancers by metabolic panel-based model. The receiver operating characteristic (ROC) curve for the discrimination of lung cancer (LC) patients from non-lung cancer patients (A). Heatmap of the least absolute shrinkage and selection operator (LASSO)-selected metabolites between LC and non-LC patients (B). The ROC curve for the discrimination of gastric cancer patients (GC) from colorectal cancer patients (CRC) (C). Contribution of the selected metabolites to the discrimination model for GC vs. CRC (D).

Figure 5

The performance of the multi-cancer classification model. The receiver operating characteristic (ROC) curves evaluating the model in discriminating tumor origins in the validation cohort (A). Confusion matrix summarizing the cancer classification results in the validation cohort (B).