Development of a Serum Metabolome-Based Test for Early-Stage Detection of Multiple Cancers

Abstract

Background: Detection of cancer at the early stage currently offers the only viable strategy for reducing disease-related morbidity and mortality. Various approaches for multi-cancer early detection are being explored, which largely rely on capturing signals from circulating analytes shed by tumors into the blood. The fact that biomarker concentrations are limiting in the early stages of cancer, however, compromises the accuracy of these tests. We, therefore, adopted an alternate approach that involved interrogation of the serum metabolome with machine learning-based data analytics. Here, we monitored for modulations in metabolite patterns that correlated with the presence or absence of cancer. Results obtained confirmed the efficacy of this approach by demonstrating that it could detect a total of 15 cancers in women with an average accuracy of about 99%.

Aims: To further increase the scope of our test, we conducted an investigator-initiated clinical trial involving a total of 6445 study participants, which included both cancer patients and non-cancer volunteers. Our goal here was to maximize the number of cancers that could be detected, while also covering cancers in both females and males.

Methods and results: Metabolites extracted from individual serum samples were profiled by ultra-performance liquid chromatography coupled to a high-resolution mass spectrometer using an untargeted protocol. After processing, the data were analyzed by our cancer detection machine-learning algorithm to differentiate cancer from non-cancer samples. Results revealed that our test platform could indeed detect a total of 30 cancers, covering both females and males, with an average accuracy of ~98%. Importantly, the high detection accuracy remained invariant across all four stages of the cancers.

Conclusion: Thus, our approach of integrating untargeted metabolomics with machine learning-powered data analytics offers a powerful strategy for early-stage multi-cancer detection with high accuracy.

Trial registration: Registration No: CTRI/2023/03/050316.

Keywords: artificial intelligence; cancer; metabolomics.

FIGURE 1

Pipeline for data processing and development of the CDAI algorithm. Figure illustrates the workflow for development of the CDAI model, which also includes the data pre‐processing steps employed. A detailed explanation of the individual steps is provided in the Section 2.

FIGURE 2

k‐fold cross validation of the CDAI algorithm. To cross‐validate the model we used the k (k = 20) fold cross validation where the train test dataset was split into 20 parts. Figure shows the results for test sensitivities for each fold along with their 95% CI. The blue line shows the average sensitivity of the CDAI model.

FIGURE 3

Differentiating cancer samples from normal controls using the CDAI algorithm. (A) This panel shows the results obtained for normal samples, and the samples from each of the cancer types tested in the validation set, using the CDAI algorithm. A scatter plot depicting the distribution of y‐scores for individual samples either in the normal control group, or in each of the 30 cancer types tested, is presented here. Whereas the normal control group is denoted as Noc, individual cancer types are abbreviated as Br, breast cancer; Endo, endometrial cancer; Cv, cervical cancer; Ov, ovarian cancer; Lung, lung cancer; AML, acute myeloid leukemia; Thy, thyroid cancer; Mel, melanoma, Colo, colorectal cancer; Kdy, kidney cancer; NHL, Non‐Hodgkin's lymphoma; Panc, pancreatic cancer; Liv, liver and bile duct cancer; Gst, gastric cancer, H&N, head and neck cancer; Osp, oesophageal cancer; Bld, bladder cancer; Brn, brain and CNS cancer; MMY, multiple myeloma; GB, gall bladder cancer; Sarc, sarcoma; Prost, prostate cancer; Tstr, testicular cancer; Vulval; vulval cancer; Anal, anal cancer; Vgnl, vaginal cancer; Penile, penile cancer; Unk. Pr, cancer of unknown primary origin; GCT, germ cell tumors; SCC, squamous cell carcinoma. Panel (B) gives the resultant confusion matrix obtained from this data, after taking a y‐score of 0 as the threshold score for distinguishing cancer‐positive samples from normal controls. Values obtained for sensitivity and specificity are also given. The ROC‐AUC probability curve plot, as an indicator of CDAI performance, is shown in panel (C). The area under the curve (AUC) obtained was 0.99.

FIGURE 4

High fidelity performance of the CDAI is maintained across cancer types, subject age groups, and cancer stages. Panel (A) gives the sensitivity of cancer detection achieved by the CDAI for the validation subset of samples in each of the 30 cancer types tested. While the abbreviation for the cancers used here is the same as than in Figure 3A, the number of cancer‐positive samples identified, against the total number of samples tested, for each of the cancer types is included in parenthesis along the X‐axis. Panel (B) gives the CDAI cancer‐detection sensitivity obtained as a function of the age‐group distribution of the cancer patients while panel (C) shows a plot of detection sensitivity versus stage of the cancer. In both the latter panels the number of samples identified as cancer positive by the CDAI, versus the total number of validation set‐samples tested in each sub‐group is given in parenthesis along the X‐axis. For all panels, bars indicate 95% CI.

FIGURE 5

Stage‐specific CDAI test performance for representative cancers. Stage‐specific sensitivities obtained using the CDAI on the validation subset of 12 representative cancers are shown here. The number of samples identified as cancer positive, against the total number of samples tested at each stage of each of the cancers is also included in parenthesis and the bars indicate 95% CI.