Noninvasive detection of any-stage cancer using free glycosaminoglycans

Abstract

Cancer mortality is exacerbated by late-stage diagnosis. Liquid biopsies based on genomic biomarkers can noninvasively diagnose cancers. However, validation studies have reported ~10% sensitivity to detect stage I cancer in a screening population and specific types, such as brain or genitourinary tumors, remain undetectable. We investigated urine and plasma free glycosaminoglycan profiles (GAGomes) as tumor metabolism biomarkers for multi-cancer early detection (MCED) of 14 cancer types using 2,064 samples from 1,260 cancer or healthy subjects. We observed widespread cancer-specific changes in biofluidic GAGomes recapitulated in an in vivo cancer progression model. We developed three machine learning models based on urine (N_urine = 220 cancer vs. 360 healthy) and plasma (N_plasma = 517 vs. 425) GAGomes that can detect any cancer with an area under the receiver operating characteristic curve of 0.83-0.93 with up to 62% sensitivity to stage I disease at 95% specificity. Undetected patients had a 39 to 50% lower risk of death. GAGomes predicted the putative cancer location with 89% accuracy. In a validation study on a screening-like population requiring ≥ 99% specificity, combined GAGomes predicted any cancer type with poor prognosis within 18 months with 43% sensitivity (21% in stage I; N = 121 and 49 cases). Overall, GAGomes appeared to be powerful MCED metabolic biomarkers, potentially doubling the number of stage I cancers detectable using genomic biomarkers.

Keywords: cancer biomarkers; liquid biopsy; metabolomics; multi-cancer early detection; prognosis.

Fig. 1.

Development of plasma, urine, and combined free GAGome MCED scores. (A) Development study overview and summary of free GAGome analysis (N_tot = 979, 553 cancers vs. 426 healthy; N_plasma = 517 cancers vs. 425 healthy; N_urine = 220 cancers vs. 340 healthy; N_combined = 184 cancers vs. 339 healthy). (B and C) Estimated means of plasma and urine GAGome features conditional on the cancer type (median N_plasma per cancer type = 30, range: 14–83; median N_urine = 50 per cancer type, range: 17–56). Dots signify credible deviations from healthy subjects (i.e., from baseline physiological levels defined by ROPE criteria [see Methods]). The vertical axis denotes independent and dependent GAGome features (measured in μg mL^-1 [except for charge, which is a.u.] or %w/w, respectively). (D–F) Free GAGome MCED scores across different stage/grade groups for plasma (N_H= 425, N_S1/LG= 178, N_S2= 54, N_S3= 57, and N_S4/HG= 217), urine (N_H= 340, N_S1/LG= 53, N_S2= 18, N_S3= 21, and N_S4/HG= 126), and for combined (N_H= 339, N_S1/LG= 44, N_S2= 16, N_S3= 19, and N_S4/HG= 105). The crossbar denotes the median and 25th/75th quantiles. Cancers with unspecified stage/grade were omitted (N = 11 and 2 for plasma and urine, respectively). Scores were capped to the interval (−6,6); see SI Appendix, Table S5 for non-visualized data points). (G) ROC curves for plasma, urine, and combined scores in the discrimination of cancers vs. healthy (N as in panel A). (H) Sensitivity at 95% specificity for the plasma, urine, and combined scores across different stage/grade groups (N as in panels D–F). Colors as in G. Error bars denote the 95% CI boundaries. (I) Sensitivity at 95% specificity for the plasma, urine, and combined scores across different cancer types (N as in SI Appendix, Table S4). Error bars denote the 95% CI boundaries. Colors as in G. (J and K) Cancer-type prediction using a Bayesian Additive Regression Trees model in the training (N = 110, five cancer types) and test (N = 74, five cancer types) sets. The numbers in the boxes represent the number of samples classified as belonging to the predicted cancer type. (L–N) Kaplan–Meier curves for OS across all cancer patients stratified into groups of “low” (undetected) vs. “high” (detected) plasma (N = 370, 13 cancer types), urine (N = 162, four cancer types), and combined (N = 152, four cancer types) scores. For each score, patients with scores greater than the 95% specificity cutoff were assigned to the “high” (black) vs. “low” (cyan) group. Hazard ratios for patients belonging to the “low” score strata and log-rank test P-values are shown under each curve. The panels show the number at risk for each group. Key: H, healthy; S1/LG, stage I or low grade; S2, stage II; S3, stage III; S4/HG, stage IV or high grade; see Table 1 for cancer types.

Fig. 2.

Validation of the pruned combined free GAGome MCED score. (A) Validation study subject flow (see also SI Appendix, Fig. S17). (B) Pruned combined free GAGome MCED scores across subjects with and without a cancer diagnosis 18 mo after the baseline visit (N_tot = 281, 110 controls vs. 158 cases; 13 cases with no stage information at diagnosis were omitted). Subjects that received a cancer diagnosis within 18 mo are grouped and colored according to the stage at diagnosis. The point and line range represent the median ± 1 standard deviation of the scores within each group. (C and D) Pruned combined free GAGome MCED scores in the subset of subjects with <4 mg/dL CRP and >1 mmol/L HDL-C (N_tot = 121, 72 controls vs. 49 cases; 4 cases with no stage information at diagnosis were omitted from panel C). Subjects that received a cancer diagnosis 18 mo after the baseline visit are grouped by the stage at diagnosis (C) and by time to death after diagnosis using the median overall survival (7.8 y) as cutoff for grouping (D). The point and line range represent the median ±1 standard deviation of the scores within each group. Key: S0, stage 0 (carcinoma in situ); S1, stage I; S2, stage II; S3, stage III; S4, stage IV.

Fig. 3.

Changes in plasma and urine free GAGomes during cancer progression in mice. (A) Experimental design overview. Principal component analysis of (B) plasma and (C) urine free GAGomes at different cancer progression timepoints. The values in parentheses on axes show the percentage of explained variance for the respective principal component. Ellipses indicate the 95% CIs for a bivariate t-distribution fitted to data points belonging to each cancer progression timepoint. Longitudinal level of (D) plasma and (E) urine 0S (non-sulfated) CS concentration per mouse at different cancer progression timepoints. Key: PC, principal component.