A proteomic classifier panel for early screening of colorectal cancer: a case control study

Abstract

Background: Diagnosis of colorectal cancer (CRC) during early stages can greatly improve patient outcome. Although technical advances in the field of genomics and proteomics have identified a number of candidate biomarkers for non-invasive screening and diagnosis, developing more sensitive and specific methods with improved cost-effectiveness and patient compliance has tremendous potential to help combat the disease.

Methods: We enrolled three cohorts of 479 subjects, including 226 CRC cases, 197 healthy controls, and 56 advanced precancerous lesions (APC). In the discovery cohort, we used quantitative mass spectrometry to measure the expression profile of plasma proteins and applied machine-learning to select candidate proteins. We then developed a targeted mass spectrometry assay to measure plasma concentrations of seven proteins and a logistic regression classifier to distinguish CRC from healthy subjects. The classifier was further validated using two independent cohorts.

Results: The seven-protein panel consisted of leucine rich alpha-2-glycoprotein 1 (LRG1), complement C9 (C9), insulin-like growth factor binding protein 2 (IGFBP2), carnosine dipeptidase 1 (CNDP1), inter-alpha-trypsin inhibitor heavy chain 3 (ITIH3), serpin family A member 1 (SERPINA1), and alpha-1-acid glycoprotein 1 (ORM1). The panel classified CRC and healthy subjects with high accuracy, since the area under curve (AUC) of the training and testing cohort reached 0.954 and 0.958. The AUC of the two independent validation cohorts was 0.905 and 0.909. In one validation cohort, the panel had an overall sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of 89.9%, 81.8%, 89.2%, and 82.9%, respectively. In another blinded validation cohort, the panel classified CRC from healthy subjects with a sensitivity of 81.5%, specificity of 97.9%, and overall accuracy of 92.0%. Finally, the panel was able to detect APC with a sensitivity of 49%.

Conclusions: This seven-protein classifier is a clear improvement compared to previously published blood-based protein biomarkers for detecting early-stage CRC, and is of translational potential to develop into a clinically useful assay.

Keywords: Colorectal cancer; Early detection; Mass spectrometry; Protein biomarker.

Fig. 1

Study design. Flow chart showing the discovery and validation phases of this study. A Discovery cohort for quantitative proteomic analysis using DDA and DIA methods. B MRM targeted proteomic assay development. C Applying the MRM assay to an independent validation cohort of patients to classify CRC and APC patients from healthy controls. D Applying the MRM assay to a blinded validation cohort, in which the diagnosis was disclosed only after the test results were given

Fig. 2

Quantitative proteomic analysis of plasma samples from CRC patients and healthy controls. A Volcano plot of DDA data obtained from discovery cohort 1. Differentially expressed proteins are shown in blue (down) or red (up) circles. X-axis shows log2-fold change of plasma proteins between CRC patients and healthy subjects, and y-axis shows log10 of statistical significance values. B Volcano plot of DIA data obtained from discovery cohort 2. The label for the differentially expressed proteins and the two axes are the same as in A. C Principal component analysis of the protein expression data in cohort 1. D Principal component analysis of the protein expression data from cohort 2. E Gene ontology analysis of up-regulated proteins in the DIA data. F Gene ontology analysis of down-regulated proteins in the DIA data

Fig. 3

Seven feature proteins selected based on DDA and DIA proteome data. A and B Boxplot of protein intensity of feature proteins selected for logistic regression. Differential expression of seven proteins between CRC patients (C) and healthy subjects H from the DDA data (A) and the DIA data (B) are shown. C Principal component analysis using the expression levels of the seven proteins in cohort 1. D Principal component analysis using the expression levels of the six proteins in cohort 2

Fig. 4

MRM quantification and logistic regression classification of CRC and healthy subjects. A ROC curves of a seven-protein logistic regression classifier (LRG1, C9, IGFBP2, CDNP1, ITIH3, SERPINA1, and ORM1) for distinguishing CRC and healthy subjects in the training and testing datasets. B ROC curve showing the performance of the seven-protein classifier in distinguishing CRC and healthy subjects in an independent validation cohort. C ROC curve showing the performance of the seven-protein classifier in distinguishing CRC and healthy subjects in a blinded validation cohort. D–F Confusion matrix showing the classification accuracy in the training (D), independent validation E, and blinded validation F cohorts

Fig. 5

Expression patterns of the protein biomarkers in different CRC stages. A Box plot of plasma concentration of the seven proteins in four different CRC stages. B Sensitivity of the seven-protein biomarker in distinguishing CRC at four different stages from healthy subjects. C Sensitivity of the seven-protein biomarker in distinguishing APC with different sizes from healthy subjects. D Sensitivity of the seven-protein biomarker in identifying precancerous lesions with different grades