Novel plasma protein signatures for early diagnosis and prognostic stratification in colorectal cancer

Summary

Colorectal cancer (CRC) remains a major global health concern, underscoring the need for reliable biomarkers for diagnosis and prognosis. In this study, untargeted plasma proteomics was conducted across two cohorts comprising 321 participants in the discovery set and 353 in the validation set. Machine learning methods were employed to select biomarkers and develop diagnostic and prognostic models, with performance evaluated by the area under the receiver operating characteristic curve (AUC). An eight-protein panel for CRC and a six-protein panel for advanced adenoma discriminated from healthy controls with AUCs of 0.932 and 0.816 in the validation set, respectively. Additionally, a combined prognostic model incorporating eight proteins and clinical factors predicted 5-year disease-free survival in CRC with an AUC of 0.744 in the validation cohort. Overall, the identified protein biomarkers have the potential to foster the development of effective blood-based tests for early detection and prognostic prediction of CRC.

Graphical abstract

Highlights

• •Untargeted proteomics and machine learning identified diagnostic and prognostic CRC markers
• •Proteomic panels showed good performance in detecting CRC and advanced adenomas
• •An eight-protein model with clinical factors modestly predicted CRC disease-free survival
• •Plasma proteins may support early CRC detection and prognostic stratification

Cancer systems biology; Machine learning; Oncology; Precision medicine; Proteomics

Introduction

Colorectal cancer (CRC) ranks as the third most common malignancy and the second leading cause of cancer-related deaths worldwide.¹ By 2040, new CRC cases will increase by approximately 63% compared to 2020, reaching 3.2 million cases, with deaths rising to 1.6 million.² While the 5-year relative survival rate for localized CRC exceeds 90%, it drops to less than 10% for distant-stage CRC.³ Therefore, early diagnosis for timely treatment is crucial to reduce mortality and improve CRC prognosis.

Colonoscopy, the gold standard for detecting CRC and its precursors, is limited by high costs, invasive nature, complex intestinal preparation, and often low compliance.⁴ Despite the widespread use of the fecal immunochemical test, its sensitivity and specificity for detecting early-stage or premalignant lesions remain limited.⁵ Additionally, stool-based DNA tests have emerged as promising approaches for CRC detection in average-risk adults, but their sensitivity for advanced precancerous lesions is only 40%–45%, meaning nearly 60% of precancerous lesions are missed.^6,7 Carcinoembryonic antigen (CEA) is a blood glycoprotein used as a biomarker for monitoring treatment response and recurrence, but it performs poorly in early CRC detection.⁸ Thus, there is a critical need for novel diagnostic and prognostic biomarkers to enhance CRC clinical practices.

Circulating proteins, which enter the blood circulation through cellular leakage or active secretion, could provide valuable insights into human health status and serve as a crucial source of biomarkers as well as potential druggable targets. Aberrant levels of proteins, including those involved in systemic inflammation,⁹ insulin-like growth factors,¹⁰ and matrix metalloproteinases,¹¹ have been shown to play a pivotal role in the initiation and progression of CRC.

With the rapid development of mass spectrometry technology, proteomics has exhibited immense potential in systematically screening protein biomarkers. Several studies have identified various blood-based protein panels with favorable diagnostic accuracy for CRC, with reported area under the curves (AUCs) generally close to or exceeding 0.8.^12,13,14 For example, a case-control study of 47 CRC cases and 212 cancer-free controls identified a five-marker panel (leucine-rich alpha-2-glycoprotein [LRG1], epidermal growth factor receptor, inter-alpha-trypsin inhibitor heavy chain family member 4, hemopexin, and superoxide dismutase 3) with an AUC of 0.86 in internal validation.¹³ Another two-stage study identified a five-protein signature (LRG1, alpha 1 antitrypsin, apolipoprotein A I, haptoglobin, and paraoxonase lactonase 3) for CRC, with an AUC of 0.79 (95% confidence interval [CI], 0.70–0.86) in the validation set.¹⁴ However, there is limited overlap among the proteins identified across different studies. Moreover, most of these studies are constrained by small sample sizes (typically <100 participants), limited protein coverage (often less than 500 proteins), insufficient external validation, and a lack of exploration into the protein profiles of precancerous lesions.¹² These limitations impede the clinical translation of protein biomarkers. Additionally, metastasis and recurrence represent the dominant factors contributing to a poor prognosis in CRC. The 5-year relative survival rates are 31.6% for local recurrence and 21.6% for metastasis,¹⁵ highlighting the clinical significance of using disease-free survival (DFS) as a prognostic endpoint for early intervention. However, few proteomic studies have adopted DFS as a prognostic endpoint for CRC patients.

In this study, we employed an untargeted proteomics approach within a stringent two-stage case-control design to analyze plasma protein profiles of 674 participants. Utilizing advanced machine learning algorithms, we developed protein biomarker panels with potential diagnostic utility for advanced adenomas (AAs) and CRC. Furthermore, we carried out a prognostic analysis among CRC patients, with DFS as the endpoint, to identify novel non-invasive biomarkers for predicting prognosis.

Results

Study design and clinical characteristics

Basic characteristics of the participants from the Guangzhou discovery cohort, Nanjing validation cohort I, and Hangzhou validation cohort II are shown in Table S1. The mean age ranged from 55.8 to 63.2 years across the three datasets. The proportion of females was 40.2% in Guangzhou, 47% in Nanjing, and 25% in Hangzhou. The results of partial least squares discriminant analysis (PLS-DA) revealed that CRC, AA, and healthy control (HC) could be distinctly separated based on their proteomic profiles (Figure S1).

Construction of diagnostic model

In the discovery cohort, 332, 144, and 258 differentially expressed proteins (DEPs) were identified for CRC versus HC, AA versus HC, and CRC versus AA, respectively (Figure 1A). In validation cohort I, the corresponding numbers were 970, 383, and 273 (Figure 1B). Across the two cohorts, 133, 15, and 32 overlapping DEPs were identified for CRC vs. HC, AA vs. HC, and CRC vs. AA (Figures 1C and 2A–2C), respectively. Notably, two proteins, cartilage oligomeric matrix protein (COMP) and alpha-1-antichymotrypsin (SERPINA3), were consistently differentially expressed across all three pairwise comparisons (Figure 2D; Tables S2, S3, and S4). To further visualize the distinct protein profiles among CRC, AA, and HC samples, we conducted a heatmap analysis combined with hierarchical clustering (Figure S2).

Figure 1.

Differentially expressed protein profiles for three pairwise groups

Volcano plots in the (A) discovery cohort and (B) validation cohort I. Proteins significantly up-regulated are in red and down-regulated in blue. (C) Scatterplots for internal validation. The proteins with FDR <0.2 in both datasets are highlighted in blue, and the top 10 proteins ranked by −log10(FDR) are labeled. AA, advanced adenoma; CRC, colorectal cancer; FDR, false discovery rate; HC, healthy control.

Figure 2.

Venn diagrams showing the overlap of DEPs between the discovery cohort and validation cohort I

(A) CRC versus HC.

(B) AA versus HC.

(C) CRC versus AA.

(D) Overlap of DEPs across all three comparisons.

AA, advanced adenoma; CRC, colorectal cancer; DEP, differentially expressed proteins; HC, healthy control.

Based on the identified DEPs, we developed diagnostic models by utilizing the random forest algorithm in the discovery cohort. The DEPs included in the three diagnostic panels were visualized using boxplots (Figure S3). Among the 133 DEPs between CRC and HC, a panel of eight proteins comprising COMP, complement C1q tumor-necrosis-factor-related protein 3 (C1QTNF3), histone H4 (H4C1), LRG1, phosphoglucomutase-1 (PGM1), plakophilin-1 (PKP1), receptor-type tyrosine-protein phosphatase eta (PTPRJ), and angiopoietin-1 receptor (TEK) exhibited high discriminatory performance in differentiating CRC from HC, with an AUC of 0.978 (95% CI: 0.960–0.992) (Table 1; Figure 3A). Among the 15 DEPs between AA and HC, a panel of six proteins including contactin-4 (CNTN4), COMP, insulin-like-growth-factor-binding protein 5 (IGFBP5), L-lactate dehydrogenase A chain (LDHA), phosphoglycerate mutase 1 (PGAM1), and osteopontin (SPP1) demonstrated excellent discriminatory ability for AA and HC, yielding an AUC of 0.945 (95% CI: 0.915–0.970) (Table 1; Figure 3B). Additionally, among the 32 DEPs between CRC and AA, a panel of 11 proteins comprising fructose-bisphosphate aldolase B (ALDOB), calumenin (CALU), COMP, cathepsin B (CTSB), C1QTNF3, ficolin-1 (FCN1), interleukin-1 receptor type 2 (IL1R2), kit ligand (KITLG), LRG1, lumican (LUM), and SERPINA3 showed strong performance in differentiating CRC from AA, achieving an AUC of 0.962 (95% CI: 0.943–0.983) (Table 1; Figure 3C).

Table 1.

The diagnostic performance of protein panels in three cohorts

Model	Cohort	Sensitivity (95% CI)	Specificity (95% CI)	AUC (95% CI)
CRC vs. HC
COMP, C1QTNF3, H4C1, LRG1, PGM1, PKP1, PTPRJ, TEK	discovery cohort	0.972 (0.938–1.000)	0.869 (0.796–0.931)	0.978 (0.960–0.992)
	validation cohort I	0.973 (0.941–1.000)	0.830 (0.761–0.899)	0.973 (0.953–0.987)
	validation cohort II	0.833 (0.667–0.964)	0.792 (0.619–0.952)	0.932 (0.857–0.989)
AA vs. HC
CNTN4, COMP, IGFBP5, LDHA, PGAM1, SPP1	discovery cohort	0.907 (0.850–0.959)	0.841 (0.764–0.908)	0.945 (0.915–0.970)
	validation cohort I	0.807 (0.695–0.899)	0.652 (0.565–0.736)	0.804 (0.729–0.866)
	validation cohort II	0.750 (0.542–0.909)	0.750 (0.565–0.909)	0.816 (0.689–0.932)
CRC vs. AA
ALDOB, CALU, COMP, CTSB, C1QTNF3, FCN1, IL1R2, KITLG, LRG1, LUM, SERPINA3	discovery cohort	0.953 (0.909–0.991)	0.841 (0.769–0.908)	0.962 (0.943–0.983)
	validation cohort I	0.750 (0.670–0.827)	0.895 (0.812–0.968)	0.904 (0.855–0.948)
	validation cohort II	0.833 (0.654–0.960)	0.792 (0.625–0.947)	0.882 (0.775–0.969)

CRC, colorectal cancer; AA, advanced adenoma; HC, healthy control; CI, confidence interval; AUC, area under the curve.

Figure 3.

Receiver operating characteristic curve analysis for protein biomarkers in three pairwise comparisons

(A) CRC vs. HC, (B) AA vs. HC, and (C) CRC vs. AA. The red dot indicates the cutoff value, with the corresponding sensitivity and specificity. The figure displays the AUC and its corresponding 95% CI. CRC, colorectal cancer; AA, advanced adenoma; HC, healthy control; CI, confidence interval; AUC, area under the curve.

We then performed an internal validation using validation cohort I. The panel comprising eight proteins achieved an AUC of 0.973 (sensitivity = 0.973, specificity = 0.830) in distinguishing CRC from HC (Table 1; Figure 3A). Additionally, the panel established to differentiate AA from HC exhibited an AUC of 0.804 (sensitivity = 0.807, specificity = 0.652) (Table 1; Figure 3B). The panel for distinguishing CRC from AA yielded an AUC of 0.904 (sensitivity = 0.750, specificity = 0.895) (Table 1; Figure 3C). The fold changes of these proteins included in the panels were either <0.85 or >1.1 in the discovery cohort or validation cohort I.

To further validate our findings, we performed an external validation using validation cohort II. The eight-protein panel discriminated CRC from HC with an AUC of 0.932 (95% CI: 0.857–0.989), a sensitivity of 0.833, and a specificity of 0.792 (Table 1; Figure 3A). The six-protein panel for discriminating AA from HC achieved an AUC of 0.816 (95% CI: 0.689–0.932), a sensitivity of 0.750, and a specificity of 0.750 (Table 1; Figure 3B). The 11-protein panel discriminating CRC from AA yielded an AUC of 0.882 (95% CI: 0.775–0.969), a sensitivity of 0.833, and a specificity of 0.792 (Table 1; Figure 3C).

We calculated Shapley additive explanations (SHAP) values to assess the contribution of the included proteins to the diagnostic performance and ranked them in descending order (Figure S4). COMP, LDHA, and C1QTNF3 emerged as the top-ranking proteins in the models for CRC vs. HC, AA vs. HC, and CRC vs. AA, respectively, highlighting their significant discriminative roles in the diagnostic models.

Pathway analysis

Pathway enrichment analysis was conducted to identify pathways significantly enriched with DEPs across the three pairwise groups (false discovery rate [FDR]<0.05) (Figure S5). In CRC vs. HC, 133 DEPs were involved in 23 pathways, with the top three being complement and coagulation cascades, glycolysis/gluconeogenesis, and the extracellular matrix (ECM)-receptor interaction (Figure S5A). For AA vs. HC, 15 DEPs were primarily enriched in glycolysis/gluconeogenesis, biosynthesis of amino acids, and carbon metabolism (Figure S5B). For the 32 DEPs in CRC vs. AA, the top three pathways included complement and coagulation cascades, ECM-receptor interaction, and hematopoietic cell lineage (Figure S5C). Notably, the ECM-receptor interaction pathway was shared across all pairwise groups, with COMP as the common DEP (Tables S2, S3, and S4). COMP levels progressively decreased with the progression of the adenoma-carcinoma sequence in both the discovery cohort and validation cohort I (p for trend <0.05) (Figure S6).

Construction of prognostic model

After a median follow-up of 60 months in the discovery cohort, 23 CRC patients (22.5%) experienced recurrence, and one patient (1.0%) died. In the validation cohort I, with a median follow-up of 26 months, 33 CRC patients (30.3%) developed recurrent disease, and seven patients (6.4%) died. The 5-year survival rate was 77.3% in the discovery cohort and 54.5% in the validation cohort I.

In the discovery cohort, 55 DFS-associated proteins were identified, of which eight proteins were selected for the calculation of DFS score (Figures S7 and S8). Four proteins, including complement component C9 (C9), immunoglobulin G (IgG) receptor FcRn large subunit p51 (FCGRT), galectin-1 (LGALS1), and 6-phosphogluconate dehydrogenase (decarboxylating) (PGD) were positively associated with poorer DFS. In contrast, ankyrin repeat domain-containing protein 26 (ANKRD26), apolipoprotein A-IV (APOA4), platelet-derived growth factor receptor beta (PDGFRB), and serum paraoxonase 1 (PON1) were linked to improved DFS. Stratification analyses by sex and disease stage showed no significant heterogeneity in the associations of the major proteins across subgroups (Figure S9).

Kaplan-Meier survival curves showed that patients in the high DFS score group had a significantly worse DFS than those in the low DFS score group, both in the discovery cohort (log rank p < 0.001) and validation cohort I (log rank p < 0.001) (Figure 4A). Additionally, time-dependent receiver operating characteristic (ROC) curves demonstrated that DFS score had high predictive accuracy for DFS, with AUCs of 0.803 for 3-year and 0.833 for 5-year DFS in the discovery cohort and 0.705 and 0.690 in validation cohort I (Figure 4B). For the individual proteins included in the DFS score, the AUC, sensitivity, specificity, positive predictive value, and negative predictive value ranged from 0.60 to 0.77, 0.50 to 0.85, 0.55 to 0.87, 0.26 to 0.48, and 0.87 to 0.94, respectively, in the discovery cohort (Table S5). In validation cohort I, the corresponding values ranged from 0.52 to 0.67, 0.28 to 1.00, 0.17 to 0.87, 0.33 to 0.58, and 0.53 to 1.00, respectively (Table S5).

Figure 4.

Evaluation of the prognostic model in discovery and validation datasets

(A) Kaplan-Meier curves and log-rank tests for DFS.

(B) Time-dependent receiver operating characteristic analysis.

AUC, area under the curve; DFS, disease-free survival.

We identified three clinical factors significantly associated with DFS (p < 0.05), including body mass index (BMI), pathological T stage, and chemoradiotherapy (Table S6). When these clinical factors were combined with the DFS-Score, the performance of predicting DFS improved, with the AUC increasing from 0.811 to 0.892 in the discovery cohort and from 0.612 to 0.744 in the validation cohort I (Figure S10; Table S7). Additionally, a nomogram was constructed to predict 3- and 5-year DFS (Figure S11), showing acceptable prognostic predictive capabilities with C-index values of 0.843 and 0.855 in the discovery cohort and 0.718 and 0.698 in the validation cohort I, respectively (Figure S12).

Discussion

In this multi-center cohort study, we conducted comprehensive plasma proteomic profiling to examine protein alterations among CRC, AA, and HC groups. Utilizing advanced machine learning methods, we identified three diagnostic panels, which consisted of eight, six, and 11 proteins, respectively, which exhibited excellent performance in distinguishing CRC and AA from HC, as well as differentiating CRC from AA. Moreover, we established a novel prognostic model, termed DFS score, which incorporated eight proteins and demonstrated robust predictive capacity for DFS. These findings suggest that the identified plasma proteins hold promise for future clinical applications.

Previous proteomic studies have identified several protein panels for CRC detection, typically comprising 3 to 10 proteins, with reported AUC values ranging from 0.72 to 0.91.^16,17,18 For example, a case-control study involving 137 CRC patients and 137 HCs performed targeted plasma proteomics profiling and identified a panel of 13 proteins with an AUC of 0.91 in a validation dataset.¹⁶ A meta-analysis found MST1/STK4 (sensitivity = 0.68, specificity = 0.78), S100A9 (sensitivity = 0.72, specificity = 0.83), and TIMP1 (sensitivity = 0.42, specificity = 0.88) as promising blood-based diagnostic biomarkers.¹⁹ However, clinical translation remains hindered, partly due to the limited overlap of proteins across different studies and challenges in validating them in independent datasets. Notably, our multi-center validated diagnostic panel, which incorporated COMP as the top-ranked contributor along with seven proteins (LRG1, C1QTNF3, H4C1, PGM1, PKP1, PTPRJ, and TEK), achieved an AUC of 0.932 (sensitivity = 0.833, specificity = 0.792) in the validation cohort, demonstrating superior diagnostic performance for CRC compared to previous studies.

In our panel for distinguishing CRC from HC, LRG1, COMP, H4C1, TEK, and PGM1 have been previously reported as biomarkers, while the remaining three proteins were newly identified. Emerging evidence highlights the diagnostic potential of LRG1 in CRC early detection.^17,18,20 LRG1 has been shown to modulate epithelial-mesenchymal transition and angiogenesis in CRC through HIF-1α activation.²¹ Additionally, LRG1 could promote cell proliferation and inhibit apoptosis in CRC cells by regulating RUNX1 expression.²⁰ These findings collectively highlight the critical role of LRG1 in CRC pathogenesis and its potential as a diagnostic biomarker.

Our study revealed that COMP was significantly downregulated in CRC compared to HC, consistent with previous findings.²² Pathway analysis indicated that COMP was involved in the ECM-receptor interaction, which is closely associated with tumor occurrence and metastasis.²³ Notably, we observed differential expression of COMP in the other two pairwise comparisons, revealing a progressive decrease in COMP expression along the adenoma-CRC sequence. Consistently, previous case-control studies reported that plasma levels of H4C1 and TEK were elevated in CRC patients compared to HC.^24,25 Supporting our findings regarding PGM1 as a biomarker for CRC, experimental evidence suggests that PGM1 knockdown may stimulate CRC cell proliferation and inhibit cell-cycle arrest and apoptosis.²⁶ The findings underscore the potential of plasma proteomics in CRC diagnosis and provide a foundation for future biological studies to elucidate the underlying mechanisms.

AAs are precancerous lesions of CRC with a high risk of malignant transformation.²⁷ Early detection and timely intervention of AA can significantly reduce CRC risk and CRC-related mortality. However, to date, very few studies have investigated blood protein biomarkers for AA. Tan C et al. performed a proteomic analysis in serum samples from 19 AA patients and 8 healthy individuals, identifying seven proteins with statistically significant differences in expression levels.²⁸ Another study using a two-stage design developed a three-protein biomarker panel for AA detection, but it exhibited limited diagnostic accuracy (AUC = 0.65; 95% CI: 0.56–0.73) in a validation set.¹³ Utilizing machine learning algorithms, our multi-center study established a six-protein panel that demonstrated robust discriminative performance for AA detection, achieving an AUC of 0.816 (sensitivity = 0.750, specificity = 0.750) in the validation cohort, providing promising biomarkers for early CRC detection.

Only a few studies have explored proteins associated with CRC prognosis. Holm M et al. found that four plasma proteins were significantly associated with overall survival (OS) in 76 CRC patients.²⁹ Surinova S et al. developed a prognostic model combining six plasma proteins with three clinical factors to predict OS, achieving an AUC of 0.72.³⁰ In our study, we established a DFS score incorporating eight plasma proteins, including two previously reported proteins (C9 and LGALS1)^31,32 and six newly identified proteins (ANKRD26, APOA4, FCGRT, PDGFRB, PGD, and PON1). Notably, the DFS score demonstrated robust predictive ability in an external dataset, with 3- and 5-year DFS AUCs of 0.705 and 0.690, respectively, surpassing traditional biomarkers like CEA (AUC = 0.65).³³ Furthermore, integrating the DFS score with clinical parameters enhanced prognostic stratification, achieving superior predictive accuracy in the validation cohort (AUC = 0.744). Further research is needed to clarify the underlying mechanisms driving the prognostic associations.

In conclusion, our study found significant changes in plasma protein profiles between CRC, AA, and HC, suggesting proteins in plasma have potential for the diagnosis of CRC. Additionally, we developed a proteomic signature that predicted prognosis in CRC patients and was validated in independent populations. Nevertheless, these findings should be interpreted as exploratory, and further validation in independent cohorts and orthogonal experimental studies is needed before these candidate proteins can be considered biologically or clinically actionable.

Limitations of the study

First, although the proteins identified in this study showed potential as diagnostic and prognostic biomarkers for CRC and were further supported by validation in independent datasets, the study population was exclusively Chinese, which may limit the generalizability of the findings to other ethnic groups. Future studies in larger, multicenter, and multiethnic populations are therefore warranted to further validate these results. Second, the identified protein panels were not validated using independent methods, such as ELISA or western blot, and thus require additional verification in future studies. Third, given the relatively modest sample size and the likely small effect sizes of individual proteins in relation to CRC or AA, the application of a stringent multiple-comparison correction might be overly conservative. We therefore adopted a relatively relaxed threshold of FDR <0.2 for the initial screening of candidate proteins, without imposing an additional fold-change cutoff. Accordingly, the findings at the individual protein level should be interpreted with caution. Fourth, this study did not investigate the biological mechanisms underlying the observed associations of these proteins with CRC and AA, and further research is needed to clarify their roles in CRC development and progression. Finally, this study should be regarded as an exploratory investigation aimed at identifying proteins with potential translational value for CRC diagnosis and prognosis. Further efforts are required to validate these findings and facilitate their clinical application.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Dong Hang (hangdong@njmu.edu.cn).

Materials availability

This study did not generate novel reagents.

Data and code availability

The proteomic data have been deposited in Figshare (https://doi.org/10.6084/m9.figshare.32586630). The code supporting the analyses has been deposited in Figshare (https://doi.org/10.6084/m9.figshare.32586630) and GitHub (https://github.com/WWChen-hub/Protein-Signatures-for-Early-Diagnosis-and-Prognostic-Stratification-in-Colorectal-Cancer). Any additional information required to reanalyze the data reported in this article is available from the lead contact upon reasonable request.

Acknowledgments

We express our gratitude to all participants in this study. This work was supported by the Noncommunicable Chronic Diseases-National Science and Technology Major Project (2024ZD0520102), National Natural Science Foundation of China (82422067), and Outstanding Youth Foundation of Jiangsu Province (BK20230005).

Author contributions

X.L., data curation, formal analysis, software, and writing – original draft; W.C., data curation, methodology, and writing – original draft; L.W., data curation, methodology, and writing – review and editing; Y.Z., investigation, software, and writing – review and editing; X.G., L.D., C.F., Y.S., L.H., and X.C., writing – review and editing; X.D., N.L., L.M., H.M., and G.J., resources and writing – review and editing; X.H. and L.D., supervision, project administration, and writing – review and editing; D.H., conceptualization, funding acquisition, supervision, and writing – review and editing.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological samples
Blood	Participants in this study	N/A
Chemicals, peptides, and recombinant proteins
Urea	Sigma-Aldrich	U5128
Dithiothreitol	Sigma-Aldrich	SLBW0160
Iodoacetamide	Sigma-Aldrich	SLBW4985
Lysyl endopeptidase	Wako	125–05061
Trypsin	Promega	V5111
Trifluoroacetic Acid	Sigma-Aldrich	302031
Formic acid	Macklin	P809715
Acetonitrile	Fisher	A955-4
Deposited data
Raw data	This paper	Figshare: https://doi.org/10.6084/m9.figshare.32586630
Software and algorithms
R version 4.3.1	R Project	https://www.r-project.org/; RRID: SCR_001905
Python version 3.10	Python Software Foundation	https://www.python.org/; RRID: SCR_008394
Spectronaut version 14.10	Biognosys	https://biognosys.com/software/spectronaut/
Xcalibur version 4.3	Thermo Fisher Scientific	https://support.thermoinformatics.com/downloads/default.aspx
Original code for data analysis	This paper	GitHub: https://github.com/WWChen-hub/Protein-Signatures-for-Early-Diagnosis-and-Prognostic-Stratification-in-Colorectal-Cancer; Figshare: https://doi.org/10.6084/m9.figshare.32586630
Other
Top14 Abundant Protein Depletion Resin	Thermo Scientific	TA263966
SOLAμ HRP column	Thermo Scientific	60209–001
Acclaim PepMap C18 Trap Column	Thermo Scientific	6041.5240
Acclaim PepMap C18 Analytical Column	Thermo Scientific	164941
Thermo U3000nano RSLC nanoLC	Thermo Fisher Scientific	5041.0010A
Q Exactive HF-X Orbitrap high-resolution mass spectrometry	Thermo Fisher Scientific	06598L

Experimental model and study participant details

Study design and cohorts

A total of 674 participants were recruited from three hospitals in China. These included 107 CRC, 107 AA, and 107 HC from the Sixth Affiliated Hospital of Sun Yat-sen University in Guangzhou between May 2019 and December 2021; 112 CRC, 57 AA, and 112 HC from Nanjing Medical University Affiliated Second People’s Hospital in Nanjing between May 2019 and November 2021; as well as 24 CRC, 24 AA, and 24 HC from Zhejiang Cancer Hospital in Hangzhou between September 2022 and October 2022 (Figure 5). The Guangzhou cohort was used as the discovery dataset, while the Nanjing and Hangzhou cohorts were used as independent validation datasets. Across the three datasets, mean age ranged from 55.8 to 63.2 years, and the proportion of female participants ranged from 25.0% to 49.1% (Table S1).

Figure 5.

Study flowchart

(A) Diagnostic analysis; (B) prognostic analysis.

The recruitment criteria for the study were as follows: 1) CRC patients were required to have pathologically confirmed CRC and no prior history of cancer; 2) Participants with AA were selected from individuals who underwent colonoscopy at the same hospitals as CRC patients. Participants were excluded if they had undergone any treatment before enrollment, presented with intestinal inflammation, infectious intestinal injury, or diarrhea, had a prior history of colorectal surgery or other malignancies, or were diagnosed with hereditary/familial CRC. AA was defined as the presence of villous or tubulovillous features, a lesion size ≥1 cm, or high-grade dysplasia; and HC were recruited from individuals who received colonoscopy and were confirmed to have no abnormalities. CRC patients, AA patients, and HC were matched by age (within 2 years), sex, and hospital.

Ethics

All participants provided informed written consent. This study was approved by the Ethics Committees of Nanjing Medical University (2020KY092), Sun Yat-sen University (2021ZSLYEC082), and Zhejiang Cancer Hospital (IRB-2023-464).

Method details

Data collection

The following variables were collected for all participants, including age, sex, BMI, drinking status, smoking status, and family history of CRC (yes or no). BMI was calculated as weight in kilograms divided by height in meters squared (kg/m²). Tumor-related variables included tumor location, tumor size, TNM stage, tumor differentiation, and treatment. Tumor location was classified as right-sided (cecum, ascending colon, hepatic flexure, and transverse colon) or left-sided (splenic flexure, descending colon, and sigmoid colon). Tumor size was dichotomized at the mean value. All patients were restaged according to the 8th edition of the American Joint Committee on Cancer TNM staging system. Tumor differentiation was categorized as well or moderately differentiated versus poorly differentiated or undifferentiated. Treatment was classified as surgery alone or chemoradiotherapy plus surgery.

Protein extraction

Proteins were extracted from plasma samples as previously described.³⁴ Briefly, plasma samples were thawed on ice and diluted with ultrapure water. Subsequently, high-abundance proteins in plasma were depleted using Top14 abundant protein depletion resin (Thermo Scientific, USA). Proteins were then denatured (200 μL of 8 M urea, Sigma-Aldrich, USA), reduced (10 mM dithiothreitol at 37°C for 30 min; Sigma-Aldrich, USA), and alkylated (10 mM iodoacetamide for 30 min at 37°C in the dark; Sigma-Aldrich, USA). Enzymatic digestion was performed using lysyl endopeptidase (Wako, Japan) and trypsin (Promega, USA) overnight at 37°C. The digestion was terminated by adding 10 μL of 10% trifluoroacetic acid (Sigma-Aldrich, USA). Finally, plasma peptides were desalted using a SOLAμ HRP column (Thermo Scientific, USA) and dried under vacuum.

Liquid chromatography-mass spectrometry analysis

High-resolution and accurate profiling data were obtained using a data-independent acquisition (DIA) liquid chromatography-mass spectrometry (LC-MS) technique, which consisted of a Thermo U3000nano RSLC nanoLC (Thermo Fisher Scientific, USA) coupled to a Q Exactive HF-X Orbitrap high-resolution mass spectrometry (Thermo Fisher Scientific, USA).³⁵ Briefly, chromatographic separation was performed using an Acclaim PepMap C18 Trap column (75 μm × 2 cm, 3 μm) and Acclaim PepMap C18 analytical column (75 μm × 25 cm, 2 μm). The mobile phase consisted of 0.1% formic acid in water (solvent A) and 0.1% formic acid in 80% acetonitrile (solvent B). Each MS cycle included one full MS scan and 60 DIA scans. The full MS scan parameters were set as follows: resolution of 60,000, automatic gain control (AGC) of 3e6, maximum injection time (IT) of 20 ms, and a scanning range of 350-1200 m/z. For DIA scans, the parameters were as follows: the resolution of 30,000; AGC target of 1e6, maximum IT of 45 ms, and a scanning range of 390-1110 m/z. MS/MS data were acquired in DIA mode using XCalibur 4.3 software (Thermo Fisher Scientific, USA).

DIA data analysis

The DIA raw files were analyzed using Spectronaut version 14.10 (Biognosys, Schlieren, Switzerland), searching against a spectral library containing 55,157 peptides and 5,328 proteins.³⁴ To ensure data quality and robustness, comprehensive quality control (QC) assessments were performed in both the discovery and validation cohorts (Figures S13 and S14). Per-sample peptide and protein quantification depths were generally consistent across cohorts, with 8278 ± 938 peptides and 1125 ± 84 proteins in the discovery cohort, and 8684 ± 1418 peptides and 1200 ± 164 proteins quantified per sample in the validation cohort I. Reproducibility was supported by median Pearson correlation coefficients exceeding 0.8 among sample groups, while the median coefficient of variation across all QC samples throughout the injection sequence remained below 20%. Batch effects were assessed by unsupervised principal component analysis, which suggested minimal batch effects. The dynamic range of protein abundance spanned across 5 orders of magnitude (Figure S15). Proteins with missing values exceeding 50% were removed, and the remaining missing values were imputed using the k-nearest neighbor method. Log-transformation and z-score standardization were applied to normalize and homogenize the data. A total of 1,508 proteins were identified in the discovery cohort, of which 1,312 were quantified and retained for analysis; in the validation cohort I, 2,118 proteins were identified, of which 1,753 were quantified and retained for analysis (Table S8).

Follow-up for prognostic analysis

For the prognostic analysis, five patients with stage IV disease were excluded. CRC cases in the discovery cohort and validation cohort I were followed up from the date of CRC diagnosis to the first occurrence of recurrence, death, or end of follow-up. Data on disease progression were collected through telephone interviews and medical records by trained staff annually. Mortality data were obtained from death surveillance systems. CRC cases were followed up until September 2022, with maximum follow-up durations of 77.7 months in the discovery cohort and 61.2 months in validation cohort I. Three CRC patients in the validation cohort I were lost to follow-up and thus excluded from the prognostic analysis. DFS was defined as the length of time since initial cancer treatment during which no evidence of cancer recurrence or death was found.³⁶ CRC recurrence encompassed both local recurrence and distant metastases.³⁷

Quantification and statistical analysis

Diagnostic analysis

PLS-DA and hierarchical cluster analysis were conducted to visualize the discrimination in plasma proteomic profiles between CRC, AA, and HC groups. Age-adjusted linear regression models were used to identify DEPs. The Benjamini-Hochberg procedure was applied to adjust for multiple testing using FDR estimation, and proteins with FDR <0.2 in both the discovery cohort and validation cohort I were considered DEPs. No predefined fold-change cutoff was imposed during DEP selection. To further explore potential biological processes, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis was performed.

Diagnostic models were constructed using the random forest algorithm and optimized through a grid search algorithm. Additional details are provided in the Methods S1. The developed models were initially tested in the internal dataset (validation cohort I) and subsequently validated in an external dataset (validation cohort II). Diagnostic performance was evaluated by calculating sensitivity, specificity, and the AUC of the ROC curve. Additionally, SHAP values were used to assess the contribution of each selected feature in distinguishing different groups.

Prognostic analysis

In the discovery cohort, univariate Cox regression analysis was performed to identify proteins associated with DFS. A prognostic proteomic signature was then developed using the least absolute shrinkage and selection operator (LASSO) Cox regression analysis. The DFS-Score for each CRC patient was calculated using the following formula:

$$DFS-Score=\sum _{j=1}^n{Expression}_j{\ast Coefficient}_j$$

where Expression and Coefficient represent the expression level of each protein and its corresponding LASSO coefficient, respectively. CRC patients in each dataset were divided into high and low DFS-Score groups based on the median DFS-Score. Survival curves were generated by Kaplan-Meier analysis and compared using the log-rank test. The predictive accuracy of DFS-Score was determined by the time-dependent ROC curve. Stratified analyses were conducted according to sex (male and female) and pathologic stage (early stage: stage 0, I, or II; late stage: stage III). A prognostic nomogram was constructed by integrating DFS-Score with clinical factors (T stage, BMI, chemoradiotherapy) identified through Cox regression analysis. All statistical analyses were performed using R (version 4.3.1) and Python (version 3.10).