Combining tertiary lymphoid structures and tumor stroma percentage for predicting prognosis and chemotherapy benefits in stage II-III colorectal cancer

Abstract

Background

The current tumor-node-metastasis (TNM) staging system is insufficient for predicting the prognosis and chemotherapeutic benefits in stage II-III colorectal cancer (CRC) patients. We aimed to evaluate whether tertiary lymphoid structures (TLS) density and tumor stroma percentage (TSP) allow the prediction of stage II-III CRC survival and chemotherapy benefits.

Methods

The intratumoral TLS (In-TLS) density, peritumoral TLS (P-TLS) density, and TSP were assessed via whole-slide imaging with hematoxylin and eosin in the training and validation cohorts. The prognostic value of the TLS density and TPS as well as their association with chemotherapy response were assessed. Furthermore, two nomograms for predicting disease-free survival (DFS) and overall survival (OS) were developed and validated.

Results

We found that low density P-TLS and high TSP were significantly associated with poor prognosis (P < 0.001). Two nomograms based on depth of invasion, lymph node metastasis, CEA level, P-TLS, and TSP were subsequently developed. The nomograms outperformed the TNM stage in prognosis estimation (C-index: training cohort - DFS, 0.747 vs. 0.649; OS, 0.763 vs. 0.635; validation cohort - DFS, 0.733 vs. 0.641; OS, 0.736 vs. 0.649; P < 0.05 for all). The nomograms could add more net benefit than the TNM stage by decision curve analysis in the two cohorts. Moreover, chemotherapy had no impact on prognosis for patients with low density P-TLS and high TSP in both high-risk stage II disease [DFS, hazard ratios (HR): 0.844 (95% CI: 0.410–1.740), P = 0.647; OS, HR: 0.719 (95% CI: 0.328–1.577), P = 0.410] and stage III disease [DFS, HR: 0.846 (95% CI: 0.520–1.375, P = 0.500; OS, HR: 0.845 (95% CI: 0.504–1.417), P = 0.524]. For the remaining patients, adjuvant chemotherapy was associated with improved survival outcomes [high-risk stage II: DFS, HR: 0.367 (95% CI: 0.173–0.781), P = 0.009; OS, HR: 0.344 (95% CI: 0.149–0.790), P = 0.012; stage III: DFS, HR: 0.432 (95% CI: 0.265–0.706), P = 0.001; OS, HR: 0.374 (95% CI: 0.213–0.659), P = 0.001].

Conclusion

The P-TLS and TSP could improve prognostic prediction and serve as potential tools for identifying patients who could benefit from chemotherapy in stage II-III CRC patients.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12885-026-16542-w.

Introduction

Colorectal cancer (CRC) ranks as the third most prevalent and the second leading cause of mortality globally [1]. The tumor-node-metastasis (TNM) staging system serves as the predominant standard for prognostic assessment and therapeutic decision-making in CRC patients [2, 3]. Adjuvant chemotherapy is recommended following radical resection for patients with high-risk stage II and stage III CRC according to the National Comprehensive Cancer Network (NCCN) guidelines [2, 3]. Nevertheless, notable heterogeneity in the prognosis of CRC patients is evident, even among those with identical TNM stage and treatment regimens [4]. This finding suggested that the current TNM staging system is unable to accurately predict prognosis and chemotherapy benefits. Consequently, there is an imperative demand for new strategies that complement the TNM staging system, striving for precise prediction of patient prognosis and chemotherapy benefits in CRC patients.

Tertiary lymphoid structures (TLS) manifest as lymphoid formations within nonlymphoid tissues, comprising a T-cell zone where mature dendritic cells interact with T cells and a follicular B-cell zone [5]. TLS typically forms at the site of tumor infiltration when tumors occur and exert regulatory control over the tumor microenvironment by recruiting circulating immune cells and bolstering local immune responses [6]. Several studies have highlighted the positive impact of TLS on tumor resistance, with high density TLS indicating favorable prognoses in diverse tumors, such as lung cancer [7], pancreatic cancer [8], and breast cancer [9]. However, the prognostic value of TLS in CRC patients remains unclear according to previous studies [10, 11]. With the advancements in image digitization and digital storage technology, whole-slide imaging (WSI) enables pathologists to obtain a comprehensive view of pathological slides [12]. This technology serves as a potent tool for assessing the correlation between TLS density and location and the prognosis of CRC patients.

The tumor stroma constitutes an indispensable component of the tumor microenvironment, actively contributing to tumor initiation, progression, metastasis, and drug resistance [13–15]. A high tumor stroma percentage (TSP) is associated with a poor prognosis in patients with several solid tumors [16, 17]. Furthermore, the TSP could serve as an independent predictor of treatment response to neoadjuvant therapy in advanced rectal cancer and breast cancer [18–20]. Importantly, TSP can be readily assessed through hematoxylin and eosin (H&E) stained slides of tumor specimens, which exhibit high robustness and minimal intraobserver variation [21].

This study aimed to evaluate the correlation between density and location of TLS and TPS with prognosis in patients with stage II-III CRC, then, develop and validate individual survival prediction models for clinical decision-making. Furthermore, we explored whether the combination of TLS with TPS could identify patients with high-risk stage II and stage III CRC who may benefit from adjuvant chemotherapy.

Materials and methods

Patients

Ethics approval was obtained from the institutional review boards of Nanfang Hospital and Xiaogan Hospital (NFEC-2023-221). Written informed consent was obtained from all participants at the time of surgery. The study was conducted following the guidelines of the Declaration of Helsinki and the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines [22].

The inclusion criteria were patients who were histologically confirmed stage II-III primary CRC, underwent curative surgical resection, harvested at least 12 lymph nodes, and completed postoperative follow-up with clinicopathological characteristics. In the training cohort, a total of 428 consecutive patients were included from January 2013 to December 2016 at Nanfang Hospital after excluding those with antitumor treatment history (n = 235), emergency resection (n = 51), 30-day mortality (n = 6), or missing crucial clinical data (n = 282). The validation cohort included 274 consecutive patients from January 2013 to December 2016 from the Xiaogan Hospital who met the same inclusion and exclusion criteria. Detailed information can be found in Supplementary Figure S1.

Baseline clinicopathological characteristics, including age, sex, tumor location, obstruction or perforation, tumor size, carcinoembryonic antigen (CEA) level, differentiation status, venous emboli and/or lymphatic invasion and/or perineural invasion (VELIPI), depth of invasion, lymph node metastasis, receipt of adjuvant chemotherapy, and survival outcomes (recurrence and/or death), were collected. Patients who met chemotherapy indications received postoperative combination chemotherapy according to NCCN guidelines (primarily 8 cycles of XELOX [capecitabine plus oxaliplatin] or 12 cycles of FOLFOX [5-fluorouracil, leucovorin, and oxaliplatin]) after informed consent.

Whole-slide image acquisition

Two gastrointestinal pathologists with more than 10 years of experience who were blinded to the prognostic information used a microscope to initially review all H&E-stained slides of each patient. Then, they selected a representative slide, which should encompass both tumor tissue and the adjacent normal tissue surrounding the tumor. Formalin-fixed paraffin-embedded samples corresponding to representative slides were sectioned at a thickness of 5 μm and subjected to H&E staining. Quality control was performed by the director of the pathology department, with reproduction and re-examination needed if H&E-stained slides exhibited fragmentation, overlap, absence, or contamination. Subsequently, the WSIs of all H&E-stained slides were scanned using an Aperio ScanScope Scanner system (Leica Biosystems) with a 20× objective and managed through Aperio ImageScope software (version 12.3.3).

Assessment of TLS density and TSP

The process of TLS density assessment was as follows: two pathologists utilized Aperio ImageScope software to delineate the WSIs of H&E-stained slides into intratumoral and peritumoral regions and calculate the TLS area within each of the two regions. Subsequently, the In-TLS density (area of TLS in the intratumoral region/area of the intratumoral region) and the P-TLS density (area of TLS in the peritumoral region/area of the peritumoral region) were determined (Supplementary Figure S2A). The cutoff values were selected using X-tile software (version 3.6.1), a validated bioinformatics tool for biomarker assessment and outcome-based cut-point optimization that identifies optimal thresholds by maximizing the chi-square value between survival groups [23–25].

Similarly, TSP was determined by two pathologists using the appropriate WSIs. First, a representative region that was most invasive was selected at low magnification (×5 objective). Subsequently, a single region containing both the stroma and tumor was observed at high magnification (10× objective), and tumor cells were present on all sides of the field of view. Previous research has indicated that a TSP of 50% is a valid cutoff value [26, 27]; hence, in this study, a TSP > 50% was considered high, and a TSP ≤ 50% was considered low (Supplementary Figure S2B). When there was disagreement about TLS or TSP, a discussion was held, and the director of the pathology department made the final decision.

Development and validation of the nomograms

In the training cohort, variables were selected separately for DFS and OS outcomes. All candidate predictors (TLS, TSP, and clinicopathological characteristics) were first subjected to univariate Cox regression analysis. Variables with P < 0.10 in univariate analysis were then entered into multivariable Cox regression models. Backward stepwise elimination was applied to identify the most parsimonious set of independent predictors, using the likelihood ratio test with Akaike’s Information Criterion as the stopping rule [28]. Cox regression, the standard for survival analysis in oncology, over machine learning algorithms for its superior interpretability and alignment with clinical practice guidelines [29]. Moreover, recent literatures demonstrated that machine learning does not consistently outperform Cox regression, while sacrificing crucial transparency for clinical translation [30, 31]. Hence, this approach prioritizes both scientific rigor and practical utility in this study. Two prediction models were developed for predicting DFS and OS and subsequently tested in the validation cohort. Collinearity assessment for the predictors in the multivariable Cox regression model was performed using variance inflation factor (VIF) and tolerance. The collinearity criteria were defined as follows: severe collinearity among independent variables was indicated when tolerance ≤ 0.1; no collinearity when VIF < 3. To increase the usability of the prognostic models, we visually presented the models as nomograms with scores linearly mapped from the beta coefficients of the multivariable Cox regression model and further developed two web-based survival rate calculators [32, 33]. The web-based calculators represent a robust translational extension of the developed nomograms, as the calculators are directly constructed based on the scoring system of the nomograms and thus share an identical predictive framework. In addition to maintaining complete consistency with the nomogram results, the web-based calculators not only facilitate efficient clinical application but also minimize errors from manual nomogram scoring.

We compared the prognostic value of the nomograms and TNM stage for predicting survival outcomes in terms of discrimination, calibration, and clinical usefulness in the training and validation cohorts. Harrell’s concordance index (C-index) and time-dependent receiver operating characteristic curve (ROC) curve were used to evaluate the discriminatory ability. Calibration curves were generated to compare the predicted survival probabilities with the actual survival probabilities. Decision curve analysis (DCA) was used to evaluate the clinical usefulness [34, 35]. To quantify the relative improvement in prediction accuracy, the prediction error curves, net reclassification improvement, and integrated reclassification improvement were assessed [36].

SHapley additive exPlanations (SHAPs) for model interpretability

To interpret how each predictor influences model prediction for prognosis, we used Shapley values [37, 38]. Artificial intelligence provides a unified method for interpreting machine learning models. On the basis of the SHAP, we were able to determine the importance of individual predictors and other clinicopathological characteristics with interpretations of how they participated in the prediction of prognosis.

Sample size

The sample size calculation formula of Cox regression for binomially distributed data, proposed by Hsieh and Lavori [39], is as follows:

N represents the sample size, P₀ is the event rate X₁ = 0, P₁ is the event rate at X₁ = 1, R is the proportion of the sample with X₁ = 1, and is the overall event rate given by . We hypothesized that the type I error α was 0.05, the type II error β was 0.10, and the minimum sample size was 155. In this study, the sample sizes of the training (n = 428) and validation (n = 274) cohorts all exceeded the minimum sample size requirement.

Statistical analysis

Statistical analyses were performed with R (version 4.1.3) and SPSS (version 25). The Mann‒Whitney test was used to compare continuous variables. The chi-square test was used to compare categorical variables. The Kaplan–Meier method and log-rank test were used to evaluate patient prognosis. Univariate and multivariable analyses were performed using the Cox proportional hazards model, and hazard ratio (HR) and 95% confidence intervals (CI) were calculated. All the statistical analyses were based on two-sided analysis, and a P value of < 0.05 was considered statistically significant.

Results

Demographics and clinical characteristics

The study design is shown in Fig. 1. A total of 702 patients were included in the study; 63.8% (448/702) were men, and 52.7% (367/702) were ≥ 55 years old. The median (interquartile range, IQR) follow-up duration was 63.00 (IQR: 44.00–66.00) months in the training cohort and 62.00 (IQR: 59.75-64.00) months in the validation cohort. Comparable prognoses were observed in both cohorts, with 5-year disease-free survival (DFS) and overall survival (OS) rates of 70.0% and 73.9%, respectively, in the training cohort (Supplementary Figure S3A) and 5-year DFS and OS rates of 69.3% and 70.0%, respectively, in the validation cohort (Supplementary Figure S3B). The two cohorts exhibited balanced clinicopathological characteristics, which justified their use as training and validation cohorts (Table 1).

Fig. 1.

Study design for the discovery and validation of TLS density and TSP based on WSI images to assess prognosis and chemotherapy benefits in stage II-III colorectal cancer patients. WSIs were available for patients in both the training cohort and the validation cohort and were used to evaluate the associations of TLS density (intratumoral and peritumoral) and TSP with survival outcomes. Then, prognostic prediction models were developed and validated. Subsequently, the TLS density combined with TPS was used to predict chemotherapy benefits. Abbreviations: WSI, whole-slide imaging; H&E, hematoxylin and eosin; P-TLS, peritumoral tertiary lymphoid structures; TSP, tumor stroma percentage

Table 1.

Characteristics of the patients in the training and validation cohorts

Variables	Training cohort (n = 428)	Validation cohort (n = 274)	P
Age (years old)			0.847
< 55	203 (47.4)	132 (48.2)
≥ 55	225 (52.6)	142 (51.8)
Sex			0.613
Male	270 (63.1)	178 (65.0)
Female	158 (36.8)	96 (35.0)
Tumor location			0.443
Right-side	100 (23.4)	71 (25.9)
Left-side	328 (76.6)	203 (74.1)
Obstruction or perforation			0.427
No	259 (60.5)	174 (36.5)
Yes	169 (39.5)	100 (36.5)
Tumor size (cm)			0.485
< 5	238 (55.6)	145 (52.9)
≥ 5	190 (44.4)	129 (47.1)
CEA level			0.562
Normal	242 (65.6)	161 (58.8)
Elevated	186 (43.5)	113 (41.2)
Differentiation status			0.820
Well and Moderate	342 (79.9)	217 (79.2)
Poor and undifferentiated	86 (20.1)	57 (20.8)
VELIPI			0.284
No	226 (52.8)	156 (56.9)
Yes	202 (47.2)	118 (43.1)
Depth of invasion			0.563
pT1-2	68 (15.6)	52 (19.0)
pT3	277 (64.7)	172 (62.8)
pT4	83 (19.4)	50 (18.2)
Lymph node metastasis			0.873
pN0	205 (47.9)	126 (46.0)
pN1	151 (35.3)	99 (36.1)
pN2	72 (16.8)	49 (17.9)
TSP			0.548
Low	253 (59.1)	167 (61.7)
High	175 (40.9)	105 (38.3)
P-TLS			0.920
Low density	250 (58.4)	159 (58.0)
High density	178 (41.6)	115 (42.0)

Values in parentheses are percentages unless indicated otherwise

Abbreviations: IQR interquartile range, CEA carcinoembryonic antigen, VELIPI venous emboli and/or lymphatic invasion and/or perineural invasion, TSP tumor stroma percentage, P-TLS peritumoral tertiary lymphoid structures

Association of TLS and TSP with prognosis

The intratumoral TLS (In-TLS) density exhibited no correlation with prognosis, while the P-TLS density could distinguish between patients with different prognoses with an optimal cutoff value of 0.0359/mm² according to X-tile software (Supplementary Figure S4). Patients were subsequently stratified into a low density peritumoral TLS (P-TLS) group and a high density P-TLS group in both cohorts. There were no significant differences in P-TLS density or TSP values between the training and validation cohorts, confirming good balance between the two cohorts (Supplementary Figure S5).

A total of 178 (41.6%) and 115 (42.0%) patients had high density P-TLS in the training and validation cohorts, respectively. In the training cohort, patients with high density P-TLS exhibited 5-year DFS and OS rates of 84.6% and 87.6%, respectively, whereas those with low density P-TLS had 5-year DFS and OS rates of 59.7% and 68.0%, respectively (both P < 0.001; Supplementary Figure S6A). Similar findings were observed in the validation cohort; the 5-year DFS and OS rates were 83.4% and 87.8%, respectively, for the high density P-TLS patients and 59.1% and 69.2%, respectively, for the low density P-TLS patients (both P < 0.001; Supplementary Figure S6B).

The relationship between the TSP and survival outcomes was further evaluated. Low TSP was observed in 250 (58.4%) and 159 (58.0%) patients in the two cohorts, respectively. Patients with a low TSP demonstrated superior survival outcomes compared to those with a high TSP (5-year DFS: 80.1% vs. 55.5%; 5-year OS: 86.1% vs. 61.7%; both P < 0.001) (Supplementary Figure S7A). These results were validated in the validation cohort (5-year DFS: 79.8% vs. 52.4%; 5-year OS: 86.4% vs. 61.9%; both P < 0.001) (Supplementary Figure S7B).

Development and validation of the nomograms

Univariate and multivariable Cox regression analyses revealed that the depth of invasion, lymph node metastasis, CEA level, P-TLS, and TSP were independent predictors of DFS and OS (Supplementary Table S1 and Table 2). The tolerance values of the five predictors were 0.971, 0.969, 0.973, 0.992 and 0.980, with the corresponding VIF values of 1.020, 1.032, 1.028, 1.008 and 1.021, respectively. Collectively, these findings confirmed the absence of collinearity among the five predictors included in the model. Among these predictors, the P-TLS and TSP were important for prognosis prediction (Supplementary Figure S8 and Table S2). Based on the results of multivariable Cox regression, we developed two nomograms that integrated the five independent predictors to predict the probability of 2-, 3-, and 5-year DFS (Fig. 2A) and OS (Fig. 2B) in stage II-III CRC patients. To enhance clinical practicality, we established two user-friendly web-based calculators (DFS: https://nomogramforcrc.shinyapps.io/NomoforDFS/; OS: https://nomogramforcrc.shinyapps.io/Nomofor/). Both patients and clinicians can input relevant predictors to obtain individualized survival probabilities (Supplementary Figure S9). Clinical applicability was demonstrated through patient stratification into low- and high-risk groups via the nomogram-predicted score median, with significantly survival separation validated in both cohorts (Supplementary Figure S10). This provides concrete risk stratification cutoff for clinical practice and complements web-based calculators for individualized risk assessment. Risk estimates can be extracted from the prediction model by SHAP values to allow an explanation of recurrence risk. The analysis identified the five most important predictors for recurrence as P-TLS, TSP, lymph node metastasis, depth of invasion, and the CEA level, which were incorporated into our nomogram and demonstrated analytical consistency (Fig. 3).

Table 2.

Multivariable Cox regression analysis for disease-free survival and overall survival in the training cohort

Characteristic	Disease-free survival		Overall survival
	HR (95% CI)	P	HR (95% CI)	P
CEA level
Normal	Ref		Ref
Elevated	1.505 (1.062–2.134)	0.022	1.621 (1.105–2.377)	0.013
Depth of invasion
pT1-2	Ref		Ref
pT3	2.994 (1.537–5.831)	0.001	2.858 (1.372–5.954)	0.044
pT4	3.900 (1.875–8.113)	< 0.001	4.145 (1.854–9.266)	0.011
Lymph node metastasis
pN0	Ref		Ref
pN1	2.228 (1.445–3.434)	0.002	2.344 (1.450–3.789)	0.001
pN2	2.812 (1.788–4.423)	< 0.001	3.083 (1.887–5.038)	< 0.001
TSP
Low	Ref		Ref
High	2.765 (1.931–3.961)	< 0.001	2.579 (1.744–3.813)	< 0.001
P-TLS
Low density	Ref		Ref
High density	0.356 (0.232–0.546)	< 0.001	0.366 (0.229–0.588)	< 0.001

Abbreviations: HR hazard ratio, CI confidence interval, Ref reference, CEA carcinoembryonic antigen, TSP tumor stroma percentage, P-TLS peritumoral tertiary lymphoid structures

DFS model: likelihood ratio test: χ²= 97.373, P < 0.001

OS model: likelihood ratio test: χ²= 83.474, P < 0.001

Fig. 2.

Developed nomograms for the prediction of DFS and OS. A Developed nomogram for DFS. B Developed nomogram for OS. The nomograms were developed in the training cohort by incorporating the CEA level, depth of invasion, lymph node metastasis, TSP, and P-TLS. Abbreviations: CEA, carcinoembryonic antigen; TSP, tumor stroma percentage; P-TLS, peritumoral tertiary lymphoid structures

Fig. 3.

The features that predict the risk of disease-free survival. The features to predict the risk of recurrence. The influence of the value of the feature itself is shown on the y-axis. The color gradient encodes the predictor value (light = high value, dark = low value); a positive SHAP value increases the risk of recurrence, while a negative SHAP value decreases risk. For P-TLS, a high density (light color) corresponds to a negative SHAP value, consistent with its favourable prognostic effect (low recurrence risk); for TSP, a high percentage (light color) corresponds to a positive SHAP value, consistent with its poor prognostic effect (high recurrence risk). Abbreviations: SHAP, SHapley Additive exPlanations; P-TLS, peritumoral tertiary lymphoid structures; TSP, tumor stroma percentage; CEA, carcinoembryonic antigen; VELIPI, venous emboli and/or lymphatic invasion and/or perineural invasion

The calibration curves exhibited satisfactory agreement between the nomogram-predicted survival probabilities and the actual survival probabilities (Fig. 4). The nomograms showed a satisfactory C-index of 0.747 (95% CI: 0.726–0.767) for DFS and 0.763 (95% CI: 0.740–0.785) for OS in the training cohort. Similar results were found in the validation cohort for predicting prognosis [0.733 (95% CI: 0.706–0.760) for DFS and 0.736 (95% CI: 0.706–0.766) for OS]. Furthermore, the area under the receiver operating characteristic curves (AUROCs) of the nomograms for predicting 2-, 3-, and 5-year DFS were 0.795 (95% CI: 0.769–0.821), 0.787 (95% CI: 0.761–0.812), and 0.787 (95% CI: 0.724–0.819), respectively, and those for predicting 2-, 3-, and 5-year OS were 0.768 (95% CI: 0.697–0.839), 0.812 (95% CI: 0.763–0.862), and 0.800 (95% CI: 0.750–0.850), respectively, in the training cohort (Supplementary Table S3). In the validation cohort, the nomograms consistently demonstrated robust predictive performance: 0.791 (95% CI: 0.730–0.853), 0.777 (95% CI: 0.715–0.839), and 0.756 (95% CI: 0.692–0.821) for predicting 2-, 3-, and 5-year DFS; 0.786 (95% CI: 0.705–0.866), 0.779 (95% CI: 0.711–0.847), and 0.767 (95% CI: 0.700-0.835) for predicting 2-, 3-, and 5-year OS (Supplementary Table S4).

Fig. 4.

Calibration curves of the nomograms for DFS and OS. A Calibration curves of 2-, 3-, and 5-year DFS and OS in the training cohort. B Calibration curves of 2-, 3-, and 5-year DFS and OS in the validation cohort. Calibration curves show the calibration of the nomograms in terms of the agreement between the predicted and actual 2-, 3-, and 5-year outcomes. Left panel: DFS; right panel: OS. Abbreviations: DFS, disease-free survival; OS, overall survival

Comparison with the TNM stage

In the training cohort, the C-index of the TNM stage was 0.649 (95% CI: 0.626–0.673) for DFS and 0.635 (95% CI: 0.662–0.688) for OS. The nomograms showed a significantly superior C-index of 0.747 (95% CI: 0.726–0.767) for DFS and 0.763 (95% CI: 0.740–0.785) for OS compared with the TNM stage (all P < 0.001). Similar results were found in the validation cohort for predicting DFS (Supplementary Table S5). In addition, the time-dependent ROC curve demonstrated that the nomograms exhibited superior discrimination compared to TNM stage for predicting DFS [nomogram: 0.787 (95% CI: 0.724–0.819) vs. TNM stage: 0.675 (95% CI: 0.763–0.812); P < 0.001] and OS [nomogram: 0.800 (95% CI: 0.750–0.850) vs. TNM stage: 0.693 (95% CI: 0.634–0.753); P < 0.001] in the training cohort, as assessed by DeLong’s test (Fig. 5A and B and Supplementary Table S3). Consistent results were observed in the validation cohort, where the nomograms also outperformed TNM stage for predicting DFS [nomogram: 0.756 (95% CI: 0.692–0.821) vs. TNM stage: 0.660 (95% CI: 0.589–0.731); P = 0.006] and OS [nomogram: 0.767 (95% CI: 0.700-0.835) vs. TNM stage: 0.663 (95% CI: 0.588–0.739); P = 0.005] (Fig. 5C and D and Supplementary Table S4). Precision-recall curves confirmed the nomogram’s superior performance in identifying true high-risk patients compared to TNM stage in both cohorts (Supplementary Figure S11). At the optimal Youden threshold for 3-year and 5-year survival outcome prediction, the nomogram achieved higher sensitivity, specificity, accuracy, positive predictive value (PPV) and negative predictive value (NPV) in both cohorts (Supplementary Tables S6-S7). DCA demonstrated that the nomograms provided greater net benefit than TNM staging across a wide range of threshold probabilities in the training and validation cohorts (Fig. 6).

Fig. 5.

Time-dependent ROC curves of nomograms and TNM stage for DFS and OS. A, B The 5-year time-dependent ROC curves of nomograms and TNM stage for DFS and OS in the training cohort. C, D Time-dependent ROC curves of nomograms and TNM stage for DFS and OS in the validation cohort. Left panel: DFS; right panel: OS. Abbreviations: DFS, disease-free survival; OS, overall survival; TNM, tumor-node-metastasis; ROC, receiver operator characteristic curve; AUROC, area under the ROC curve

Fig. 6.

Decision curve analysis of nomograms and TNM stage for DFS and OS. A, B Decision curves of the nomograms and the TNM stage for predicting DFS and OS in the training cohort. C, D Decision curves of the nomograms and the TNM stage for predicting DFS and OS in the validation cohort. The y-axis represents the net benefit, and the x-axis represents the different threshold probabilities. Left panel: DFS; right panel: OS. Abbreviations: DFS, disease-free survival; OS, overall survival; TNM, tumor-node-metastasis

We further evaluated the improvement in prognosis prediction accuracy of the nomograms and TNM stage. The results revealed a net reclassification improvement of 0.243 (95% CI: 0.145–0.333) for DFS and 0.261 (95% CI: 0.168–0.354) for OS and an integrated discrimination improvement of 0.043 (95% CI: 0.012–0.084) for DFS and 0.045 (95% CI: 0.016–0.089) for OS in the training cohort. Similar results were observed when the nomograms were compared with the TNM stage in the validation cohort (Supplementary Tables S8-S9). The prediction error curves further demonstrated that the nomograms had better accuracy than the TNM stage (Supplementary Figure S12).

Relationship of the combination of P-TLS with TPS and chemotherapy benefits

In our cohort, 627 patients met chemotherapy indications; 337 (53.7%) received chemotherapy while 290 (46.3%) did not, with no significant clinicopathological differences between these groups (Supplementary Table S10). Adjuvant chemotherapy significantly improved the prognosis of high-risk stage II and stage III CRC patients (Supplementary Figure S13). We further investigated the relationship between P-TLS combined with TPS and adjuvant chemotherapy benefits. Patients were classified into class 1 (low density P-TLS + high TSP), while other patients were classified into class 2. The results confirmed that class 1 and class 2 were significantly associated with prognosis, regardless of whether patients received adjuvant chemotherapy (Supplementary Figure S14).

We further discovered that patients in class 2 could prolong survival in high-risk stage II and stage III CRC patients after the Bonferroni correction test (all P < 0.0125) [high-risk stage II: DFS, HR: 0.367 (95% CI: 0.173–0.781), P = 0.009; OS, HR: 0.344 (95% CI: 0.149–0.790), P = 0.012; stage III: DFS, HR: 0.432 (95% CI: 0.265–0.706), P = 0.001; OS, HR: 0.374 (95% CI: 0.213–0.659), P = 0.001]; however, no significant improvement in DFS or OS was observed in high-risk II and stage III CRC patients in class 1 [high-risk stage II: DFS, HR: 0.844 (95% CI: 0.410–1.740), P = 0.647; OS, HR: 0.719 (95% CI: 0.328–1.577), P = 0.410; stage III: DFS, HR: 0.846 (95% CI: 0.520–1.375, P = 0.500; OS, HR: 0.845 (95% CI: 0.504–1.417), P = 0.524] (Figs. 7 and 8). A statistical interaction test was conducted between class and chemotherapy, which confirmed that there was a significant interaction effect (all P < 0.05) for the impact on DFS and OS in high-risk stage II and stage III patients (Supplementary Table S11).

Fig. 7.

Adjuvant chemotherapy benefits in high-risk stage II patients in terms of DFS and OS. A, B Comparison of DFS and OS according to receipt of adjuvant chemotherapy in class 1 (low density P-TLS + high TSP) of high-risk stage II patients. C, D Comparison of DFS and OS according to receipt of adjuvant chemotherapy in class 2 (other patients) of high-risk stage II patients. P values were calculated by the log-rank test. Abbreviations: CT, chemotherapy; DFS, disease-free survival; OS, overall survival; HR, hazard ratio

Fig. 8.

Adjuvant chemotherapy benefits in stage III patients in terms of DFS and OS. A, B Comparison of DFS and OS according to receipt of adjuvant chemotherapy in class 1 (low density P-TLS + high TSP) of stage III patients. C, D Comparison of DFS and OS according to receipt of adjuvant chemotherapy in class 2 (other patients) of stage III patients. P values were calculated by the log-rank test. Abbreviations: CT, chemotherapy; DFS, disease-free survival; OS, overall survival; HR, hazard ratio

Discussions

Due to the significant heterogeneity of CRC, there is considerable variability in clinical outcomes even among patients at the same stage who are receiving similar treatment regimens. Therefore, precise prognosis assessment and classification of adjuvant chemotherapy benefits are essential for making appropriate treatment choices. In this study, we found that the P-TLS and TPS were independent prognostic factors for OS and DFS in stage II-III CRC patients, even after adjustment for clinicopathological characteristics. Then, we developed and validated two nomograms along with corresponding web-based survival rate calculators that integrated the P-TLS, TPS, CEA level, depth of invasion, and lymph node metastasis for predicting patient prognosis. The two nomograms provided satisfactory discrimination, calibration, and clinical usefulness in both cohorts. Importantly, we demonstrated that the combination of P-TLS with TPS could identify patients likely to benefit from adjuvant chemotherapy among high-risk stage II and stage III CRC patients.

TLS resemble secondary lymphoid organs and are classically described as aggregations of lymphocytes that develop at inflamed sites in the presence of autoimmune disease, infection, or tumors [5, 6]. TLS provide an important microenvironment for antitumor immune responses by facilitating antigen presentation and generating effector and memory T cells. TLS plays an important role in tumor progression and prognosis and is emerging as a potential biomarker for predicting survival outcomes in patients with various tumors. Increasing evidence suggests that high TLS density is an independent favorable prognostic indicator for lung cancer, pancreatic cancer, and breast cancer [7, 9]. Conversely, some studies have reported that TLS has no or limited relationship with improved survival. Regarding the location of TLS, P-TLS often have a favorable prognostic effect on lung cancer, pancreatic cancer, and ovarian cancer but are also associated with invasive metastasis in breast cancer. Furthermore, the In-TLS is considered a favorable predictor of lung cancer and hepatocellular cancer [40, 41]. These results suggest the heterogeneity of the antitumor effects exerted by the TLS in different tumors [42], and previous studies have focused primarily on local TLS density, which is a potential reason for the current contradictions in the prognostic value of TLS in CRC. The WSI is a powerful tool for comprehensively analyzing pathological slides. Therefore, in this study, In-TLS and P-TLS densities were evaluated separately using the WSI. The results showed that the In-TLS was not correlated with prognosis, while patients with high density P-TLS had significantly better survival than those with low density P-TLS. Currently, there are no universally accepted clinical standards for TLS detection and counting, and the methodology of this study may serve as an alternative for clinical workers. The ATOMIC trial demonstrated that adding atezolizumab to mFOLFOX6 significantly improved 3-year DFS in stage III dMMR patients, which highlighted the therapeutic potential of targeting immune checkpoints in colon cancer [43]. Notably, TLS represent organized lymphoid aggregates that facilitate antitumor immune responses, and our data showing high P-TLS density as a favorable prognostic factor may help identify patients most likely to benefit from immunotherapy.

The tumor stroma undergoes dynamic changes during tumor progression, forming regions that facilitate cancer progression and invasion [13]. The TPS has been identified as a prognostic biomarker for CRC patients and is simple to assess, cost-effective, and highly robust according to routine pathological examinations [44]. The most widely accepted cutoff value for the TPS is 50%, which has the highest predictive value for gastrointestinal cancer [15, 21, 26]. A meta-analysis indicated that a low TPS is associated with a poorer prognosis, consistent with the findings of this study [45]. Normal stroma may delay or prevent tumor development, while abnormal stromal components can promote tumor growth [46]. The tumor-activated stroma may induce epithelial cell disruption, tumor invasion, and immune evasion of tumor cells. The extracellular matrix can alter the tumor microenvironment, and high matrix metalloproteinase expression may promote tumor expansion and metastasis [47].

Several researchers have postulated prediction models for the prognosis of CRC patients [48]. Xue et al. constructed a radiomics signature for predicting survival outcomes based on CT imaging in 286 patients, achieving a C-index of 0.677–0.782 [49]. Molecular biomarkers, such as lncRNAs [50], circulating tumor cells [51], and methylation-based signature [52], also demonstrate potential predictive value in prognosis prediction. However, the additional technical requirements for these methods are often unattainable in some pathology departments. These molecular biomarkers extraction from serum samples also increases costs for patients, limiting clinical utility. Compared to those of other models, the nomograms constructed in this study were based on the P-TLS, TPS, and clinicopathological characteristics. These predictors can be derived from routine H&E-stained slides and clinical reports without increasing the economic burden on patients. Furthermore, to facilitate the translation of the established nomogram into clinical practice, we further developed two web-based survival rate calculators for better visualization and use by clinicians and patients.

In addition, we also investigated whether the combination of the P-TLS with TPS could identify patients who were more likely to benefit from adjuvant chemotherapy. Adjuvant chemotherapy is usually recommended for patients with high-risk stage II and stage III CRC [2, 3]. However, despite receiving the same chemotherapy regimen, approximately 30% of patients still experience recurrence [53]. These results suggest that some patients do not benefit from chemotherapy and may also suffer from toxic side effects. This study revealed that patients with class 1 (low density P-TLS + high TSP) did not have improved prognoses for either high-risk stage II or III disease, whereas patients with class 2 (other patients) experienced prolonged survival from adjuvant chemotherapy. These results suggested that the P-TLS and TSP combination may predict the benefit from adjuvant chemotherapy in patients with high-risk stage II and stage III disease.

POLE-mutant endometrial cancer with increased numbers of tumor-infiltrating lymphocytes may be correlated with reduced chemotherapy benefit or even potential harm [54, 55]. However, this study demonstrated that patients with high P-TLS density derive significant benefits from adjuvant chemotherapy, consistent with the Immunoscore. Immunoscore is based on CD3 + and CD8 + T-cell densities at tumor center and invasion margin. An international multicenter study demonstrated that patients with high Immunoscore benefit from chemotherapy [56]. These findings suggested that TLS not only predict favorable prognosis but also may serve as potential biomarkers for CRC treatment selection. Such differential responses underscore tumor-specific immune‒chemotherapy interactions and emphasize that biomarker-guided treatment selection may vary across tumor types, necessitating individualized approaches.

This study has several limitations. First, this was a retrospective study, and selection bias could not be avoided. Second, the underlying mechanisms through which the P-TLS and TPS combination influences the benefits of adjuvant chemotherapy are not yet fully understood. Third, microsatellite instability (MSI)/mismatch repair (MMR) status—now a standard stratification variable for adjuvant chemotherapy decisions in CRC—was not routinely tested in our retrospective cohort. Omission of MSI/MMR status may confound the chemotherapy-benefit subgroup analysis, as dMMR/MSI-H CRC patients have distinct immunological features and chemotherapy responsiveness compared to pMMR/MSS patients. Finally, the chemotherapy subgroup analysis, particularly for high-risk stage II class 1 patients, is limited by a modest sample size and requires further validation. Therefore, a prospective, international, and multicenter clinical trial incorporating molecular analyses is necessary to enhance predictive performance and validate model robustness.

Conclusions

The present study demonstrated that the P-TLS and TPS were significantly associated with prognosis, and the nomograms could serve as potential tools for guiding individualized care in stage II-III CRC patients. Moreover, the combination of P-TLS with TPS might be a useful predictive tool for determining the benefit from chemotherapy in high-risk stage II and stage III CRC patients.

Abbreviations

CEA

carcinoembryonic antigen

CI

confidence intervals

CRC

colorectal cancer

C-index

Harrell’s concordance index

DCA

decision curve analysis

DFS

disease-free survival

HR

hazard ratios

H&E

hematoxylin and eosin

In-TLS

intratumoral tertiary lymphoid structures

IQR

interquartile range

MMR

mismatch repair

MSI

microsatellite instability

NCCN

National Comprehensive Cancer Network

NPV

negative predictive value

OS

overall survival

P-TLS

peritumoral tertiary lymphoid structures

PPV

positive predictive value

PR

Precision recall

ROC

receiver operating characteristic curve

SHAPs

SHapley Additive exPlanations

STROBE

Strengthening the Reporting of Observational Studies in Epidemiology

TNM

tumor-node-metastasis

TSP

tumor stroma percentage

VELIPI

enous emboli and/or lymphatic invasion and/or perineural invasion

WSI

whole-slide imaging

Authors’ contributions

K.Y., X.D., H.L., W.J., and J.Y. conceived and designed the study; K.Y., H.L., X.D., L.J. and Z.L. collected and assembled the data; W.J., and J.Y. were responsible for data analysis and interpretation. All authors wrote and approved the paper.

Funding

This work was supported by grants from the National Natural Science Foundation of China (82503329 and 82273360), the GuangDong Basic and Applied Basic Research Foundation (2025A1515011769 and 2023A1515110639), the Fujian Province University-Industry Cooperation Project (2024Y4018), the Joint Funds for the innovation of Science and Technology, Fujian province (2024Y9282), the Postdoctoral Fellowship Program of CPSF (GZC20231069), the Fellowship of China Postdoctoral Science Foundation (2024M751321), the Science and Technology Program of Guangzhou (2023A04J2393), the Health Research Project of Hunan Provincial Health Commission (20257625), the Science and Technology Planning Project of Guangzhou City (2024A04J524), the Natural Science Foundation of Hunan Province of China (2026JJ60296), the Natural Science Plan of Xiaogan City, Hubei Province, 2025 (XGKJ2025010036), the President Foundation of Nanfang Hospital, Southern Medical University (2022B021 and 2023B016), and the Provincial Natural Science Foundation Innovation and Development Joint Fund of Hubei Province (2026AFC0678).

Data availability

The datasets used during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

Ethics approval was obtained from the institutional review boards of Nanfang Hospital and Xiaogan Hospital (NFEC-2023-221). Written informed consent was obtained from all participants at the time of surgery.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.