Study populations

From January 2016 to April 2023, 298 patients were eligible for inclusion in the research, 179 (60.07%) of these patients were male. For the fibrosis grade, 101 (33.89%) had no renal fibrosis, 100 (33.56%) had mild renal fibrosis, 55 (18.46%) had moderate renal fibrosis, and the other 42 (14.09%) patients had severe fibrosis (Table 2). IgA nephropathy, membranous nephropathy, and diabetic nephropathy were the most common pathological types in the cohort, with 72 (23.83%), 67 (22.48%), and 34 (11.41%) individuals, respectively. Other baseline characteristics are shown in Table 1 and Table 2.

Table 1 Baseline characteristics and biochemical measurements of patients enrolledTable 2 Baseline clinical characteristics and biochemical measurements in patients classified by fibrosis gradesLYVE1 was associated with renal fibrosis

A total of 2 patients had serum sLYVE1 concentrations outside the range of detection of the standard curve, and this missing value was addressed using multiple imputation. To assess the relationship between sLYVE1 and renal fibrosis, subjects were divided into four groups according to Banff classification. Notably, sLYVE1 levels rose with increasing grade of renal fibrosis. In detail, compared with the non-fibrosis (340.42 (255.87, 519.92)) or mild fibrosis (331.82 (238.79, 449.33)) groups, sLYVE1 concentrations were significantly increase in the moderate (467.83 (365.68, 547.89)) or severe fibrosis (609.22 (510.45, 752.74)) groups, and sLYVE1 expression was also significantly different between the moderate and severe fibrosis groups. However, there was no significant difference in non-fibrosis group and mild fibrosis group, which is consistent with the fact that lymphangiogenesis tends to occur in the middle to late stages of renal fibrosis (Fig. 1a). Therefore, next we focused on the diagnostic ability of sLYVE1 for moderate to severe fibrosis (MSF) and severe fibrosis (SF), with mild fibrosis being excluded from the exploration.

Fig. 1

a Serum sLYVE-1 levels in relation to grade of renal fibrosis. Results are presented as mean ± SD. One-way ANOVA was performed to compare differences between groups. b ROC curve of sLYVE1 to predict moderate-to-severe fibrosis. c ROC curve of sLYVE1 to predict severe fibrosis

The receiver operating characteristic curve (ROC) was plotted to search for a diagnostic cut-off of serum sLYVE1 for MSF and SF. Area under the curve (AUC) was calculated as 0.791 (95% CI 0.648–0.817) and 0.846 (95% CI 0.746, 0.840) with cut-off values of 405.25 ng/mL and 498.55 ng/mL, respectively, in MSF and SF group (Fig. 1b, c).

Further subgroup analyses based on renal pathology and complications were also performed. In the IgA nephropathy, membranous nephropathy, diabetic nephropathy, and lupus nephritis subgroups, sLYVE1 levels were higher in patients with moderate and severe fibrosis than in patients in the non-mild and mild groups, irrespective of pathology type (Fig. 2a–d). Similarly, in subgroup analyses based on the two most common comorbidities (hypertension and diabetes mellitus), sLYVE1 also both increased with increasing fibrosis degree (Fig. 2e, f).

Fig. 2

a–d Subgroup analysis according to different pathology types: a IgA nephropathy; b membranous nephropathy; c diabetic nephropathy; d lupus nephritis. e, f Subgroup analysis according to complications: e hypertension; f diabetes

Elevated serum sLYVE1 was not linked to decreased renal function

To clarify that the elevation of sLYVE1 was indeed due to fibrosis and not to circulating accumulation of serum small molecules due to impaired renal function, we performed a stratified analysis based on GFR. Serum sLYVE1 levels were significantly higher in patients with moderate-to-severe fibrosis than in patients with non-to-mild fibrosis in all subgroups with different GFR levels (Fig. 3a). Furthermore, no correlation between sLYVE1 and GFR was found after classifying patients according to fibrosis grade (Fig. 3b–e). Serum light chains have a similar molecular weight to LYVE1, and their levels correlate with abnormal plasma cell function rather than fibrosis, making them an ideal control for this study. Stratified analyses based on the presence or absence of fibrosis (Fig. 3f) or grade of fibrosis (Fig. 3g) showed no significant changes in the kappa light chains and λ light chains of the patients. These results demonstrate the increase in sLYVE1 levels is due to fibrosis rather than impaired renal filtration function.

Fig. 3

a Serum sLYVE1 levels were higher in patients with moderate-to-severe fibrosis than in patients with non-mild fibrosis, independent of GFR classification. b–d In all four fibrosis classes, sLYVE1 was not associated with GFR. e, f Serum kappa and lambda light chains did not significantly change with renal fibrosis

Feature selection and nomogram models

Since ROC curve analysis showed that serum sLYVE1 alone was a good indicator in determining MSF and SF, we moved on to investigate whether it could assist the diagnostic ability of traditional clinical tests. Therefore, a total of 298 patients were randomly divided into a training cohort and a validation cohort by a ratio of 2:1. As shown in Table 3, there were no statistical differences in baseline clinical indicators between the 2 cohorts.

Table 3 Baseline characteristics and biochemical measurements of two cohortsPredictive models incorporating sLYVE1 had better performance for MSF

The results of the stepwise regression suggested that the model consisting of sLYVE1, gender, GFR, and hypoalbuminemia had the lowest AIC value and VIF value below 3, indicating no collinearity between variables in the model. Binary logistic regression confirmed that high sLYVE1, males, and hypoalbuminemia were independent predictors of MSF and were associated with an increased risk of MSF (Table 4). Although the indicator of nephrotic syndrome-level-proteinuria (NS-proteinuria) had a p-value of >0.05 in the univariate logistic regression analysis, it was also included in the model due to its importance in clinical diagnosis.

Table 4 Results of logistic regression analysis of clinical indicators for moderate-to-severe fibrosis

To assess whether sLYVE1 enhances the diagnostic power of traditional clinical indicators, two models were developed. Model 1 consisted of traditional clinical indicators only, while model 1 incorporated sLYVE1 constituted model 2. Nomograms of models 1 and 2 are plotted in Fig. 4a and b.

Fig. 4

Models of predicting moderate-to-severe fibrosis. a, b Nomogram of models 1 (without sLYVE1) and 2 (with sLYVE1); c ROC curves of two models in the training cohort. d ROC curves of two models in the validation cohort. e DCA curves of two models in the training cohort. f DCA curves of two models in the validation cohort

We then validated the nomograms with ROC curves and calibration curves and used DCA curves to assess their clinical value. The calibration curves demonstrated good consistency in both the training cohort (Supplementary Fig. 2a, 2b) and the validation cohort (Supplementary Fig 2c, 2d). The C-index for model 2 was higher than model 1 in both the training (0.847 vs. 0.658) and validation sets (0.798 vs. 0.660), indicating better accuracy for model 2 (Fig. 4c, d). Also, DCA curves of model 2 exhibited larger net clinical benefits than model 1 in two cohorts (Fig. 4e, f).

Predictive models incorporating sLYVE1 had better performance for SF

To explore the diagnostic value of sLYVE1 as an aid in SF, we performed a similar assessment to that in MSF. Since the model with sLYVE1, hypoalbuminemia had minimal AIC value and low VIF (Table 5), they were incorporated into the models (Fig. 5a, b).

Table 5 Results of logistic regression analysis of clinical indicators for severe fibrosisFig. 5

Models of predicting severe fibrosis. a, b Nomogram of models 1 (without sLYVE1) and 2 (with sLYVE1); c ROC curves of two models in the training cohort. d ROC curves of two models in the validation cohort. e DCA curves of two models in the training cohort. f DCA curves of two models in the validation cohort

Calibration curves and DCA curves were displayed in Fig. 5c–f and Supplementary Fig 3, indicating that the model had a better consistency accuracy and greater net clinical benefits with the inclusion of sLYVE1. Furthermore, the results of ROC analysis revealed a better discriminate power of model 2 than model 1 in two sets (training set: 0.856 vs. 0.603; validation set: 0.835 vs. 0.655).