A serum biomarker panel and miniarray detection system for tracking disease activity and flare risk in lupus nephritis

Abstract

Introduction: Lupus nephritis (LN) leads to end stage renal disease (ESRD), and early diagnosis and disease monitoring of LN could significantly reduce the risk. however, there is not such a system clinically. In this study we aim to develop a biomarker-panel based point-of-care system for LN.

Methods: Immunoassay screening combined with genomic expression databases and machine learning techniques was used to identify a biomarker panel of LN. A quantitative biomarker-panel mini-array (BPMA) system was developed and the sensitivity, specificity, reproducibility, and stability of the were examined. The performance of BPMA in disease monitoring was validated with machine models using a larger cohort of LN. The BPMA was also used to determine LN flare using a machine-learning generated flare score (F-Score).

Results: Among 32 promising LN serum biomarkers, VSIG4, TNFRSF1b, VCAM1, ALCAM, OPN, and IgG anti-dsDNA antibody were selected to constitute an LN biomarker Panel, which exhibited excellent discriminative value in distinguishing LN from healthy controls (AUC = 1.0) and active LN from inactive LN (AUC = 0.92), respectively. Also, the 6-biomarker panel exhibited a strong correlation with key clinical parameters of LN. A multiplexed immunoarray was constructed with the 6-biomarker panel (named BPMA-S6 thereafter). An LN-specific 8-point standard curve was generated for each protein biomarker. Cross-reaction between these biomarkers was minimal (< 1%). BPMA-S6 test results were highly correlated with those from ELISA (Spearman's correlation: fluorescent detection, rs = 0.95; colorimetric detection, rs = 0.91). The discriminative value of BPMA-S6 for LN was further validated using an independent cohort (AUC = 0.94). Using a longitudinal cohort of LN, the derived F-Score exhibited superior discriminative value in the training dataset (AUC = 0.92) and testing dataset (AUC=0.82) to distinguish flare vs remission.

Conclusion: BPMA-S6 may represent a promising point-of-care test (POCT) for the diagnosis, disease monitoring, and assessment of LN flare.

Keywords: biomarker panel; disease monitoring; flare assessment; lupus nephritis; point-of-care diagnostics.

Figure 1

Construction and optimization of the lupus biomarker-panel miniarray (BPMA-S6). (A) Schematic of the slide construction for BPMA-S6. (B) Active control image of antibody/antigen contactless spotting. (C) Boxplot of signal intensities across different serum dilution ratios. (D) Array images showing the performance buffers in both fluorescent and colorimetric groups. (E) Prototype of the 3D-printed BPMA washer. (F) Customized portable colorimetric microscope. (G) Smartphone application developed for BPMA-S6.

Figure 2

Assaying 30 serum samples with BPMA-S6. (A) Representative BPMA array images in fluorescent and colorimetric settings using serum samples from LN and HC. (B, C) A total of 30 serum samples from active LN (LN-A, N = 10, red), inactive LN (LN-I, N = 10, green), and healthy controls (HC, N = 10, blue) were tested using BPMA fluorescent (B) and colorimetric (C) signals. (D) Linear regression and Spearman’s correlation analyses were performed to assess the correlation between BPMA and ELISA measurements. Asterisks indicate the level of statistical significance: n.s. p > 0.05; ^* p < 0.05; ^** p < 0.01; ^*** p < 0.001; ^**** p < 0.0001.

Figure 3

Omics selection for the LN serum biomarker panel. (A) Heatmap of the ELISA-based normalized protein levels for 17 promising LN serum biomarkers from LN-Active (N = 49), LN-Inactive (N = 13), and healthy control (N = 23) groups, with row clustering performed using Euclidean distance. (B) Barplot of AUC values (upper) for individual biomarkers distinguishing LN from HC, or LN-Active (LN-A) from LN-Inactive (LN-I), and heatmap of the Least Absolute Selection Shrinkage Operator (LASSO) weight scores (lower) for the biomarker panel distinguishing LN from HC, or LN-A from LN-I. (C) Correlation plot of serum biomarker levels with clinical parameters, where the color of each square represents Spearman’s correlation coefficient value, and significance levels are indicated with asterisks. (D) Gene expression profiles of validated biomarkers at six tissue-specific bulk-seq databases, with square colors representing fold-changes between lupus and HC, and significance levels indicated with asterisks. (E) Potential cellular origins of the selected biomarkers were identified using a lupus scRNA database comprising 1.2 million PBMCs. The color of each square represents the normalized expression level. (F) Protein–protein interaction (PPI) network among the 17 promising biomarkers constructed using the STRING database, where node colors indicate the associated pathways and edge colors represent predicted functional associations: yellow for text-mining evidence, black for coexpression evidence, purple for experimental evidence, and blue for sequence similarity evidence. ^* p < 0.05; ^** p < 0.01; ^*** p < 0.001; ^**** p < 0.0001.

Figure 4

Evaluation of the BPMA-S6 for disease diagnostics and monitoring. (A) Principal component analysis (PCA) of a cross-sectional cohort of 89 subjects, including 36 with active lupus nephritis (LN-Active, red), 19 with inactive lupus nephritis (LN-Inactive, green), 11 with chronic kidney disease (CKD, blue), and 23 healthy controls (HC, purple). (B, C) Six machine-learning models based on different learning principles were trained using threefold 100-iteration cross-validation to address binary classification problems: LN vs. HC (B) and LN-Active vs. LN-Inactive (C). (D) Confusion matrix of a fine-tuned multiclass classification Random Forest model evaluated on a testing dataset of 36 samples. The model was trained using 60% (53 samples) of the data for four-group classification.

Figure 5

LN flare assessment with ML-based Flare-Score. (A, B) Performance of the Flare-Score in tracking disease progression in the training dataset (A) and testing dataset (B) for four representative LN patients. A total of eight patients with 33 visit samples were used to train the LASSO model to generate the Flare-Score, and seven patients with 18 visit samples were used to test the model. The x-axis represents the visiting month, while the y-axis shows the Flare-Score and disease activity indexes. Dot colors indicate LN flare status. (C, D) The discriminatory abilities of the Flare-Score and BPMA-S6 panel markers in distinguishing LN flare from remission were evaluated in the training dataset (C) and testing dataset (D). (E) The optimal cut-point for the Flare-Score (red dot, 10.05) was determined based on the highest combined sensitivity and specificity across both datasets. (F) Distribution of Flare-Score diagnoses in the flare (left) and remission (right) groups; the red line on the x-axis indicates the optimal Flare-Score cut-point (10.05), and the y-axis shows the density of flare (left) and remission (right) samples.