Investigation of ferroptosis-associated molecular subtypes and immunological characteristics in lupus nephritis based on artificial neural network learning

Abstract

Background: Lupus nephritis (LN) is a severe complication of systemic lupus erythematosus (SLE) with poor treatment outcomes. The role and underlying mechanisms of ferroptosis in LN remain largely unknown. We aimed to explore ferroptosis-related molecular subtypes and assess their prognostic value in LN patients.

Methods: Molecular subtypes were classified on the basis of differentially expressed ferroptosis-related genes (FRGs) via the Consensus ClusterPlus package. The enriched functions and pathways, immune infiltrating levels, immune scores, and immune checkpoints were compared between the subgroups. A scoring algorithm based on the subtype-specific feature genes identified by artificial neural network machine learning, referred to as the NeuraLN, was established, and its immunological features, clinical value, and predictive value were evaluated in patients with LN. Finally, immunohistochemical analysis was performed to validate the expression and role of feature genes in glomerular tissues from LN patients and controls.

Results: A total of 10 differentially expressed FRGs were identified, most of which showed significant correlation. Based on the 10 FRGs, LN patients were classified into two ferroptosis subtypes, which exhibited significant differences in immune cell abundances, immune scores, and immune checkpoint expression. A NeuraLN-related protective model was established based on nine subtype-specific genes, and it exhibited a robustly predictive value in LN. The nomogram and calibration curves demonstrated the clinical benefits of the protective model. The high-NeuraLN group was closely associated with immune activation. Clinical specimens demonstrated the alterations of ALB, BHMT, GAMT, GSTA1, and HAO2 were in accordance with bioinformatics analysis results, GSTA1 and BHMT were negatively correlated with the severity of LN.

Conclusion: The classification of ferroptosis subtypes and the establishment of a protective model may form a foundation for the personalized treatment of LN patients.

Keywords: Ferroptosis; Immune; Lupus nephritis; Machine learning; Molecular subtype; Protective model.

Fig. 1

The flow chart of this study

Fig. 2

Identification of differential expressed ferroptosis regulators in LN patients. A, B The PCA plot exhibiting the expression profiles of GSE32591 and GSE127797 before (A) and after (B) correction of batch effect. C Venn diagram exhibiting 10 differentially expressed FRGs in LN patients. D Volcano plot depicting the mRNA expression levels of FRGs between healthy individuals and LN patients. E Heatmap exhibiting the differentially expressed FRGs between the aforementioned groups. F, G Correlation plot (F) and network diagram (G) of 10 differentially expressed FRGs. Positive correlations were marked in blue and negative correlations were marked in red. The size of the rectangle showing the specific value of the correlation coefficients

Fig. 3

Immunological characteristics of normal subjects and LN patients and correlations between the characteristic FRGs and immune cells. A Stack chart exhibiting the relative proportions of 22 infiltrated immune cells from LD and LN samples. B Box plots exhibiting the alterations in infiltrated immune cells between LD and LN groups. C Heatmap exhibiting the correlations between characteristic FRGs and distinct immune cell compositions. *p < 0.05, **p < 0.01, ***p < 0.001. Positive correlations were marked in red and negative correlations were marked in blue. The size of the circle showing the specific value of the correlation coefficients

Fig. 4

Characterization of molecular subtypes based on characteristic FRGs. A Consensus matrix in LN patients (k = 2). Both the rows and columns of the matrix represent samples. The consistency matrix ranges from 0 to 1, from white to dark blue. B Consensus CDF when k = 2 to 6. C Delta area under CDF curve. D Consensus clustering score exhibiting the subtype score when k = 2–6. E t-SNE exhibiting that LN patients were classified into two distinct ferroptosis subtypes

Fig. 5

Identification of distinct biological functions and signaling pathways between subtypes. A, B GSVA exhibiting distinct biological processes (A) and signaling pathways (B) between subgroups

Fig. 6

Association between ferroptosis subtypes and immunological features. A Stack chart exhibiting the abundances of 22 immune cell subpopulations from Subtype 1 and Subtype 2. B Box plots exhibiting the differences in the relative abundance of infiltrated immune cell types between the two ferroptosis subtypes. C Box plots exhibiting the differences in the immune score between the two ferroptosis subtypes. D Box plots exhibiting the mRNA expression of immune checkpoints between two ferroptosis subtypes

Fig. 7

Construction of a protective model for prediction of LN patients. (A, B) Venn diagram (A) and heatmap (B) exhibiting 12 ferroptosis subtype-specific feature genes in LN patients. C A total of 9 characteristic genes were obtained using the using the Boruta feature selection algorithm. D The visualization of the neural network machine learning model. E Assessment of the classification performance of the neural network machine learning model in the validation cohort. F Representative nomogram based on NeuraLN and immune score for predicting LN progression. G Representative calibration curves for assessing the diagnostic accuracy of the nomogram

Fig. 8

Association between NeuraLN scoring model and immune microenvironment. A Heatmap exhibiting the expression profiles of the 9 model genes in the validation cohort. B Box plots exhibiting the differences in immune score between the low-and high-NeuraLN group. C Box plots exhibiting the differences in the relative abundances of infiltrated immune cell types between the low-and high-NeuraLN group. D Heatmap exhibiting the correlations between model genes and distinct immune cell compositions. *p < 0.05, **p < 0.01, ***p < 0.001

Fig. 9

Identification of distinct biological functions and signaling pathways between low-and high-NeuraLN group. A, B GSEA exhibiting distinct signaling pathways (A) and biological processes (B) between low-and high-NeuraLN group

Fig. 10

External validation of the alterations in model genes based on clinical specimens. A Representative immunohistochemistry images of the 9 model genes including ALB, BHMT, CUBN, DPYS, GAMT, GSTA1, HAO2, PAH, and SLC27A2. B Violin plots exhibiting the quantitative results of 9 model genes expression between controland LN patients. C Correlation analysis between the clinical characteristics and the expression of ALB, BHMT, GAMT, GSTA1, and HAO2 in LN glomerular tissues