ELAV-Like Protein 2

Network analysis was performed in Banner and BLSA datasets separately to identify modules of co-expressed proteins. For the Banner network, missing proteins were imputed using the k-nearest neighbor imputation function in R. Then, batch effects were removed using Combat^{64 (link)}, and age at death, sex, and PMI were regressed from the proteomic profiles using Bootstrap regression. Weighted gene co-expression network analysis (WGCNA)^{68 (link)} was used on normalized protein abundance to define protein co-expression networks. For BLSA, we used the BLSA networks from Seyfried et al.^{23 (link)}, which were previously built using proteins measured from the precuneus and prefrontal cortex in the same individual. We defined hub proteins, i.e., highly connected proteins, for each of the modules as those with intramodular kME in the top 90th percentile among the proteins in the corresponding module^{68 (link)}. Gene ontology (GO) enrichment analysis was performed on each protein co-expression module using GO Elite and Fisher exact test^{69 (link)} to glean a deeper biological understanding of these modules.

Raw data are available in the UNC Dataverse (https://doi.org/10.15139/S3/XHHUFB).
Mean and standard deviation (SD) were calculated for age and BMI. We used ANOVA and post hoc Dunnet's test to detect differences between groups. Relative percentages were calculated for sex, race, ethnicity, and smoking status. Any nasal immune mediators that were below the limit of detection by ELISA were set to half the minimum detected value for all analyses. We used ANOVA to detect differences in the mean levels of each nasal immune mediator (ng/ml) across the four cohorts. We reported unadjusted p‐values and p‐values adjusted with the Benjamini and Hochberg method to control the false discovery rate (FDR).
For mediators found to differ among groups by ANOVA, we determined the percent difference in average mediator concentration for each cohort compared to the healthy nonsmoking cohort. For these comparisons, the concentrations (ng/ml) were log‐transformed and we used Dunnett's test to compensate for multiple comparisons. We also ran multinomial logistic regression models to assess the association between the concentration for these mediators in the nasal fluid and the odds of being in the COPD, CRS, and smoking cohort compared to being in the healthy cohort. We ran unadjusted models and additionally adjusted models for sex. Due to lack of overlap between ages of the elderly COPD group and the other groups, we were unable to control for age statistically in this study.
In order to assess if nasal inflammatory mediators differed by anti‐inflammatory medication or respiratory symptoms, we ran sensitivity analyses within the COPD group to investigate differences in mediator concentrations by inhaled corticosteroid (ICS) use and more symptom variability (defined as ≥10 days of worse‐than‐baseline respiratory symptoms over 4 months of observation) using t‐tests (Alvarez‐Baumgartner et al., 2022 ). These sensitivity analyses were possible only in the COPD subgroup because ICS use was an exclusion criterion in the healthy and smoking groups, and data on ICS use and respiratory symptoms were only collected in the group with COPD.
In order to identify modules (clusters) of correlated mediators and their correlation with airway disease and smoking status, we conducted a weighted gene co‐expression network analysis (WGCNA) using the R4.1.1 WGCNA package (Langfelder & Horvath, 2008 (link)). A soft‐power threshold of 6 was chosen after determining that the scale‐free topology fit index curve flattened at 6. Modules were constructed with a merging threshold of 0.25 and minimum module size of 2, resulting in six modules. We evaluated the association of each cluster with cohort membership and display these results graphically with a heat map. Finally, we identified the specific mediators driving the clusters (hub proteins), defined as having a high cohort significance (correlation between healthy cohort membership and mediator concentration >0.2) and having strong module membership (correlation of module eigengene and gene expression profile >0.8), based on (Langfelder & Horvath, 2008 (link)).

The symbols for c2.cp.kegg.v7.4 were obtained from the Molecular Signatures Database (MSigDB, https://www.gsea-msigdb.org/gsea/index.jsp). To investigate the deferentially expressed pathways between high- and low-risk groups, the “GSVA” R software was utilized. “GOplot” and “enrichplot” R programs were used to visualize the findings of the Gene Ontology (GO) analysis, which was conducted via the R package “org.Hs.eg.db” [20 (link)]. An analysis of the protein-protein interaction (PPI) network was published on the Search Instrument for the Recovery of Interacting Genes (STRING) v11.0 website (https://cn.string-db.org/). A protein-protein interaction (PPI) analysis was conducted via the Cytoscape v3.7.2 software, and hub genes were screened via the “cytohubba” application. In order to assess these genes, twelve factors will be considered. The genes with the greatest degree of value were determined as hub genes. A survival analysis and a single sample gene set enrichment analysis (ssGSEA) were conducted on the gene with the highest frequency. As part of the Human Protein Atlas (HPA), immunohistochemistry was utilized to confirm the hub gene signature's expression of proteins (https://www.proteinatlas.org).