Unsupervised gene expression analyses identify IPF-severity correlated signatures, associated genes and biomarkers

We used the microarray data from the IPF cohort in the Lung Genomics Research Consortium’s (LGRC) website (http://​www.​lung-genomics.​org; also deposited in data repository GEO - GSE47460 [11]). Among 582 subjects in dataset GSE47460, 12 had clinical and pathological designations as “controls”, and 131 had clinical and pathological diagnoses of “UIP/IPF”. We selected these 143 subjects for our cluster analysis, differential analysis, and to train classifiers. Demographic and clinical characteristics of the selected cohort are summarized in Table 1. There was no statistically significant difference in age between control and IPF patients, but there were more males in the IPF group. The predicted forced expiratory volume in one second (FEV1), forced vital capacity (FVC), and diffusing capacity of the lungs for carbon monoxide (D_LCO) were significantly lower in UIP/IPF patients compared to those of the control group. For evaluating the classifier performance and assessing the relevance of our identified IPF sub types, we used three independent IPF cohorts (GSE24206 [8], GSE10667 [9] and GSE53845 [1]).

We used the Scikit-learn [12] package in Python for clustering analysis and PCA, and the limma [13] package in R for differential analysis. Data was first preprocessed by aggregating redundant transcript, log2-tranformed and median normalized across each gene, resulting an expression data matrix of 14,110 genes by 143 samples (or subjects). For PCA, the principal components were calculated using only expression data containing only IPF samples, and expression data from control and IPF patients were projected onto these principal components. Then, hierarchical clustering using the Ward linkage method on Euclidean space was performed on the transformed matrix. The number of clusters was chosen as the smallest number that allowed maximal difference in average FEV1, FVC, and D_LCO values among clusters. Following clustering, differential analysis was performed across IPF clusters and control with Benjamini–Hochberg false discovery rate (FDR) correction. Differentially expressed genes (DEG) were defined as those with FDR-adjusted p-value ≤0.05 and absolute log2 fold-change ≥1 when compared to control.

We used the Scikit-learn package in Python to build and evaluate logistic regression classifiers to evaluate classification power of each IPF gene set. The datasets from the training and validation cohort were median normalized and scaled to (0, 1) across each gene. In order to assess classifier accuracy and reduce over-fitting, we included a 2-fold cross-validation step before training the final classifier using all samples in the training cohort. Then, classifiers were used to predict IPF status in each validation cohort. Receiver-Operating-Characteristic (ROC) curve, overall accuracy, sensitivity and specificity were used to evaluate classifier performance.

We used ToppFun of the ToppGene Suite [14] for functional enrichment analysis and ToppGene for candidate gene prioritization. For candidate gene prioritization, we used ‘GO: Biological Process’, ‘GO: Cellular Component’, ‘Human Phenotype’, ‘Mouse Phenotype’, ‘Pathway’ and ‘Disease’ as features to compute the similarity. The significance threshold was set as FDR-adjusted p-value ≤0.05. We used “known” IPF genes from the Orphanet [15] and DisGenet [16] databases as training sets to rank the differentially expressed genes in IPF and identify and prioritize novel candidate genes for IPF.

To examine genes associated with UIP/IPF, we first performed differential gene expression analysis comparing 131 UIP/IPF patients with 12 control subjects and identified 988 differentially expressed genes. However, among these genes there were several distinctive gene expression patterns that defined distinct subsets of IPF patients (Additional file 1: Fig. S1a). To determine whether this molecular heterogeneity correlated with disease severity, we grouped UIP/IPF patients based on their available clinical FVC and D_LCO measurements (FVC ≥ 55% or D_LCO ≥ 40%: mild-to-moderate IPF; otherwise: severe IPF) and repeated the differential analysis comparing each of the phenotype-based UIP/IPF sub-groups with the control group. This resulted in 1175 and 1167 DEGs in IPF patients grouped by D_LCO and FVC measurements, respectively. Surprisingly, we observed that even among UIP/IPF patients within the same FVC or D_LCO sub group, expression of the genes was still highly variable (Additional file 1: Fig. S1b and c). This finding was further corroborated with results from principal component analysis (PCA) of all patient samples using 3657 (first quartile) most variable genes, wherein FVC or D_LCO subgroups could not be separated from each other by the first two principal components (Additional file 2: Fig. S2a and b). The heterogeneity within the gene expression profile of UIP/IPF patients and their poor concordance with markers of IPF severity motivated us to take an unbiased approach of clustering the IPF patients based on their gene expression profile to detect (a) potential IPF subgroups and (b) identify DEGs that correlate with lung function.

We performed ward clustering followed by PCA on the gene expression profiles of 131 UIP/IPF patients, and identified 6 distinct patient clusters (C1 through C6) (Additional file 2: Fig. S2c). Subgroups of differential IPF severity, as reflected by the average of clinical measures (D_LCO, FEV1, and FVC), were arranged in descending order from patient clusters C6 to C1. D_LCO values in patient clusters C5 orC6 were significantly lower than those in C1, C2, C3, or C4, and significantly higher in patient clusters C1 or C2 than those in C3, C4, C5, or C6. On the other hand, FVC and D_LCO values did not differ significantly between C1 and C2, C3 and C4, or between C5 and C6 (Fig. 1a–c ; Additional file 3: Table S1). These results suggested that the patient subgroups C1 and C2 had modest changes while C5 and C6 had a significant decline in lung function compared to control. Whereas patient subgroups of C3 and C4 exhibited intermediate changes in their lung function compared to those with mild (C1 and C2) and severe (C5 and C6) disease phenotypes based on lung function tests.

To examine transcriptomic differences between these patient clusters, we performed differential analysis comparing IPF patient clusters with control using the R package ‘limma’ and identified 2968 DEGs (Additional file 4: Table S2). Interestingly, these DEGs included nearly all DEGs identified by earlier grouping methods using (a) pooled IPF patients; (b) patients grouped based on D_LCO measurements; and (c) patients grouped based on FVC measurements (Additional file 5: Fig. S3) and expression of these 2968 DEGs within each of the six patient clusters was more homogenous compared to earlier methods.

Substantial differences in gene expression profiles were found between patient clusters that had similar disease severity, namely C1 vs. C2, C3 vs. C4 and C5 vs. C6 (Fig. 1d ). Three major gene expression modules (Gm) can be identified from the total 2968 DEGs. Gm1 was up-regulated in patient cluster C1, C3, C4, C6 and moderately in C5, Gm2 was up-regulated in patient cluster C1, C3, C5 and C6, while Gm3 was down-regulated in patient cluster C3, C5, C6 and moderately in C1 and C4. To gain functional insight into these genes, we performed enrichment analysis on them and found Gm1 was enriched in processes such as ‘extracellular matrix organization’, ‘regulation of cell migration’ and ‘collagen catabolic process’, Gm2 was enriched in processes such as ‘cilium’ and ‘cilium assembly’ and Gm3 was enriched in ‘angiogenesis’ and ‘lung alveolar morphology’. Taken together, these results show that UIP/IPF patient subgroups stratified by disease severity can be distinguished using gene expression profiles-based clustering, which in turn reveals the involvement of different molecular pathways in the pathogenesis and severity of fibrotic lung disease.

The number of DEG found in each of the patients’ clusters ranged from 262 genes (patient cluster C2) to 2117 genes (C5). About 34% of the total DEGs were unique to one patient subgroup while the remaining were found to be overlapping with the others (Fig. 2a and b ). All IPF subgroups shared a set of 145 DEGs, which were named the IPF core gene set. Among them, genes involved in ‘proteinaceous extracellular matrix’ and ‘regulation of epithelial to mesenchymal transition (EMT)’ were up-regulated, while hemoglobin genes such as HBA1, HBA2, HBD, HBG1 and HBQ1 were down-regulated (Fig. 3 ). The most severe patient subgroups C5 andC6 shared DEGs in C1 and C3, which were enriched in processes including ‘mitotic nuclear division’, ‘cilium assembly’, ‘epithelial/endothelial migration’ and ‘tube development’. Patient clusters C5 andC6 also uniquely expressed 840 DEG, with 448 genes in C5 and 392 genes inC6. Subsets of C5 unique genes were enriched in pathways such as ‘cilium assembly’ and ‘tube development’, which were also perturbed in less-severe IPF subgroups C1 and C3. This suggests a potential IPF progression path of C1➔C3➔C5 characterized by increasing expression of cilium-associated genes, and is in consistent with a previous study that reported a positive correlation of cilium-associated gene expression and increased IPF severity [10]. TheC6-specific gene set included inflammatory response genes such as HMOX1, IL1R1, IL20RB, IL36G, SELE, SERPINF2, TNFRSF21 and TNFRSF6B, but not genes enriched in cilium-associated pathways (Fig. 3 ). This suggests IPF can alternatively progress via up-regulating inflammation genes without further up-regulation of cilium-associated genes, and is consistent with a recent report showing increased inflammation in rapid progressive IPF [17].

To further validate and assess the relevance of our identified IPF subgroups, we used three independent IPF cohorts (GSE24206, GSE10667 and GSE53845). By utilizing multiple testing datasets, we determined if the gene sets reveal key differences in gene expression between IPF and normal, and between severe IPF (explant) and usual IPF (biopsy).

Out of 145 genes in the core IPF gene set (Additional file 6: Fig. S4), 133 were found in all three validation cohorts. We used the LGRC dataset with 12 controls and 131 IPF patients as the training set, and trained a logistic regression classifier for classification of IPF patients. Then, IPF status (normal or IPF) was predicted by using the classifier on each validation dataset, where the decision threshold was set to provide at least 90% sensitivity. The ROC curve, sensitivity, specificity, and overall accuracy in each of the validation datasets are shown in Fig. 4 . Specificity, sensitivity and accuracy in all three validation datasets were >90%. Similarly, we validated the C5 andC6 unique gene sets (Additional file 7: Fig. S5). The C5 unique gene set (448 genes) performed poorly in differentiating severe IPF from usual IPF in all validation datasets (data not shown). Hence, this gene set was not considered for further analysis. However, although classifiers built on theC6 unique gene set failed to distinguish AEIPF from IPF, they could moderately differentiate IPF explant from IPF biopsy, indicating the unique gene expression profile of patient clusterC6 were also present in severe, end-stage IPF. Taken together, these results demonstrate that the core IPF gene set is a robust gene signature to separate IPF from control. The advanced IPF gene set (c6 unique gene set) on the other hand can differentiate advanced IPF, but not AEIPF, from stable IPF.

To identify genes that were most functionally relevant to biologic processes perturbed in IPF, we employed a systems biology approach to rank each gene in the core and advanced gene sets based on their functional similarity to one of the two training sets comprising genes known or implicated to be involved in IPF [15, 16] (Additional file 8: Table S3) and their gene expression fold change in IPF compared to control. Specifically, functional similarity was calculated using the ToppGene Suite’s gene prioritization tool [14]. Genes in the core and advanced IPF gene sets that were also in the training set (“known” IPF genes) were removed from ranking. The remaining 133 and 382 genes from the core and advanced gene sets respectively were then ranked separately based on either similarity score or fold change compared with normal. The rankings of each of the genes were aggregated using the rank product method [18]. Top 10% ranked genes in the advanced and core IPF gene set are shown in Table 2. Twenty-two of these genes had been shown to be differentially expressed in IPF patients compared with healthy control or involved in IPF pathogenesis. Enrichment analysis of the novel candidates showed these genes were involved in pathways often perturbed in IPF. For example, up-regulated genes in the advanced set such as SERPINF2, MMP14, DMP1 and CTSL, were enriched in ‘extracellular matrix organization’. On the other hand, genes involved in leukocyte activation such as BLM, RAG1, PRKCZ, LBP, and MMP14 were only present in the advanced gene set.

Among the genes in the core IPF gene set, 60 of them encode secreted proteins (based on Uniprot [19] annotation) or were previously found in bronchoalveolar lavage fluid (BALF). Given these genes’ classification power and their potential clinical utility, we decided to prioritize candidate IPF BALF biomarkers among them. We first ranked these genes based on magnitude of the coefficients from a logistic regression model. Then, we built a series of logistic regression models, each trained on up to 50 top ranked genes, to determine the threshold for marker selection (Additional file 9: Fig. S6). In the end, we identified 11 putative biomarkers, including HMGCS2, CHL1, DAO, CRTAC1, EDN1, WNT10A, HBEGF, IL6, CCK, EPHA3 and SEMA3E, in the core IPF gene set (details in Additional file 10: Table S4) which is the smallest biomarker set that allowed >0.8 specificity and 0.9 sensitivity in distinguishing IPF from healthy control. Notably, three core IPF markers, HMGCS2, CHL1 and SEMA3E, were also differentially expressed compared with control in BALF of bleomycin-treated mouse and their direction of dysregulation was consistent with our study [20].