Transcriptomic Responses to Ivacaftor and Prediction of Ivacaftor Clinical Responsiveness

Abstract

Ivacaftor is a drug that was recently approved by the U.S. Food and Drug Administration for the treatment of patients with cystic fibrosis (CF) and at least one copy of the G511D mutation in the CFTR (CF transmembrane conductance regulator) gene. The transcriptomic effect of ivacaftor in patients with CF remains unclear. Here, we sought to examine whether and how the transcriptome of patients is influenced by ivacaftor treatment, and to determine whether these data allow prediction of ivacaftor responsiveness. Our data originated from the G551D Observational Study (GOAL). We performed RNA sequencing (RNA-seq) on peripheral blood mononuclear cells (PBMCs) from 56 patients and compared the transcriptomic changes that occurred before and after ivacaftor treatment. We used consensus clustering to stratify patients into subgroups based on their clinical responses after treatment, and we determined differences between subgroups in baseline gene expression. A random forest model was built to predict ivacaftor responsiveness. We identified 239 genes (false discovery rate < 0.1) that were significantly influenced by ivacaftor in PBMCs. The functions of these genes relate to cell differentiation, microbial infection, inflammation, Toll-like receptor signaling, and metabolism. We classified patients into “good” and “moderate” responder groups based on their clinical response to ivacaftor. We identified a panel of signature genes and built a statistical model for predicting CFTR modulator responsiveness. Despite a limited sample size, adequate prediction performance was achieved with an accuracy of 0.92. In conclusion, for the first time, the present study demonstrates profound transcriptomic impacts of ivacaftor in PBMCs from patients with CF, and provides a pilot statistical model for predicting clinical responsiveness to ivacaftor before treatment.

Clinical Relevance

In the present study, for the first time, we examined changes in the transcriptomic profile of patients with cystic fibrosis after ivacaftor treatment, and developed a statistical method to predict a patient’s responsiveness to ivacaftor before treatment using the patient’s transcriptomic data and clinical measurements.

Cystic fibrosis (CF) is a genetic disorder caused by mutations in the CFTR (CF transmembrane conductance regulator) gene. Approximately 2,000 CFTR mutations are associated with CF, and a minority of these account for ∼90% of CFTR alleles (1). G551D-CFTR, represents 4% of CFTR disease causing and is a class III mutation characterized by apical localization but abnormal gating. Recently, ivacaftor was approved by the U.S. Food and Drug Administration for the treatment of patients with CF (age 6 and older) and at least one copy of the G511D mutation, as this drug has been shown to potentiate gating of the mutant channel. Phase 3 studies showed that the drug significantly reduced sweat chloride, improved CF symptoms, and improved lung function (forced expiratory volume in 1 second [FEV₁]) (2, 3). One previous study followed patients with CF before and after ivacaftor treatment and demonstrated marked improvement in the patients’ lung and gastrointestinal functions (4). To further understand the action mechanism of ivacaftor, as well as the adverse side effects of the treatment, there is a need to evaluate global gene expressions before and after ivacaftor treatment. In addition, we are interested in developing baseline gene expression signatures for the precise prediction of ivacaftor responsiveness.

To address these issues, we acquired archived biospecimens and clinical data from a multicenter clinical study (GOAL [G551D Observational Study]). We performed RNA sequencing (RNA-seq) experiments and transcriptomic analyses on peripheral blood mononuclear cells (PBMCs) obtained before and after ivacaftor treatment. Our analyses indicate that ivacaftor had profound transcriptomic impacts on the expression of a subset of genes in peripheral blood, and it is feasible to predict responsiveness to ivacaftor in patients with CF by evaluating their baseline PBMC gene signatures. Our data provide justification for further prospective studies to evaluate these types of predictive models.

Methods

PBMC Samples for this Study

Samples for RNA-seq analysis were archived by GOAL (4), a longitudinal cohort study that involves 28 centers in the Cystic Fibrosis Therapeutics Development Network (ClinicalTrials.gov Identifier NCT01521338). Samples of plasma, serum, buffy coat, urine, and expectorated sputum were collected from patients and stored at the Cystic Fibrosis Foundation Therapeutics Biorepository. Clinical assessments were performed at one or two baseline visits, and then at 1, 3, and 6 months after initial ivacaftor use. The buffy coat samples used in this study, which were essentially all of the available samples from the GOAL study, were obtained at one baseline visit and at 1 month after ivacaftor use. The participants or their guardians provided written informed consent, and the study was approved by the Institutional Review Board of the University of Pittsburgh.

RNA Extraction, RNA-Seq Library Construction, and Next-Generation Sequencing

The buffy coats were stored at −80°C, thawed on ice for 2 hours, and transferred into PAXgene tubes (PreAnalytiX). RNA was extracted using the PAXgene RNA extraction kit, and ∼500 ng total RNA per sample was used for library prep using the TruSeq RNA Access Library Prep Kit (Illumina, Inc.). Libraries were then quantified and normalized to 2 nM before they were sequenced on a NextSeq 500 sequencer using NextSeq 500 High Output Kit −150 cycles (Illumina, Inc.), generating ∼60 million paired-end reads for each sample.

Preprocessing of the RNA-Seq Data

Quality controls for raw fastq files were performed with FastQC (5). Low-quality reads and 3′ adapters were trimmed with Trim Galore! (6) and Cutadapt (7). Then, the trimmed reads were aligned to human genome reference (hg19) with the RNA-seq aligner STAR (8). Subsequently, the gene expression in each sample was quantified by counting the number of read fragments that were uniquely mapped to genes using the featureCounts (9).

Consensus Clustering of Patients Based on Responsiveness

Consensus clustering (10) is a statistical method that helps to detect underlying clusters within a dataset. We used the consensus clustering method (R package ConsensusClusterPlus) (11) to assign patients to “good” and “moderate” responder groups based on their clinical responses to ivacaftor, including FEV₁, body mass index (BMI), and Cystic Fibrosis Questionnaire-Revised (CFQR) respiratory scores. For more details on the use of consensus clustering, please refer to the data supplement.

Differential Expression Analysis

We performed a differential expression (DE) analysis using DESeq2 (12) on count data obtained at baseline and 1 month after treatment, adjusting for pair and batch effects. We conducted a similar DE analysis comparing baseline gene expression levels between the good and moderate responders identified by consensus clustering, adjusting for batch effect and age. The Benjamini-Hochberg procedure was applied to control for multiple comparisons and generate adjusted P values for each gene.

Pathway Enrichment Analysis

We performed a pathway enrichment analysis on the top genes from the DE analysis using the Ingenuity Pathway Analysis (IPA) pathway toolset (13) and our in-house code for Kyoto Encyclopedia of Genes and Genomes pathways (14). This analysis helped us obtain a regulatory picture of the underlying biological processes.

Statistical Modeling for Predicting Responsiveness

We built a random forest model (15) to predict whether a patient with CF would be a good responder to the drug based on his or her baseline clinical features. More details can be found in the data supplement.

Results

Identification of Differentially Expressed Genes between Patients at Baseline and 1 Month after Initial Use of Ivacaftor

In this study, we analyzed 93 RNA-seq samples from 56 subjects before and after treatment, as shown in the Venn graph in Figure 1A. These 56 subjects represent the entire GOAL population well. In the post-treatment patients who underwent RNA-seq in the GOAL study, we observed expected improvement in CF-related phenotypes, such as increases in FEV₁ and CFQR respiratory scores, and decreases in sweat chloride secretion, as shown in Table 1. Specifically, in our study subjects, we observed trends in the course of changes of the four most important clinical measures (mean FEV₁% predicted, mean BMI change from baseline, mean sweat chloride, and mean CFQR respiratory score change from baseline) (Figure E1 in the data supplement) similar to those observed in Reference 4, which used the entire GOAL study population. The underlying mechanism for the drug’s effectiveness is unclear. Thus, we first investigated whether there were any significant changes in the overall transcriptomic profile or specific gene expression levels after the treatment. A principal component analysis was performed based on the entire transcriptomic profiles of the 93 samples. Figure 1B presents a two-dimensional principal component analysis plot with distinct color indexes for pre- and post-drug samples. It clearly shows that the drug had little effect on the overall gene expression profile at Month 1, as the two groups of samples are evenly mixed. We then performed a DE analysis and successfully identified 102 genes (adjusted P < 0.05) and 239 genes (adjusted P < 0.1) whose expression was significantly changed from baseline to Month 1. The volcano plot in Figure 1C shows the –log₁₀ (adjusted P values) versus log₂ (fold change of post-drug samples vs. predrug samples) of all genes detected by the RNA-seq analysis. The differentially expressed genes (DEGs) are depicted in red circles, and the remaining non-DEGs are plotted as black circles. Figure 1D is a heat map illustrating the relative expression of each DEG. The DEG expression profiles look dramatically different before and after the treatment, although the hierarchical clustering did not perfectly separate the 93 samples in terms of ivacaftor treatment. The top 10 DEGs (ranked by adjusted P values) are listed in Table 2. Overall, this analysis suggests that although the overall transcriptomic profiles for these patients were not substantially altered after ivacaftor treatment, the expression of a subgroup of genes was associated with the drug treatment.

Figure 1.

RNA-sequencing (RNA-seq) analyses show that ivacaftor specifically affects expression in a group of genes after 1 month of drug use. (A) Venn diagram of the 93 RNA samples from 56 subjects in this study. Baseline RNA samples were obtained from 43 subjects and Month 1 RNA samples were obtained from 50 subjects. Both baseline and Month 1 RNA samples were obtained from 37 subjects. (B) Principal component analysis (PCA) demonstrates the variance of all genes expressed within each of the 93 RNA samples. The red color indicates samples obtained from patients at 1 month after ivacaftor use, and the blue color indicates samples obtained from patients at baseline. (C) A volcano plot shows the log₂ (fold change) and −log₁₀ (adjusted P value [Padj]) of all the genes compared between baseline and Month 1 samples. Statistically significant differentially expressed genes (DEGs) are represented in red. The statistical criterion for a gene to be considered differentially expressed was an adjusted P value ≤ 0.1. (D) A heat map illustrates the expression levels of the 239 DEGs as highlighted in C. Each column is a sample, and each row is a gene. The value is the z-score of the normalized gene expression count. For any given gene, the red color represents gene expression that is greater than the overall mean, and the blue color represents gene expression that is less than the overall mean. Hierarchical clustering of genes and samples are represented by the dendrogram to the left and across the top of the heat map.

Table 1.

Demographics and Cystic Fibrosis–related Characteristics of the Subjects at Baseline and Month 1

	Baseline (n = 43)	Month 1 (n = 50)	P Value
Female, n (%)	23 (53)	28 (56)	0.97
Age, mean (SD)	21.0 (12.0)	20.2 (9.9)	0.96
FEV1, mean (SD)	2.5 (1.0)	2.7 (1.1)	0.34
FEV1% predicted, mean (SD)	85.4 (26.6)	90.1 (24.5)	0.28
BMI, mean (SD)	21.4 (4.8)	21.6 (4.5)	0.78
BMI percentile, mean (SD)	53.8 (29.0)	58.7 (26.1)	0.78
Sweat chloride, mean (SD)	103.3 (9.1)	60.7 (22.4)	1 × 10−15
CFQR respiratory, mean (SD)	72.0 (15.9)	82.6 (12.1)	6 × 10−4

Definition of abbreviations: BMI = body mass index; CFQR = Cystic Fibrosis Questionnaire-Revised; FEV₁ = forced expiratory volume in 1 second.

Table 2.

List of the Top 10 Differentially Expressed Genes (Out of 239 Differentially Expressed Genes) between Subjects at Baseline and 1 Month 1

Gene Symbol	Gene Name	Chromosome	Fold Change	P Value
MARC1	Mitochondrial amidoxime reducing component 1	1	0.70	2.3 × 10−7
GYG1	Glycogenin 1	3	0.72	5.4 × 10−7
KLF9	Kruppel-like factor 9	9	1.30	7.9 × 10−7
LILRA6	Leukocyte immunoglobulin-like receptor, subfamily A (with TM domain), member 6	19	0.72	1.6 × 10−6
FUT7	Fucosyltransferase 7 (α-(1,3) fucosyltransferase)	9	0.70	1.7 × 10−6
PLOD1	Procollagen-lysine, 2-oxoglutarate 5-dioxygenase 1	1	0.78	1.8 × 10−6
ALPL	Alkaline phosphatase, liver/bone/kidney	1	0.68	2.2 × 10−6
TLR4	Toll-like receptor 4	9	0.70	2.3 × 10−6
PGD	Phosphogluconate dehydrogenase	1	0.76	2.6 × 10−6
DSC2	Desmocollin 2	18	0.70	2.8 × 10−6

Definition of abbreviation: TM = transmember.

All of the listed genes have adjusted P values < 0.01.

Validation of the RNA-Seq Results

To validate the data from RNA-seq, we performed real-time RT-PCR on selected top DEGs based on P values (Figure 2). The real-time RT-PCR data for MARC1 (with lowest P value) and TLR4 (presented in 9/19 pathways significantly affected by ivacaftor treatment as shown in Table 3) were consistent with our RNA-seq findings.

Figure 2.

Experimental validations of the two top DEGs ([A] MARC1 and [B] TLR4) from RNA-seq using RT-PCR. The samples were obtained from five patients at their baseline and 1 month post-treatment stages. The overall trends of RT-PCR of the two genes are consistent with our RNA-seq findings. One-sided paired t tests were performed on the pre-ivacaftor and post-ivacaftor samples for each gene. The P values are 0.064 and 0.040 for MARC1 and TLR4, respectively. MARC1 = mitochondrial amidoxime reducing component 1; TLR4 = Toll-like receptor 4.

Table 3.

Enrichment of Ingenuity Pathway Analysis Pathways (P < 0.001) for the Differentially Expressed Genes between Subjects at Baseline and 1 Month

Ingenuity Pathway Analysis Pathways	Number of DEGs (%)	−Log10 (P Values)	Genes
LXR/RXR activation	8	4.5	TLR4, LY96, CD14, SERPINA1, IRF3, HMGCR, MMP9, IL18RAP
Phagosome formation	8	4.2	FCAR, PIK3C2B, TLR4, CR1, ITGAM, RHOU, FCGR1A, FCGR1B
Osteoarthritis pathway	10	4.2	TLR4, TGFBR1, S100A9, RBPJ, NAMPT, TCF3, ALPL, MMP9, EP300, IL18RAP
Regulation of the epithelial–mesenchymal transition pathway	9	3.9	PIK3C2B, NOTCH2, TGFBR1, EGR1, RBPJ, TCF3, MAP2K1, MMP9, PSEN1
Erythropoietin signaling	6	3.8	SOCS1, PIK3C2B, FOS, CBL, EPOR, MAP2K1
Prolactin signaling	6	3.7	SOCS1, PIK3C2B, FOS, SOCS2, MAP2K1, EP300
Acute phase response signaling	8	3.5	SOCS1, FOS, HP, RBP7, SOCS2, SERPINA1, TCF3, MAP2K1
Role of macrophages, fibroblasts, and endothelial cells in rheumatoid arthritis	11	3.4	SOCS1, PIK3C2B, TLR4, FOS, TRAF4, TCF3, IL17RA, MAP2K1, FCGR1A, EP300, IL18RAP
Role of JAK2 in hormone-like cytokine signaling	4	3.4	SOCS1, EPOR, SOCS2, SIRPA
Pentose phosphate pathway (oxidative branch)	2	3.2	PGD, H6PD
Hepatic fibrosis/hepatic stellate cell activation	8	3.2	TLR4, LY96, TGFBR1, KLF6, CD14, MMP9, TIMP2, IL18RAP
Molecular mechanisms of cancer	12	3.2	PIK3C2B, FOS, CBL, TGFBR1, RHOU, ABL1, RBPJ, TCF3, MAP2K1, PSEN1, PRKAR1A, EP300
IGF-1 signaling	6	3.1	SOCS1, PIK3C2B, FOS, SOCS2, MAP2K1, PRKAR1A
Production of nitric oxide and reactive oxygen species in macrophages	8	3.1	PIK3C2B, TLR4, FOS, CYBA, RHOU, SERPINA1, MAP2K1, SIRPA
IL-8 signaling	8	3.0	PIK3C2B, FOS, ITGAM, RHOU, IQGAP1, MAP2K1, CSTB, MMP9
MIF regulation of innate immunity	4	3.0	TLR4, FOS, LY96, CD14
Colorectal cancer metastasis signaling	9	3.0	PIK3C2B, TLR4, FOS, TGFBR1, RHOU, TCF3, MAP2K1, MMP9, PRKAR1A
Granzyme A signaling	3	3.0	ANP32A, HMGB2, EP300
Toll-like receptor signaling	5	3.0	TLR4, FOS, LY96, TRAF4, CD14

Definition of abbreviations: DEGs = differentially expressed genes; IGF-1 = insulin-growth factor; JAK2 = Janus kinase 2; LXR/RXR = liver X receptor/retinoid X receptor; MIF = macrophage migration inhibitor factor.

The number and percentage of DEGs in each pathway, as well as the symbols of the DEGs, are listed in the table. The percentages of the DEGs are calculated from the number of DEGs in the specific pathway and the total number (239) of DEGs found between subjects at baseline and Month 1.

We also used one-sided paired t tests to compare the PCR data for samples obtained before and after ivacaftor treatment. The P values for MARC1 and TLR4 were 0.064 and 0.040, respectively.

Pathway Enrichment Analysis Using DEGs

To better understand the molecular functions and mechanisms of relevance among these DEGs, we performed a pathway enrichment analysis. An IPA pathway analysis was performed to detect molecular pathways significantly associated with the DEGs detected in the DE analysis. As shown in Table 3, 47 of out 239 DEGs were mapped to 19 molecular pathways (P < 0.001), which were associated with cell differentiation, microbial infection, inflammation, Toll-like receptor signaling, and metabolism. These associations likely contributed to the changes observed in the bacterial colonizations in these patients after ivacaftor treatment (4).

Clustering Patients into Good or Moderate Responders Based on Clinical Responses to Ivacaftor

We aimed to develop baseline gene expression signatures by RNA-seq for prediction of ivacaftor responsiveness. Instead of analyzing clinical responses individually, we stratified the 43 patients with baseline clinical and RNA-seq measures into good or moderate responders based on their multiple clinical responses using consensus clustering. As explained in the Methods section, we obtained 12 individual ranks for each subject based on their relative changes in the four most important clinical features from baseline to three post-treatment visits (Months 1, 3, and 6). In addition, we calculated the average rank for one subject in each feature by taking the average of individual ranks over the three visits. In total, each subject had 12 individual ranks and four average ranks. Figure 3A shows graphical presentations of consensus matrices obtained under four conditions: either hierarchical clustering or K-means based on either 12 individual or four averaged ranks. It is clearly shown that all four consensus clustering results led to two distinct groups. We noticed that the group assignments for the 43 subjects were similar but not identical among the four consensus clustering results. As shown in the Venn graphs of Figure 3B, we found that the four clustering methods identified 14 overlapping subjects in one group and another 24 overlapping subjects in the other group. The membership of the remaining five nonoverlapping subjects changed according to the specific clustering method used. Thus, for a conservative and robust analysis, we focused only on the 38 overlapping subjects. According to Table 4, the 14 overlapping subjects were significantly more responsive to the drug in terms of relative changes in CFQR respiratory score and FEV₁% predicted, than the other 24 subjects. Therefore, we defined the 14 subjects as good responders and the other 24 subjects as moderate responders.

Figure 3.

Patients were clustered into a “good” or “moderate” responder group based on their clinical responses to ivacaftor. (A) Graphical presentations of consensus matrices obtained under four conditions: hierarchical clustering based on 12 individual ranks of relative changes in clinical phenotypes (hc12), K-means based on the 12 individual ranks (km12), hierarchical clustering based on four ranks that were averaged over three visits (hc4), and K-means based on the four averaged ranks (km4). Each pixel represents the proportion in which the two corresponding subjects were clustered in one group out of the total number of times that both were sampled. The color intensity grows as the proportion increases. (B) Venn diagram depicting the distribution of the good responders (left) and the distribution of the moderate responders (right) under the four consensus clustering in A.

Table 4.

Relative Changes of Phenotypes in the 14 Good and 24 Moderate Responders

Relative changes	Good (n = 14)	Moderate (n = 24)	P Values
FEV1% predicted, mean (SD)
Month 1	17.80 (8.80)	1.83 (6.44)	6.7 × 10−6
Month 3	15.25 (8.36)	2.52 (6.68)	6.8 × 10−5
Month 6	18.42 (12.49)	1.31 (6.42)	1.8 × 10−4
BMI, mean (SD)
Month 1	1.90 (2.30)	1.09 (2.59)	0.33
Month 3	1.70 (3.34)	1.65 (3.81)	0.97
Month 6	1.33 (6.68)	3.66 (4.25)	0.25
Sweat chloride, mean (SD)
Month 1	−36.47 (21.73)	−50.43 (19.53)	0.06
Month 3	−45.74 (19.55)	−55.00 (15.53)	0.14
Month 6	−51.55 (17.42)	−58.94 (17.66)	0.22
CFQR resp, mean (SD)
Month 1	52.00 (50.22)	3.09 (12.51)	3.0 × 10−3
Month 3	52.92 (29.58)	5.48 (12.16)	3.4 × 10−5
Month 6	46.68 (38.46)	0.89 (18.83)	6.6 × 10−4

Identification of DEGs between the Good and Moderate Responders

The good and moderate responders were grouped purely based on relative changes in clinical phenotypes associated with CF. We next examined whether the two groups of subjects had DEGs before exposure to the drug after adjustment for the age effect. The DE results revealed 65 DEGs (adjusted P < 0.01) between the two groups (Figure 4). There are substantial differences in expression profiles between the two groups. The top 10 DEGs (ranked by adjusted P values) are listed in Table 5.

Figure 4.

RNA-seq analyses demonstrate significant differences in gene expression between good and moderate responders. (A) A volcano plot shows the log₂ (fold change) and −log₁₀ (Padj) of all the genes compared between good and moderate responders at the baseline. The statistically significant DEGs are represented in red. The statistical criterion for a gene to be considered differentially expressed was an adjusted P value ≤ 0.01. (B) A heat map illustrates the expression levels of the top 65 DEGs between the good and moderate responders. The setting is the same as in Figure 1D.

Table 5.

List of the Top 10 Differentially Expressed Genes (Out of 65 Differentially Expressed Genes) between Good and Moderate Responders

Gene Symbol	Gene Name	Chromosome	Fold Change	P Values
SLC1A3	Solute carrier family 1, member 3	5	2.59	2.9 × 10−8
FOXF2	Forkhead box F2	6	2.81	4.5 × 10−8
MCU	Mitochondrial calcium uniporter	10	1.46	2.9 × 10−7
MCEMP1	Mast cell–expressed membrane protein 1	19	2.46	3.0 × 10−7
ZNF438	Zinc finger protein 438	10	2.03	6.6 × 10−7
ITGB4	Integrin, beta 4	17	2.51	6.9 × 10−7
ANXA3	Annexin A3	4	2.47	7.9 × 10−7
GCOM1	GRINL1A complex locus 1	15	2.48	8.5 × 10−7
GALNT14	Polypeptide N-acetylgalactosaminyltransferase 14	2	2.50	1.1 × 10−6
ZDHHC19	Zinc finger, DHHC-type containing 19	3	2.44	1.1 × 10−6

Definition of abbreviation: GRINL1A = glutamate receptor-like protein 1A.

All of the listed genes have adjusted P values < 0.01. Fold change is comparing good responders with moderate responders.

Moreover, we plotted a heat map in Figure E2 for a list of CFTR modifier genes (16–19) that were found among the DEGs (adjusted P < 0.1). There is an apparent difference in the general expression profiles of these modifier genes between the good and moderate responders (Table E4). We also performed a gene set enrichment analysis (20) for the identified CFTR modifier gene set and obtained a positive normalized enrichment score (1.44) with a false discovery rate of 0.01, as illustrated in Figure E4. Similarly, we performed a rotation gene set test analysis (21) for this gene set and obtained a P value of 0.03. Both studies suggest that the genes were differentially expressed between the good and moderate responders.

Predicting Good or Moderate Responders

We built a random forest model to predict whether a patient would be a good or moderate responder based on the baseline features of the patient. Due to the limited sample size, we only included the relevant clinical features (i.e., FEV₁% predicted, BMI, sweat chloride, and CFQR respiratory score) and the fragments per kilobase of transcript per million mapped reads (FPKM) value of one selected gene, PPARG, which ranks high among the significant CFTR modifier genes listed in Table E1, and also appears in the top two IPA pathways in Table E6. It was previously recognized as a potential therapeutic target for treating CF (22). Figure 5A is a variable-importance plot showing the magnitude of importance of each predictor in the model, with the x-axis representing the mean decrease in the Gini index. The Gini index is a measure of total variance across two responder classes. For each predictor, we added up the total amount that the Gini index was decreased by splits over this specific predictor, and averaged the total by the number of trees involved in the random forest model. The mean decrease in the Gini index is large when the predictor plays a major role in prediction. This analysis shows that the CFQR respiratory score, PPARG (the mean FPKM levels in good and moderate responders were 0.302 and 0.112, respectively), and FEV₁% predicted values at baseline were the three most important predictors. We performed a 10-fold cross-validation and obtained an average test error of 0.108 with an SE of 0.030. The receiver operating characteristic curve of the prediction model is shown in Figure 5B, with an area under the curve of 0.95.

Figure 5.

A random forest model for predicting good or moderate responders. (A) Graphical representation of the variable importance in the random forest model to predict good or moderate responders. It depicts the mean decrease in the Gini index for each predictor relative to the largest. The Gini index is a measure of total variance across two responder classes. A smaller value of the Gini index indicates that the subjects classified in one cluster are predominantly from the same class. For each predictor, we added up the total amount that the Gini index was decreased by splits over this specific predictor, and averaged the total by the number of trees involved in the random forest model. The most important predictors were the Cystic Fibrosis Questionnaire-Revised (CFQR) respiratory score, PPARG gene expression, and forced expiratory volume in 1 second (FEV₁)% predicted value. (B) Receiver operating characteristic (ROC) curve for the random forest classifier of good or moderate responders based on predictors in A. The true-positive rate is the sensitivity: the fraction of responders who are correctly identified. The false-positive rate is 1-specificity: the fraction of nonresponders who are classified incorrectly as responders. The ideal ROC curve hugs the top-left corner, indicating a high true-positive rate and a low false-positive rate. The area under the curve is 0.95. BMI = body mass index; PPARG = peroxisome proliferator activated receptor-γ.

We also examined the contribution of PPARG to prediction accuracy (i.e., the proportion of true-positive and true-negative predictions out of all predicted samples), as shown in Table 6. With only clinical baseline features as predictors, the accuracy is 0.82. When we include PPARG, the accuracy increases to 0.92. In addition, under the baseline clinical model, the average 10-fold cross-validation test error is 0.183 with an SE of 0.048, and its area under the curve is 0.87. Therefore, PPARG improved the prediction results in terms of accuracy and test error.

Table 6.

Contingency Table for Predicted and True Classifications for Good and Moderate Responders

	Clinical		Clinical + PPARG
	Predicted Positive	Predicted Negative	Predicted Positive	Predicted Negative
True positive	22	2	24	0
True negative	5	9	3	11

Definition of abbreviation: PPARG = peroxisome proliferator activated receptor-γ.

Two predicted models are shown: the “clinical” model contains only baseline clinical features as predictors, and the “clinical + PPARG” model includes both clinical features and fragments per kilobase of transcript per million mapped reads values of PPARG at baseline.

We further examined the random forest classification performances for the genes listed in Table E5 (including PPARG) and found that the model using PPAGR (together with the clinical features) was the best model with the lowest prediction errors. The results are summarized in Table E6. We included only PPARG (without adding more genes) in our final random forest classification model in part because we wanted to use as few predictors as possible to prevent model overfitting, especially given the small sample size. On the other hand, we found that the random forest model with clinical features only (with a test error of 0.183) performed better than many models, including a gene predictor. In fact, only four genes (PPARG, S100A8, SLC1A3, and NOS1) presented smaller test errors (0.108, 0.162, 0.170, and 0.150, respectively) than 0.183. We further examined the performance of random forest model incorporating these four genes, and found that the test error was 0.108, which is the same as the test error obtained using PPARG alone. Therefore, we included only PPARG in our final classification model.

In addition, we also compared the prediction performances of the random forest and logistic regression models using the genes from Table E5. The random forest model with PPARG still outperformed all logistic regression models. The results are presented in Table E7.

Discussion

For the first time, this paper reveals the specific transcriptomic responses in PBMCs after initiation of ivacaftor treatment in patients with CF and a G511D mutation. The results suggest that ivacaftor improves symptoms, possibly by influencing cell differentiation, microbial infection, inflammation, Toll-like receptor signaling, and metabolism. Due to the high cost of the drug throughout a patient’s life (approximately $300,000/yr), there is an urgent need to help patients with CF decide whether the drug is effective or not. Also for the first time, we identified good and moderate responders to ivacaftor, and discovered signature gene expressions that distinguish the two types of patients with CF. Lastly, we built a random forest statistical model to predict whether a new patient with CF would be a good responder based on his or her characteristics before treatment.

The goal of personalized medicine is to customize healthcare, including medical decisions and practices, drugs, and treatments, for each patient. In addition to facilitating more precise and effective medication use, this model also has the potential to reduce costs associated with healthcare. Genetic and genomic markers have been recently identified and included in drug labeling with great impact. For example, abacavir, a drug for treating HIV, causes hypersensitivity, which can be fatal, in ∼5–8% of subjects among populations of European descent (23, 24). This adverse reaction was found to be associated with HLA*B5701 carriage, and studies showed that genetic screening before treatment improved the safety (23, 24). Ivacaftor was recently approved by the U.S. Food and Drug Administration for the treatment of patients with CF and at least one copy of the G511D mutation. However, methods are needed to evaluate CFTR functional activity and assess new therapies. Precision medical care for patients with CF also requires a predictor of responsiveness to ivacaftor. One of our main goals in this study was to develop ivacaftor responders’ gene signatures as predictors that could lead to the development of clinical diagnostics to accelerate personalized medicine in the future. However, when we performed an association analysis of the RNA-seq data with individual clinical phenotypes, we found that no signatures were correlated with a specific outcome measure (e.g., sweat chloride). A possible explanation for this is the inadequate association between gene signatures and a single phenotypical feature. Thus, when we clustered the subjects by four different phenotypes, we were able to detect gene signatures associated with clinical outcomes. With a larger sample size in the future, we could screen more good and moderate responders for potential gene signatures, and even identify signatures associated with a specific clinical outcome.

One caveat of our study is that we examined the PBMC transcriptomic responses 1 month after the patients received ivacaftor. Thus, the transcriptomic changes we saw may or may not be a direct effect of ivacaftor on PBMCs. This is supported by the finding that the PBMCs expressed very low levels of CFTR (FPKM = 0.06). In addition, we believe that the changes likely were not specific to CFTR modulation itself because we did not observe any gene expression changes that significantly correlated with sweat chloride changes. Indeed, one major action of the drug was to improve the lung function of these patients, and the improved lung function could have had a secondary impact on the PBMC transcriptome. However, when we looked at the correlation between lung function and the top 20 DEGs’ FPKM values, we found that the FEV₁% predicted values were not significantly associated with 18 out of 20 DEGs (adjusted P value > 0.05) by running a mixed-effects model (which accounts for the pair effect) between FEV₁% predicted values and each DEG expression. This suggests that the PBMC transcriptomic changes were not entirely driven by the improvement in the patients’ lung function.

We noticed that the moderate responders had better clinical features than the good responders at baseline, corresponding to the gene expression profile differences (Tables 5 and E4) between the two groups. In particular, two DEGs, SLC26A9 and SLC6A14, were previously identified as the top two most significant genes associated with meconium ileus in a large-scale genome-wide association study of patients with CF (25). After treatment with ivacaftor for 1 month, we could not find any significant differences in gene expression profiles between the two cohorts. Also, the CF-related clinical features between the two groups were similar (Figure E3). This suggests that ivacaftor effectively improved both the transcriptomic and clinical profiles in the good responders.

To the best of our knowledge, our samples are the only available PBMC samples from ivacaftor-treated patients with CF before and after treatment. Furthermore, these samples are the only samples available from the Cystic Fibrosis Foundation Biorepository. This is the first study to use these samples and report significant transcriptomic changes in patients with CF undergoing ivacaftor treatment. Our RNA-seq data are available from the Gene Expression Omnibus (accession number GSE128723), and will serve as an independent and external cohort for other studies in the future.

The off-target effects of CFTR modulators have not been thoroughly investigated. In fact, a decrease in amiloride responses in human bronchial epithelial cells due to chronic ivacaftor treatment was observed recently (26), and ivacaftor has been shown to have antimicrobial properties in vitro (27), suggesting that it may display off-target effects in vivo. A possible explanation for the antimicrobial effect is that ivacaftor structurally resembles quinolone antibiotics. However, most fluoroquinolone derivatives have immunomodulatory effects, such as inducing IL-2 synthesis while inhibiting the synthesis of IL-1 and TNF-α (28), which suggests that ivacaftor may have immunomodulatory functions. Although most research on CF pathogenesis has focused on epithelia, recent work has shown that CFTR is expressed on human and murine monocytes and macrophages, and that CFTR deficiency could alter these cells’ functions (29, 30). Thus, it is possible that these CFTR modulators also have an impact on immune cell functions. Indeed, our data suggest that ivacaftor-sensitive biological pathways/biomarkers can be identified in patients’ peripheral blood after treatment. With the development of RNA-seq technology, including single-cell RNA-seq in the future, both global and cell-type–specific effects of CFTR modulators can be uncovered by transcriptomic and epigenomic analyses in peripheral blood from patients before and after treatment.