Data Collection
This study was reviewed by the Vanderbilt University Medical Center’s (VUMC) Institutional Review Board and determined to be non-human subjects research (IRB # 180498). All methods were carried out in accordance with relevant guidelines and regulations. The cohort was derived from BioVU, a clinical practice-based biobank at Vanderbilt University Medical Center, a tertiary care center. BioVU uses de-identified electronic health records (EHRs)—including demographics, clinical notes, medical history, problem lists, medications, and diagnostic and procedure codes—which are linked with stored DNA samples.12,13 Informed consent to donate leftover blood samples from routine lab testing to BioVU for research was obtained from all subjects and/or their legal guardian(s) at time of lab collection.
For this retrospective cohort, we identified new users of azathioprine with at least one ICD9 or ICD10 code in their EHR during follow-up, as described below. We excluded patients whose genetic data did not pass quality control or whose primary indication for azathioprine was an organ transplant. Given that PrediXcan has been validated primarily in patients of European ancestry,14 we restricted the cohort to patients with EHR-reported White race; in previous work, we determined EHR-reported White race was highly concordant with European ancestry among new users of azathioprine in BioVU (i.e., >99% of patients identified as having White race had predominantly European ancestry).15
Chart review was completed to confirm azathioprine use and gather covariates. Patients entered the cohort on the first day of new azathioprine use, defined as no previous mention of azathioprine or mercaptopurine use in the EHR. Follow-up ended on the first of the following: 1) day of azathioprine discontinuation; 2) last confirmed azathioprine prescription or use as per the EHR plus 90 days; (3) lost to follow-up; (4) day of death; or (5) end of the study (December 31, 2018). We abstracted additional data from the EHR such as reported race, sex, age, initial daily dose of azathioprine, and indication.
As previously described,4 we genotyped patients using the Illumina Infinium Expanded Multi-Ethnic Genotyping Array (MEGA)EX platform, imputed the results using Michigan imputation servers,16 and applied standard genetic quality controls. We then used PrediXcan to estimate the expression of TPMT and NUDT15 in liver tissue; we pre-specified liver tissue as the most appropriate target given that most of azathioprine metabolism is hepatic.17 We first assessed whether the cohort had sufficient variability in TPMT and NUDT15 to deploy PrediXcan successfully and proceeded with the remaining steps accordingly. PrediXcan predicts gene expression in tissues by combining MASHR v8 GTEx weights with estimated dosages from the Michigan Imputation Server. We then confirmed that the MASHR v8 SNPs did not replicate SNPs used in CPIC guidelines nor were they in linkage disequilibrium. Given that the Michigan Imputation Server estimates dosage by summing the posterior probabilities of having the alternate allele in the haplotype, we anticipated that the predicted liver expression would have clusters of values.8,18,19
Using clinical pharmacology databases (e.g., Micromedex), the Food and Drug Administration package insert for azathioprine, and peer-reviewed publications, we compiled a non-exhaustive list of nine common types of adverse effects associated with azathioprine use: infections, gastrointestinal intolerance, hematologic toxicity, dermatologic symptoms, skin malignancies, non-skin malignancies, hepatotoxicity, constitutional symptoms, and pulmonary symptoms.20-22 We then identified 51 phecodes, a method for assigning phenotypes using ICD9 and ICD10 diagnosis codes in claims data, corresponding to these nine adverse event types (Table S4).23 We gathered ICD9 and ICD10 diagnosis codes for each patient during follow-up and designated patients with ≥2 non-consecutive instances of an adverse event type phecode as potential cases. After assessing whether any of these phecode groups were associated with TPMT or NUDT15 liver expression (described further below), we reviewed EHRs for adverse event types that had a pre-specified p<0.10, using clinical notes to confirm that the adverse effect could be reasonably attributed to azathioprine. We reviewed all potential cases for leukopenia, rash, and non-basal cell skin cancer. Leukopenia was defined as white blood cell count <4.0 K/μL, which was not be attributed to other causes. Cases of rash were limited to those with an ICD code for rash which was not be attributed to other causes (e.g., indication for azathioprine). Skin cancer cases were biopsy-proven or reported in dermatology documentation. Review was completed by a clinician using a standardized adjudication form (Figure S3).24,25 Ambiguous cases were subject to final adjudication by the senior author. All adjudications were completed blinded to genotype and predicted expression. We did not review charts for potential cases not identified by a phecode.
Statistical Analysis
For demographic and clinical characteristics, we present categorical variables as numbers and percentages, and continuous variables as means and standardized differences. We used Chi-squared tests to compare categorical variables and Kruskal-Wallis tests for continuous variables.
We first assessed whether the predicted gene expression had sufficient variance for analysis (i.e., not clustered in one node) and the relationship between the SNP(s) used to generate the predicted expression and those included in definitions of CPIC metabolizer phenotypes; we used the Kruskal-Wallis test to assess the association between the predicted expression of TPMT in liver tissue and CPIC guideline-based TPMT metabolizer groups (i.e., normal/indeterminate and poor/intermediate) as described previously.4 We next analyzed the association between predicted expression of TPMT and the 51 phecodes corresponding to the adverse effects of azathioprine using logistic regression. Phecodes associated with adverse events with a p<0.10 were deemed “potential” associations and progressed to the next step of clinical review for case validation. Finally, we grouped the predicted expression of TPMT in liver tissue into tertiles and conducted logistic regression to assess the association between predicted expression and adverse events confirmed by review. We also completed logistic regressions adjusted by (1) age at baseline, (2) sex, and (3) initial dose of azathioprine for each of the phenotypes.
We gathered laboratory results during azathioprine use, including white blood cell counts. As such, to assess the completeness of the phecodes, we compared the incidence of leukopenia phecodes to instances of patients with a laboratory result of white blood cell count <4.0 K/μL, calculating positive and negative predictive values. Additionally, we assessed the TPMT metabolizer status of confirmed leukopenia cases, as defined by current CPIC guidelines, to ascertain whether any identified cases were among normal metabolizers (and could thus show additive clinical utility). Finally, we directly tested the underlying SNP(s) used to predict expression in liver tissue to determine any associations with significant outcomes in the primary analysis.
Analyses were conducted using STATA version 17.026 and R version 4.1.0.27