Genetic variation associated with thyroid autoimmunity shapes the systemic immune response to PD-1 checkpoint blockade

Abstract

Activation of systemic immune responses using PD-1 checkpoint inhibitors is an essential approach to cancer therapy. Yet, the extent of benefit relative to risk of immune related adverse events (irAE) varies widely among patients. Here, we study endocrine irAE from 7 clinical trials across 6 cancers where atezolizumab (anti-PD-L1) was combined with chemotherapies and compared to standard of care. We show that atezolizumab-induced thyroid dysfunction is associated with longer survival. We construct a polygenic risk score (PRS) for lifetime risk of hypothyroidism using a GWAS from the UK Biobank and apply this PRS to genetic data collected from 2,616 patients of European ancestry from these trials. Patients with high PRS are at increased risk of atezolizumab-induced thyroid dysfunction and lower risk of death in triple negative breast cancer. Our results indicate that genetic variation associated with thyroid autoimmunity interacts with biological pathways driving the systemic immune response to PD-1 blockade.

Fig. 1. Thyroid irAE are common and associated with longer overall survival during treatment with atezolizumab as monotherapy or in combination.

a Fraction of patients that developed hypothyroidism and hyperthyroidism irAE across atezolizumab trials and their corresponding control arms. Low designates Common Terminology Criteria for Adverse Events (CTCAE) grading of 1 and 2. High designates CTCAE grade >2. b Left, results of an individual participant data (IPD) meta-analysis showing the 95% confidence intervals (CI) and point estimate for a time-dependent covariate in a Cox model associating occurrence of a given endocrine irAE and overall survival in patients treated with atezolizumab (N = 3552) or with standard of care treatments in the control arms (N = 2523) across 7 clinical trials from the safety evaluable population. IPD meta-analysis p-values for a two-sided Wald test that the logarithm (log) of the random effect estimate of a time-dependent covariate in a Cox model, stratified across trials, is non-zero for hypothyroidism p = 5.26 × 10⁻¹⁵, hyperthyroidism p = 1.02 × 10⁻⁴, type-1 diabetes p = 0.06, adrenal insufficiency p = 0.4, and hypophysitis p = 0.93 in atezolizumab treated patients. IPD meta-analysis p-values for patients in the control arms for hypothyroidism p = 0.0039, hyperthyroidism p = 0.03, type-1 diabetes p = 0.76, and adrenal insufficiency p = 0.29 obtained by the same test. Subpanels to the right show the 95% CI around the point estimate of the HR for this time-dependent covariate for hyperthyroidism and hypothyroidism split by each individual trial arm. Trials names are abbreviated as follows: imv211 = IMvigor211; impXXX = IMpowerXXX; imm151 = IMmotion151; impas130 = IMpassion130. Abbreviations for treatment combinations are coded as follows: Atezo = atezolizumab monotherapy; A = atezolizumab; C = carboplatin; P = paclitaxel; NabP = Nab-paclitaxel; B = bevacizumab; SUN = sunitinib; E = etoposide. Meta-analysis: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.

Fig. 2. Genetic variation associated with lifetime risk of thyroid autoimmunity contributes to risk of thyroid irAE in atezolizumab treated cancer patients.

a Random effect point estimate and 95% CI for IPD meta-analysis HR expressed in unit normalized PRS for the occurrence of hypothyroidism irAE estimated in a mixed effect Cox model with genotype eigenvectors as fixed effect covariates using data from (N = 1584) atezolizumab and (N = 1302) standard of care treated European ancestry cancer patients across 7 clinical trials. PRSs are abbreviated as follows: hypoT = hypothyroidism; LDpred2 = hypothyroidism PRS constructed by beta-shrinkage; T1D = type-1 diabetes; VIT = vitiligo. IPD meta-analysis p-values for a two-sided Wald test that the mixed effect Cox model estimated log-HR is non-zero for hypoT p = 7.52 × 10⁻⁹, LDpred2 p = 5.49 × 10⁻⁹, T1D p = 0.67, and VIT p = 1.10 × 10⁻⁶ in atezolizumab treated patients and for hypoT p = 0.68, T1D p = 0.31, and VIT p = 0.38 in the control arms stratified by arm. Subpanels show the HR also expressed in unit PRS estimated using a univariable Cox model for each trial arm. b Left, cumulative incidence plot comparing risk of hypothyroidism irAE in patients treated with atezolizumab with above and below-median values for the hypothyroidism PRS. The median value was computed across all patients including those in the control arms. Right, cumulative incidence for the same comparison of patients in the control arms. Shaded regions provide the 95% CI. c Random effect point estimate and 95% CI for the IPD meta-analysis HR expressed in normalized unit PRS for the time to occurrence of hyperthyroidism irAE estimated using a mixed effect Cox model with genotype eigenvectors as fixed effect covariates in (N = 1584) atezolizumab and (N = 1302) standard of care treated European ancestry cancer patients across 7 clinical trials. Meta-analysis p-values for a two-sided Wald test that the estimated log-HR is non-zero for hypoT p = 0.016, T1D p = 0.67, and VIT p = 0.0012 in atezolizumab treated patients and hypoT p = 0.57, T1D p = 0.87, and VIT p = 0.85 in standard of care treated patients stratified by trial arm. d Estimated importance of variants from the hypothyroidism PRS retained in a survival lasso model for time to hypothyroidism irAE in atezolizumab treated patients. The genes whose transcription start sites (TSS) are spanned by the credible set to which the lasso retained variant belongs are provided with no trailing parentheses. The two closest genes in genomic distance to a TSS from credible set ends are indicated by trailing parenthesis containing distance in kilobases (kb). Only genes who have TSS within 500 kb are reported. Dash designates credible sets that span more than 3 TSS. Trials names are abbreviated as follows: imv211 = IMvigor211; impXXX = IMpowerXXX; imm151 = IMmotion151; impas130 = IMpassion130. Abbreviations for treatment combinations are coded as follows: Atezo = atezolizumab monotherapy; A = atezolizumab; C = carboplatin; P = paclitaxel; Nab P = Nab-paclitaxel; B = bevacizumab; SUN = sunitinib; E = etoposide. Meta-analysis: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.

Fig. 3. Genetic variation associated with lifetime risk of thyroid autoimmunity affects hypothyroidism irAE risk during atezolizumab treatment and acts independently of pre-treatment risk factors.

a IPD meta-analysis assessing the association between hypothyroidism irAE and potential pre-treatment risk factors in a multivariable mixed effects Cox model fit to data from (N = 3234) atezolizumab and (N = 2297) standard of care treated cancer patients in the safety evaluable population with pre-treatment thyroid hormone measurements across the 7 clinical trials analyzed stratified across arms. Measurements were normalized across patients by normalization to the quantiles of a standard normal distribution and modeled as random effects. Point estimates and 95% CI for HR for hypothyroidism expressed in unit normalized hormone levels after fitting the model. TSH = pre-treatment measured thyroid-stimulating hormone; fT4 = free thyroxine; fT3 = free triiodothyronine. Gender is encoded as 1 = female and 0 = male. p-values for a two-sided Wald test that the log-HR is non-zero for fT4 p = 0.44, TSH p = 9.93 × 10⁻¹⁴, and gender p = 0.012 in atezolizumab treated patients and fT4 p = 0.25, TSH p = 3.26 × 10⁻⁸, and gender p = 0.00054 in standard of care treated patients. b Random effect point estimate and 95% CI for the IPD meta-analysis hazard ratio expressed in normalized unit PRS for the time to occurrence of hyperthyroidism irAE estimated using a mixed effect Cox model with genotype eigenvectors as fixed effect covariates in (N = 1584) atezolizumab and (N = 1302) standard of care treated European ancestry cancer patients across 7 clinical trials. TSHgwas uses a PRS constructed from a GWAS of TSH levels in individuals not receiving any medication for thyroid dysfunction. hypoT(adj) computes the association between a hypothyroidism irAE and the hypothyroidism PRS adjusted for measured pre-treatment TSH levels and gender using these as additional fixed effect covariates in the model. Meta-analysis p-values for a two-sided Wald test that the estimated log-HR is non-zero for hypoT(adj) p = 3.91 × 10⁻⁷ and TSHgwas p = 0.11 in atezolizumab treated patients and hypoT(adj) p = 0.66 and TSHgwas p = 0.36 in the standard of care treated patients stratified across trial arms. c Cumulative incidence plot comparing risk of hypothyroidism in atezolizumab patients with and without all of the pre-treatment risk factors identified for hypothyroidism irAE in atezolizumab treated cancer patients. d Positive predictive value and sensitivity (also known as precision and recall) for hypothyroidism irAE and population hypothyroidism occurrence across thresholds for the PRS in atezolizumab treated patients and estimated by 4-fold cross validation in the UK Biobank respectively. Curves were also generated for subgroups that have increasing incidence of hypothyroidism irAE. Meta-analysis: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.

Fig. 4. Polygenic risk for hypothyroidism is associated with lower risk of death in European ancestry TNBC patients treated with atezolizumab and nab-paclitaxel.

Kaplan–Meier (KM) plot for OS for triple-negative breast cancer (TNBC) patients of European ancestry treated with atezolizumab plus nab-paclitaxel (top) and placebo plus nab-paclitaxel (bottom) from IMpassion130. Patients were split into two groups on the basis of median split of the hypothyroidism PRS (hypoT) across all European ancestry IMpassion130 patients with germline genetic data. Censoring events are denoted by vertical dashes. Dashed horizontal and vertical lines designate the median survival time. Shaded regions provide the 95% confidence intervals. impas130 = IMpassion130. A = atezolizumab; NabP = Nab-paclitaxel.