Data sources and study design

Three separate sources of data were used for this study. Results from the individual data sources were consistent with each other and were combined to gain more statistical power. The first data source consists of nine in-house historical atezolizumab advanced NSCLC clinical trials. They are the five phase II or III atezolizumab monotherapy studies,3 5 20–22 and four phase III atezolizumab combo-therapy studies.23–26 Patients were consented for treatment and data usage, and were treated in accordance with the Declaration of Helsinki. In-house clinical trial data were used in accordance to internal processes and guidelines. The trial data were analyzed with all patients treated, regardless of intention-to-treat status.

The second source of data was the US nationwide, de-identified, EHR-derived Flatiron Health (FH) database.27 It is a longitudinal database comprizing patient-level structured (eg, laboratory values and prescribed medications) and unstructured data (eg, biomarker reports), curated via technology-enabled abstraction. During the study period, the de-identified data originated from approximately 280 US cancer clinics (~800 sites of care).

The third source of data was the integrated FH/Foundation Medicine (FMI) Clinico-Genomic Database (CGDB).27 28 It consists of patients from the FH database who had genomic data derived from FMI comprehensive genomic profiling. CGDB provides de-identified patient-level data (from FH database) linked to FMI genomic data from the same patient.28 Since CGDB patients were part of the FH database, to avoid including duplicated patients from the two databases, we excluded the patients in FH database with FMI tests. Patient treatment data between January 2011 and February 2020 (data collection cut-off date) from the two databases were used for the analyses. Both databases consist of retrospective observational de-identified anonymized patient-level data; as such, this study was exempt from informed consent and Institutional Review Board requirements. Contracts governing the access and proper use of data were signed and the rules were respected. Patients from these combined sources were then selected according to various criteria, such as data availability, as illustrated in the top part of figure 1A.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Patient selection and treatment outcome group definitions. (A) Patient selection for the study cohorts. Blue numbers in parentheses denote numbers of patients combined from the three data sources (CGDB, the nine in-house atezolizumab studies and FH database) at each step of patient selection and analysis. The numbers at the end of patient selection step with ‘baseline biomarker data’ show the numbers of patients with the most commonly available biomarker analyzed (white blood cell (WBC)). For the other biomarkers, the patient numbers are illustrated in each analysis figure. In each biomarker analysis, patients were stratified into high and low biomarker groups, and the correlation between biomarker levels and treatment outcome groups was studied. Patients with intermediate biomarker levels (eg, second tertile) were not analyzed. The biomarker high and low groups were balanced on baseline characteristics using propensity score-based SMRW method. Patients in the bottom ‘response’ analysis were also included in the ‘progression’ analysis since they had both response and progression outcome data. The ‘progression’ analysis included additional patients from the FH database because disease progression data is available. (B) Kaplan-Meier plot of overall survival of FH patients treated with CPI or chemotherapy with their first CPI or chemotherapy started during 2015 to February 2020 when the therapies were both used. Dotted vertical line indicates the cut-off used to exclude early deaths, when the two survival curves can be well differentiated. The two groups were balanced on their baseline characteristics using propensity score-based SMRW weights. (C) Treatment outcome group definitions. Each group is represented by a schematic patient journey. Same definition is used for CPI or chemotherapy, but CPI is used as an example for illustration. Patients were aligned on their first CPI treatment day (↓). Disease progression during the first 14 days following CPI initiation was excluded as recommended by Flatiron Health (grayed period). Example of disease progression (x) and/or tumor response events (⬤) are indicated on the journey. The analysis of response to CPI (blue box) used both tumor response and disease progression data. A durable response was defined as having a CR or PR within the study duration. The analysis of progression on CPI (yellow box) used only progression data. Progressors were defined as patients with a progression event during the study duration. Footnotes for figure 1A: *Patients who died within 18 weeks from treatment start were removed (see Methods and Discussion for rationale). ^£Under the durable outcome definition (figure 1C), durable response or clinical benefit requires the entire study duration (270 days in this study) to confirm the durability. ^ΨPatient characteristics are provided in table 1. #Number of patients with the most commonly available biomarker (WBC) from FH, CGDB and trials, respectively. ^§Analysis performed on patients using durable response definition in figure 1C. ^¥Analysis performed on patients using durable clinical benefit definition in figure 1C. CGDB, Clinico-Genomic Database; CPI, checkpoint inhibitors; CR, complete response; FH, Flatiron Health; NSCLC, non-small cell lung cancer; PD, progressive disease; PR, partial response; SMRW, Standardized Mortality Ratio Weighting.

For the study design, for each biomarker, patients were stratified into high and low biomarker groups. For PD-L1, they were PD-L1 high versus PD-L1 negative. For biomarkers with continuous values, they were third tertile versus first tertile. For TMB, the first and third tertiles were equivalent to TMB of ≤4.35 and ≥10.44 mut/Mb, respectively. The correlation of biomarker levels with treatment outcome was explored (‘Analysis’ part of figure 1A). The high and low biomarker groups were balanced on patients’ baseline characteristics that are known prognostic factors that can confound the biomarker effect (age, sex, race, disease stage at initial diagnosis, smoking history, histology and Eastern Cooperative Oncology Group (ECOG) performance status (PS)) using multivariate propensity score-based algorithm29 30 (see ‘Statistical analysis’ section for detail). All patient counts in the result plots are weighted counts. Two sets of analyses were performed: (1) response analysis based on durable response definition which delineates ‘responders’ and ‘non-responders’; (2) progression analysis based on durable clinical benefit definition which delineates ‘non-progressors’ and ‘progressors’ (figure 1AC). To distinguish CPI-specific predictive effect versus general prognostic effect of biomarkers, analysis of standard-of-care chemotherapy cohorts was also performed.

To focus on the association of the biomarker with specific drug response rather than with overall survival (OS), which depends on diverse prognostic factors, we removed patients who died early regardless of the treatment received. It has been previously reported31 that the CPI and chemotherapy survival curves did not differentiate well until after 18 weeks (also confirmed in our data set: figure 1B). We therefore restricted the analysis to patients who were alive beyond 18 weeks. We confirmed that the inclusion of early deaths is a confounding effect in the sensitivity analysis (online supplemental figure 1).

Patient treatment and demographics

For the real-world data (RWD), patients with no record of CPI or other immunotherapy were assigned to the chemotherapy cohort if they were treated with cytotoxic monotherapy or combinations with other cytotoxics or targeted therapies. Patients in the CPI cohorts were treated with nivolumab, (~35%) pembrolizumab (~40%), atezolizumab (~20%) or durvalumab (~5%). Assignment to CPI cohort was regardless of other therapies before or after the CPI line. The analysis used the initial line of exposure to CPI or chemotherapy for the treatment of NSCLC. In CPI-treated patients, slightly greater than 50% of the patients had their first CPI in the first line and ~35% in the second line, and among the chemotherapy patients, ~95% had their first chemotherapy in the first line. All RWD patients included in our analysis were at an advanced stage when they started the treatment. There were 1114 stage I–III patients from the atezolizumab studies, which constitute <5% of the 24,152 patients selected for the study, figure 1A.

In the atezolizumab trials, there were a mix of chemo-refractory and chemo-naïve patients. All CGDB data and seven of nine atezolizumab studies,22 24 included patients with epidermal growth factor receptor (EGFR) and anaplastic lymphoma kinase (ALK) mutations. To determine whether we can account for the impact of EGFR or ALK mutation on the relationship between PD-L1 and/or TMB on one side and the response or progression on the other side, we analyzed the FH data set. EGFR or ALK mutation (prevalence of 6.7% and <1%) test results were available for 70% or 66% of the patients, respectively. The impact of the mutation status on the biomarker effect of CPI response was minimal, potentially due to the low mutation prevalence, (≤0.01 max difference in the ORs) and was therefore not pursued further.

Outcome measures

Tumor response data included Response Evaluation Criteria in Solid Tumours(RECIST) response assessments from atezolizumab trials and real-world response (rwR)32 data from CGDB.28 Disease progression was defined using the following data extracted from the databases: real-world progression (rwP)33 abstracted by FH from patient records, death or change to a treatment line of a different drug category. The durable response definition dichotomized patients into responders (defined as having rwR or RECIST complete or partial response (CR/PR) without progression during the entire study duration of 270 days) and non-responders (having progressive disease (PD) or stable disease (SD) in the trials or disease progression in RWD during the study). The duration of 270 days was to ensure prolonged response that would include multiple tumor scans and a duration for which balanced groups can be obtained. Patients who had a response and had no PD/SD/disease progression but lost to follow-up before the study ended were not included in the analysis.

The durable clinical benefit definition dichotomized the patients into non-progressors and progressors, similar for clinical benefit rate. Specifically, non-progressors had no PD (in trials) or disease progression (in RWD) throughout the study duration, and progressors had PD or disease progression. The main difference between the two definitions is that the durable clinical benefit definition considers durable SD as non-progressors, whereas the durable response definition requires durable CR or PR to be ‘responders’. The two treatment outcome definitions are illustrated in figure 1C.

Biomarker measurements

PD-L1 on tumor tissue was detected by immunohistochemistry using anti-PD-L1 antibody clones 22C3, SP142, SP263 or 28–8. Sensitivity analysis using the different antibody clones showed similar patterns of association with response. They were combined to increase power. Samples were considered ‘negative’ if they had a percentage tumor cell with PD-L1 staining <1%, ‘low/mod’ if it was ≥1% and <50%, and ‘high’ it was ≥50%. The Roche trials used the SP142 PD-L1 assay (Ventana Medical Systems, Tucson, Arizona, USA). FH and CGDB patients used mainly 22C3, SP142 or SP263 assays (results extracted from the medical records in the databases). Tumor PD-L1 expression measured closest to and no more than 1-year prior to the index start date of treatment was used for the analyses. In case of multiple records from the same date, the records from the primary site and the most commonly used antibody were selected.

TMB was quantified by FMI using FoundationOne panel next-generation sequencing,34 and defined as the number of non-driver somatic coding mutations detected per megabase of a tumor genome extrapolated from targeted sequenced genome. TMB data from the most recent specimens collected before start of treatment was used. Research Use Only calculations based on FMI’s research algorithm at the time of collection were analyzed.28 35

Measurements from standard blood analytes were extracted from the three data sources. Data from the most recent measurement no more than 1 year before the start of treatment was used. In case of multiple values from the same date, the median was used.

Statistical analysis

The statistical analysis approach we used comprised of two steps. In the first step, patient groups with high and low biomarkers were balanced on their baseline characteristics using the multivariate propensity score-based cohort balancing SMRW (Standardized Mortality Ratio Weighting) algorithm to minimize potential confounders between these comparison groups.30 36 The second step used univariate analysis but made use of the weights coming from the cohort balancing for each patient. Baseline characteristics included age, sex, race, disease stage at the initial diagnosis, smoking history, histology and ECOG PS. These are available demographics and covariates that can potentially impact the outcome (having durable response or not). The descriptive statistics of the baseline characteristics are in table 1 (for the 24,152 patients with relevant outcome data) and online supplemental table 1 (for the initial 71,850 patients) and online supplemental table 2 (for the 8121 patients analyzed for tumor PD-L1).

There were differences in patient characteristics between different data sources, especially between atezolizumab trials and RWD (FH and CGDB). Since patients from the different data sources were present in both the high and low biomarker groups, the differences between these two analysis groups were readily balanced with methods like SMRW. All the results from the association and survival analyses presented in the main body of the paper used balanced cohorts (by applying SMRW weights). It was noted that there were typically no major differences in results between balanced and unbalanced cohorts. An example for unbalanced cohort has been provided in online supplemental figure 2.

Association of biomarkers with the treatment outcomes was explored using Fisher’s exact test. Null hypothesis was complete independence between the groups. Sensitivity analysis using multivariate logistic regression adjusted by the same set of baseline covariates used in propensity score-based balancing mentioned above produced very similar results (ORs were mostly ≤ ±0.03 compared with Fisher’s exact test using propensity score balanced data). In the mosaic plot for visualizing the association (vcd R package37), colors reflect Pearson’s χ² residuals and indicate extent of deviation from expected frequency under complete independence. Forest plot for summarizing the ORs from the Fisher’s exact tests was performed using forestplot R package.30 38 Statistical significance of difference between ORs was estimated using z-test when independence assumption was true (using the asymptotic normal distribution for log odds). Otherwise, bootstrapping was used to estimate the mean and variance of the log OR differences, which were used to calculate the p value. Overlapping and unique patients in each of the data set were sampled separately and then recombined accordingly to capture the dependence between the two data sets. Survival analysis used univariate Cox proportional hazards model with individual biomarkers investigated in separate models. The models were corrected for any potential imbalance in baseline characteristics between the low and high biomarker groups using the weights from the propensity score-based SMRW outputs. The survival curves were performed using Kaplan-Meier method.39 40 All analyses were conducted using R software V.3.6 or 4.0.