Using genomic DNA from resected tumor tissue, CGP was performed retrospectively as described by Milbury et al. (11) using a clinical trial assay version of FoundationOne^® CDx to identify patient-specific alterations in the ctDNA cohort with the FoundationOne^® Tracker assay ( Figure 1 ). This assay leverages a tissue-informed personalized ctDNA monitoring approach for evaluating therapeutic response and detecting molecular response for patients across tumor types. CGP was performed on patient samples in a Clinical Laboratory Improvement Amendments–certified, College of American Pathologists–accredited, New York State–approved laboratory (Foundation Medicine, Cambridge, MA). Approval for this study, including a waiver of informed consent and a Health Insurance Portability and Accountability Act of 1996 waiver of authorization, was obtained from the Western Institutional Review Board (protocol no. 20152817). Briefly, genomic DNA was extracted from formalin-fixed paraffin-embedded samples, and DNA was end-repaired and A-tailed, and adapters were ligated and hybridization-based captured, followed by sequencing on the Illumina^® HiSeq 4000 (12). Tumor mutational burden (TMB) was calculated from 0.8 megabase (Mb) of sequenced DNA, excluding driver and germline alterations (13). A novel, proprietary algorithm was used to predict somatic probability to minimize the selection of non-tumor–derived variants (germline, CH-derived, and sequencing artifacts). This algorithm includes a novel logistic regression model to predict probability of a variant being somatic (somatic probability score) based on the difference between the observed variant allele frequency (VAF) and the inferred expected germline VAF. This algorithm directly infers the expected germline allele frequency from known germline single-nucleotide polymorphisms located on adjacent genomic regions expected to have the same copy number with the variant in question. The algorithm then filters variants based on the somatic probability score, allele frequency and annotation, and comparison with databases of known single-nucleotide polymorphisms and CH variants. Short variants (SVs) with high somatic probability were selected and submitted for primer design for the monitoring assay.

Multiplex polymerase chain reaction (PCR) sequencing workflow has been described previously (4, 6, 7). Briefly, the personalized amplicon design pipeline was used to design multiplex PCR primer pairs for each variant derived from the CGP result for each patient. The approach can select coding non-silent alterations in cancer-associated genes [termed pathogenic for alterations with known or likely oncogenic significance or termed variant of unknown significance (VUS) coding for alterations with unknown significance] as well as intronic or synonymous alterations (termed non-coding and synonymous) for monitoring (4–7). Predicted amplicons were ranked, and 2–16 were selected for each custom patient-specific panel. Blood sample processing, cell-free DNA library construction, and sequencing were described previously (6). VAFs were determined for each of the targeted variants. Samples with ≥2 variants above the detected threshold were defined as ctDNA-positive. Absolute ctDNA levels [reported as mean tumor molecules of plasma per milliliter (MTM/mL)] in the plasma were determined by normalizing VAF to the plasma volume and input cell-free DNA for each sample. All passing targets were used in the MTM/mL calculation, including those that were undetected.

The IMvigor010 trial (NCT02450331) was a global, phase 3, open-label, randomized, and multicenter study that evaluated the efficacy and safety of adjuvant atezolizumab (up to 1 year or until recurrence of urothelial carcinoma or unacceptable toxicity) compared with observation in patients with muscle-invasive urothelial carcinoma who are at a high risk for recurrence after surgical resection. Imaging assessments for disease recurrence were performed at baseline and every 12 weeks for 3 years and every 24 weeks for years 4–5 and at year 6 (14). Details of the study have been published previously (6, 14). As study specimens have been analyzed previously (6), we investigated residual samples when available. Because residual samples were analyzed post-hoc with variable sample availability, a comparative analysis with the previously published MRD study was not undertaken as part of this initial analysis.

DFS was the primary endpoint of IMvigor010 defined by local (pelvic) or urinary tract recurrence, distant urothelial carcinoma metastasis, or death from any cause (14). The secondary efficacy endpoint was OS, defined as the time from randomization to death from any cause (14). The study was performed according to the Good Clinical Practice and the Declaration of Helsinki. Protocol approval was obtained from independent review boards or ethics committees at each site. An informed written consent was obtained from all patients.

This work was not part of the original statistical analysis plan from IMvigor010; therefore, the work is exploratory. The analysis includes the prognostic value of ctDNA positivity and the effect of atezolizumab on DFS and OS compared with placebo. Nominal P-values are given as an indicator of probability but do not represent statistical significance. Hazard ratios (HRs) for recurrence or death were estimated using a univariable Cox proportional hazards model. Survival distributions for DFS and OS were estimated using a Kaplan–Meier method and tested for difference between patient groups using the log-rank test.

A multivariable Cox proportional hazards model was used to estimate the HR between ctDNA-positive and ctDNA-negative patients in the observation arm adjusting for key clinical variables. A series of univariable Cox proportional hazards models with ctDNA status, nodal status, programmed cell death-ligand 1 (PD-L1) status, tumor stage, prior neoadjuvant chemotherapy, and number of lymph nodes were evaluated one at a time, followed by a multivariable Cox proportional hazards model, including only the variables that were statistically significant (P < 0.05) in the univariable models. Descriptive statistics were used to summarize clinical characteristics. All statistical analyses were performed in R version 4.0.2 (https://www.R-project.org).

For CGP analysis, the frequency of genomic alterations was assessed from metastatic bladder specimens in the FoundationCORE^™ database. To select a cohort representative of metastatic disease, specimens that were collected from the bladder, surrounding tissue (ureter, pelvis, urethra, penis, iliac crest, vulva, and anus), soft tissue, or lymph nodes were excluded. Comparisons of genomic alteration frequency between the IMvigor010 cohort and the FoundationCORE^™ cohort were made with Fisher exact tests adjusted for multiple comparisons using the Benjamini–Hochberg procedure. Homologous recombination deficiency was defined by pathogenic mutations in any of the genes BRCA1, BRCA2, ATM, BARD1, BRIP1, CDK12, CHEK1, CHEK2, FANCL, PALB2, RAD51B, RAD51C, RAD51D, and RAD54L.

We evaluated a personalized MRD assessment and recurrence-monitoring assay based on tissue-informed CGP, which uses a computational algorithm to filter out non-tumor–derived variants, thus avoiding the germline sampling step that is associated with WES methods. The tumor-derived variants were used to measure plasma-based ctDNA levels in a personalized assay at a single MRD time point with a median of 11 weeks postsurgery ( Figure 1 ) and with a median follow-up of 28 months. In the IMvigor010 trial, 809 patients were enrolled as follows: 403 in the observation arm and 406 in the atezolizumab arm [intention-to-treat (ITT) (15) population]. There were 473 patients eligible for CGP (58% of the ITT population), and, ultimately, 396 of these 473 patients were included in the biomarker-evaluable population (BEP) as follows: 184 patients in the observation arm and 212 patients in the atezolizumab arm ( Figure 2 ). Baseline clinical characteristics of the BEP were well balanced between arms and similar to the ITT population ( Supplementary Table 1 ). At the postsurgical time point, the ctDNA-positive population included 66 patients in the observation arm [(66/184 (36%)] and 63 patients in the atezolizumab arm [63/212 (30%)]. Survival outcomes were similar between the ITT population and the BEP ( Supplementary Figure 1 ).

CGP analysis of 396 patients with MIBC across both arms revealed potentially actionable alterations, including FGFR2/3 SVs and fusions (13%), ERBB2 SVs and amplifications (13%), and PIK3CA SVs (20%) ( Figures 3A, B ). TMB high (≥10 mutations/Mb) was identified in 37% (145/392) of patients. Among the 145 TMB high patients, only five were microsatellite instability high, three had unknown status, and 137 were microsatellite stable. Comparing CGP results with a real-world cohort with metastatic bladder cancer (16, 17), a similar distribution was seen, although a few substantial differences were found for PTEN [odds ratio (OR) = 0.30], FGFR3 (OR = 2.07), TP53 (OR = 0.62), and TERT (OR = 0.62) (P < 0.05 for all) ( Figure 3C ). Together, this analysis demonstrates that patients with MIBC harbor multiple potentially actionable alterations in FGFR2/3, ERBB2, and PIK3CA and frequently exhibit high TMB.

To maximize the number of monitorable alterations, the assay included coding cancer-associated alterations (termed pathogenic or VUS coding) and intronic or synonymous alterations (termed non-coding and synonymous) for tracking of 2–16 variants using a personalized assay (4–7). The distribution and the number of monitorable variants were comparable between patients in the observation and atezolizumab arms of IMvigor010 ( Figure 4A ). The median number of monitorable variants was 12 in each arm. As expected, most ctDNA-positive patients experienced disease progression, which was more pronounced in the observation arm ( Figure 4B ). Of the 129 ctDNA-positive patients, 107 (82.9%) experienced disease progression, which included 60/66 (90.9%) in the observation arm and 47/63 (74.6%) in the atezolizumab arm. The plasma VAF distribution of the different tracked variant statuses (mean VAF by status: non-coding and synonymous = 2.53%, pathogenic = 2.11%, and VUS coding = 1.92%, difference not significant by a Kruskal–Wallis test) was similar across all samples ( Figure 4C ), and the distribution of MTM/mL values was similar between observation and atezolizumab arms ( Supplementary Figure 2 ). Alterations across a wide range of VAFs (range, 0.01%–68.6%; median = 0.3%) were detected in ctDNA-positive samples, and monitorable variant statuses were similarly distributed across all samples, exhibiting no obvious bias. In sum, ctDNA-positive samples were readily identified across a range of patients with a differing number of variants designed, VAF, alteration status, and treatment arms.

Analysis of DFS in the observation arm showed that the ctDNA-positive patients had a higher risk of disease recurrence than the ctDNA-negative patients ( Figure 5A ; Supplementary Table 2 ). The median DFS was 3.0 months (95% CI, 2.9–5.5) in the ctDNA-positive population and was not reached in the ctDNA-negative population (observation arm DFS HR = 5.77; 95% CI, 3.84–8.67; P < 0.0001). Of the 66 ctDNA-positive patients in the observation arm, 60 had relapsed at time of analysis; therefore, the positive predictive value was 91%. Of the 118 ctDNA-negative patients in the observation arm, 75 had not relapsed at time of analysis, resulting in a negative predictive value of 64%. In this study, the clinical sensitivity of this ctDNA assay to detect radiologic relapse, corresponding to the true positive rate, was 58% (60/103 in the observation arm) and the specificity was 93% (75/81 in the observation arm). Analysis of OS in the observation arm showed that the ctDNA-positive patients had a higher risk of death than the ctDNA-negative patients ( Figure 5A ; Supplementary Table 2 ). The median OS was 14.1 months (95% CI, 10.5–23.0) in the ctDNA-positive population and was not reached in the ctDNA-negative population (observation arm OS HR = 5.81; 95% CI, 3.41–9.91; P < 0.0001).

Furthermore, patients who were positive for ctDNA had improved DFS and OS with adjuvant atezolizumab compared with patients undergoing observation ( Figures 5A, B ; Supplementary Table 2 ). The median DFS was 5.9 months in ctDNA-positive patients who received adjuvant atezolizumab versus 3.0 months in ctDNA-positive patients who underwent observation (HR = 0.56; 95% CI, 0.38–0.83; P = 0.003) ( Figure 5A ; Supplementary Table 2 ). The median OS was 23.1 months in ctDNA-positive patients who received adjuvant atezolizumab versus 14.1 months in ctDNA-positive patients who underwent observation (HR = 0.66; 95% CI, 0.42–1.05; P = 0.081) ( Figure 5B ; Supplementary Table 2 ). When the ctDNA-positive population was stratified by high TMB (≥10 mutations/Mb) versus low TMB (<10 mutations/Mb), the direction, magnitude, and strength of associations with DFS and OS were in line with what would be anticipated based upon the cohort powering ( Figure 5C , Supplementary Figures 3, 4 ; Supplementary Table 3 ), and prior TMB assay performance in other bladder cancer cohorts (18–20). In addition, we found that stratifying the ctDNA-positive population by PD-L1 status demonstrated a benefit in DFS and OS for high versus low PD-L1 in the atezolizumab arm ( Supplementary Figures 5, 6 ). We performed an exploratory biomarker analysis of FGFR3, ERBB2, PIK3CA, CDKN2A alterations, MDM2 amplification, and homologous recombination deficiency in the ctDNA-positive population comparing atezolizumab with the observation arm, which suggested that FGFR3 wild-type, but not pathogenic mutants, favored atezolizumab, consistent with the JAVELIN 100 analysis (19) ( Supplementary Figure 7 , univariable analysis; Supplementary Figure 8 , multivariable analysis). No significant difference in DFS or OS was found between arms for ctDNA-negative patients (DFS HR = 1.28; 95% CI, 0.88–1.88; P = 0.195; OS HR = 1.25; 95% CI, 0.72–2.15; P = 0.432) ( Figures 5A, B ; Supplementary Table 2 ).

To determine which prognostic factors were the most significantly associated with DFS and OS, we performed univariable ( Supplementary Figure 9 ) and multivariable exploratory analyses ( Figure 6 ). These analyses confirmed that ctDNA status independently identified patients with improved outcomes in the observation arm and, of the clinical variables, was the strongest predictor of both DFS and OS.