Seiden, Michael V.

PURPOSE A multi-cancer detection test using a targeted methylation assay and machine learning classifiers was validated and optimized for screening in prospective, case-controlled Circulating Cell-free Genome Atlas (ClinicalTrials.gov identifier: NCT02889978 ) substudy 3. Here, we report test performance in a subgroup of participants with symptoms suspicious for cancer to assess the test's ability to potentially facilitate efficient diagnostic evaluation in symptomatic individuals. METHODS We evaluated test performance (sensitivity, specificity, and accuracy of cancer signal origin [CSO] prediction accuracy) in participants with clinically presenting cancers (CPCs) and noncancer with underlying medical conditions and among two subgroups (65 years and older and GI cancers). Overall survival (OS) of participants who had a cancer signal detected/not detected was compared with SEER-based expected survival. RESULTS A total of 2,036 cancer and 1,472 noncancer participants were included. Specificity was high in all noncancer participants (99.5% [95% CI, 98.4 to 99.8]). In participants with CPCs, the overall sensitivity was 64.3% (95% CI, 62.2 to 66.4) and the overall accuracy of CSO prediction in true positives was 90.3%. For GI cancers, the overall sensitivity was 84.1% (95% CI, 80.6 to 87.1). In participants 65 years and older, test performance was similar to that of all participants. Individuals with cancers not detected had a significantly better OS than that expected from SEER ( P < .01). CONCLUSION This test detected a cancer signal with high specificity and CSO prediction accuracy and moderate sensitivity in symptomatic individuals, with especially high performance in participants with GI cancers. The survival analysis implied that the cancers not detected were less clinically aggressive than cancers detected by the test, providing prognostic insights to physicians. This multi-cancer detection test could facilitate efficient workup and stratify cancer risk in symptomatic individuals.

Abstract Introduction: A test that detects cancer signal across multiple cancer types and predicts signal (tissue) origin (SO) could aid in more efficient diagnostic workup and shorten time to cancer diagnosis in individuals with signs and symptoms.Methods: The Circulating Cell-free Genome Atlas (CCGA; NCT02889978) study is a prospective, longitudinal, multicenter, case-control study to develop and validate a multi-cancer detection test. The 2nd CCGA substudy utilized a targeted methylation-based cell-free DNA assay and machine learning algorithm and included assessment of test performance (sensitivity and SO prediction accuracy) in a subgroup of participants with clinically presenting cancers (CPCs) that were undiagnosed prior to blood draw. Specificity was assessed in the noncancer group and subgroups with confounding (nonmalignant) conditions (CCs; eg, cirrhosis) and noncancer participants enrolled in hematology clinics (HCs).Results: Specificity was 99.5% (95% confidence interval: 98.2-99.9%; 396/398), 93.8% (71.7-99.7%; 15/16), and 99.3% (96.0-100.0%; 136/137) for the noncancer group, CCs subgroup, and HCs subgroup, respectively. Overall sensitivity among those with CPCs was 66.4% (62.2-70.3%; 344/518). Sensitivity of cancer signal detection increased with increasing clinical stage (Table). SO prediction accuracy was 91.7% (88.3-94.3%; 300/327) among CPC participants with cancers detected, excluding those with multiple or unknown primaries. The test demonstrated prognostic value as detected cancer participants had worse survival probability than those not detected. Conclusions: This multi-cancer detection test detected cancer signals and predicted SO in individuals with CPCs with high specificity. These findings support further clinical development of this multi-cancer detection test that could accelerate the diagnostic resolution of symptomatic cancers. Table. Sensitivity by Clinical Stage Across Cancer Type in Clinically Presenting CancersClinical StagePositive Test/Total Cancer; Sensitivity (95% CI)All*344/518; 66.4% (62.2-70.3%)I33/122; 27.0% (20.0-35.5%)II60/102; 58.8% (49.1-67.9%)III103/121; 85.1% (77.7-90.4%)IV136/147; 92.5% (87.1-95.8%)Not expected to be staged9/21; 42.9% (24.5-63.5%)Non-informative2/4; 50.0% (15.0-85.0%)CI, confidence interval.*One participant who had a positive test result had multiple primaries with clinical stage I and not-expected-to-be-staged. Citation Format: Alan H. Bryce, Minetta C. Liu, Michael V. Seiden, David D. Thiel, Donald Richards, Carlos Becerra, Kathryn N. Kurtzman, Xiaoji Chen, Tony Wu, Quan Zhang, Jingjing Gao, Nan Zhang, Earl Hubbell, Arash Jamshidi, Eric T. Fung, Eric A. Klein. Performance of a cell-free DNA-based multi-cancer detection test as a tool for diagnostic resolution of symptomatic cancers [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr LB058.

Abstract Introduction: A multi-cancer early detection (MCED) test as a complement to existing screening tests could increase the number of cancer cases detected in a population, potentially improving patient outcomes and survival as well as decreasing harmful and aggressive treatments. The Circulating Cell-free Genome Atlas study (CCGA; NCT02889978) was designed to develop and validate a blood-based MCED test analyzing plasma cell-free DNA (cfDNA) to detect cancer signals across multiple cancer types and simultaneously predict their signal origin. Here, the results of the third and final pre-specified CCGA validation sub-study for a refined MCED test in a large cohort in preparation for clinical use are reported. Methods: CCGA is a prospective, multicenter, case-control, observational study with longitudinal follow-up (overall population N=15,254). In this sub-study (n=5309), key primary objectives were to evaluate test performance for cancer signal detection (specificity, overall sensitivity, sensitivity by clinical stage) and signal origin prediction (accuracy). cfDNA from evaluable samples was analyzed using a targeted methylation bisulfite sequencing assay and a machine learning algorithm. The classifier was trained to target a specificity of 99.4% and locked before analysis of the independent validation set. Overall, 4077 participants comprised the independent validation set with confirmed status (cancer: n=2823; non-cancer: n=1254 with non-cancer status confirmed at year-one follow-up). MCED test results are reported for this confirmed status set. Results: Mean (SD) age in the cancer and non-cancer groups was 62.6 (11.76) and 56.2 (12.63) years, respectively. Specificity for cancer signal detection was 99.5% (1248/1254; 95% confidence interval: 99.0-99.8%). Overall sensitivity for cancer signal detection was 51.5% (1453/2823; 49.6-53.3%); sensitivity increased with stage (Stage I: 16.8% [14.5-19.5%], Stage II: 40.4% [36.8-44.1%], Stage III: 77.0% [73.4-80.3%], Stage IV: 90.1% [87.5-92.2%]). Stage I-III sensitivity was 67.6% (593/877; 64.4-70.6%) in a pre-specified set of 12 high-signal cancers accounting for ~63% of annual US cancer deaths [1] and was 40.7% (863/2118; 38.7-42.9%) in all cancers. Cancer signals were detected across &gt;50 cancer types [2]. Overall accuracy of signal origin prediction in true positives was 88.7% (87.0-90.2%). Conclusions: In this pre-specified, large-scale, clinical validation sub-study of CCGA, the MCED test detected cancer signals across &gt;50 cancer types, which is critical to maximize the number of cancer cases detected in a population. This MCED test performed with high specificity and high accuracy of signal origin prediction. These data lay the foundation for population-scale clinical implementation of this test. 1.US Mortality Data 1969-2016 (www.seer.cancer.gov); based on 2015-2016. 2.Amin et al. CA Cancer J Clin. 2017;67:93e99. Citation Format: Eric A. Klein, Donald Richards, Allen Cohn, Mohan Tummala, Rosanna Lapham, David Cosgrove, Gina Chung, Jessica Clement, Jingjing Gao, Nathan Hunkapiller, Arash Jamshidi, Kathryn Kurtzman, Michael V. Seiden, Charles Swanton, Minetta C. Liu. Clinical validation of a targeted methylation-based multi-cancer early detection test [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr LB013.

Abstract Purpose: We recently reported the development of a cell-free DNA (cfDNA) targeted methylation (TM)-based sequencing approach for a multi-cancer early detection (MCED) test that includes cancer signal origin prediction. Here, we evaluated the prognostic significance of cancer detection by the MCED test using longitudinal follow-up data. Experimental Design: As part of a Circulating Cell-free Genome Atlas (CCGA) substudy, plasma cfDNA samples were sequenced using a TM approach, and machine learning classifiers predicted cancer status and cancer signal origin. Overall survival (OS) of cancer participants in the first 3 years of follow-up was evaluated in relation to cancer detection by the MCED test and clinical characteristics. Results: Cancers not detected by the MCED test had significantly better OS (P &lt; 0.0001) than cancers detected, even after accounting for other covariates, including clinical stage and method of clinical diagnosis (i.e., standard-of-care screening or clinical presentation with signs/symptoms). Additionally, cancers not detected by the MCED test had better OS than was expected when data were adjusted for age, stage, and cancer type from the Surveillance, Epidemiology, and End Results (SEER) program. In cancers with current screening options, the MCED test also differentiated more aggressive cancers from less aggressive cancers (P &lt; 0.0001). Conclusions: Cancer detection by the MCED test was prognostic beyond clinical stage and method of diagnosis. Cancers not detected by the MCED test had better prognosis than cancers detected and SEER-based expected survival. Cancer detection and prognosis may be linked by the underlying biological factor of tumor fraction in cfDNA.

Abstract Background: The Circulating Cell-free Genome Atlas study (NCT02889978) is a multi-center, case-control, observational study with longitudinal follow-up (n=15,254; 56% cancer, 44% non-cancer) to support development of a cell-free DNA (cfDNA) multi-cancer early detection test. Previously, we reported that a targeted methylation assay detected and localized &gt;20 cancer types at &gt;99% specificity in individuals with cancer.1,2 Here, we report prediction of cancer (presence/absence) and tissue of origin (TOO) in individuals enrolled with clinical suspicion of cancer but without pathologic diagnosis or treatment at time of enrollment. Methods: Plasma cfDNA from blood samples collected prior to clinical diagnosis was subjected to targeted methylation sequencing. Samples were divided into a training set and an independent validation set to train and validate a machine learning classifier to assess cancer and predict TOO. Performance was assessed in a subset of participants enrolled with suspicion of cancer; subsequently, cancer was confirmed by evaluating a pathologic specimen. Results: Participants being evaluated for suspicion of cancer were classified as confirmed cancer (&gt;20 cancer types; n=164 in training, n=75 in validation) or confirmed non-cancer (n=49 training, n=15 validation). In the confirmed non-cancer group, all training and validation samples were correctly predicted as non-cancer (100% specificity). In the confirmed cancer group, cancer detection across all stages was 40.2% (66/164; 95% confidence interval [CI], 32.7-48.2%) in training and 46.7% (35/75; 95% CI, 35.1-58.6%) in validation. Excluding stage I renal cancers (where detection/tumor fraction is low in plasma and which comprised 20% of participants in this subset) detection across stages was 50.4% (66/131; 95% CI, 41.5-59.2%) and 59.3% (35/59; 95% CI, 45.7-71.9%), respectively. In stages II and above, detection was 70.7% (58/82; 95% CI, 59.6-80.3%) and 78.9% (30/38; 95% CI, 62.7-90.4%), respectively. For detected cancers, TOO was predicted in 93.9% (62/66) samples in training and 100% (35/35) in validation. Of those with a TOO call, accuracy was 85.5% (53/62; 95% CI, 74.2-93.1%) and 97.1% (34/35; 95% CI, 85.1-99.9%), respectively. Conclusion: A cfDNA multi-cancer detection test has shown the potential to predict cancer and TOO in individuals with suspicion of cancer ahead of histologic diagnosis with performance comparable to those with confirmed cancer at the time of blood collection. This was achieved with high specificity and TOO accuracy. The high specificity suggests that the false positive rate could be comparable in populations with average versus higher risk (suspicion) of cancer. These findings suggest that a cfDNA multi-cancer detection test could accelerate the diagnostic resolution of suspicion of cancer. References: 1. Oxnard GR, et al. ASCO Breakthrough Meeting 2019; Abstract 44. 2. Oxnard GR, et al. ESMO Annual Meeting 2019; Abstract 5639. Citation Format: David D. Thiel, Xiaoji Chen, Kathryn N. Kurtzman, Jessica Yecies, Tony Wu, Quan Zhang, Hai Liu, Nan Zhang, Eric T. Fung, Michael V. Seiden, Minetta C. Liu, Geoffrey R. Oxnard, Earl Hubbell, Alexander M. Aravanis, Anne-Renee Hartman, Eric A. Klein. Prediction of cancer and tissue of origin in individuals with suspicion of cancer using a cell-free DNA multi-cancer early detection test [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr CT021.

Abstract Background: Earlier cancer detection could identify tumors when they are more treatable and thus may improve clinical outcomes. A blood-based multicancer early detection test using circulating tumor cell-free DNA (cfDNA) offers the potential to address this unmet need. Our previous discovery work identified whole-genome bisulfite sequencing as outperforming whole-genome and targeted sequencing approaches for multicancer detection across stages. We thus developed a targeted methylation (TM) assay for multicancer detection and tissue-of-origin (TOO) localization. Methods: Participants were from the Circulating Cell-free Genome Atlas (CCGA; NCT02889978) and STRIVE (NCT03085888) studies, both prospective, multicenter, observational studies with longitudinal follow-up. cfDNA samples spanning &gt;20 cancer types of all stages were divided into a cross-validated training set and an independent validation set; there were 3,583 evaluable samples in training (1,530 cancer; 2,053 noncancer) and 1,803 in validation (678 cancer; 1,125 noncancer). Performance in a prespecified set of 12 high-signal cancers was also reported (anus, bladder, colon/rectum, esophagus, head and neck, liver/bile-duct, lung, lymphoma, ovary, pancreas, plasma cell neoplasm, stomach). This multicancer early detection test evaluated cfDNA for the presence of cancer and localization of TOO; the TM classifier was trained to target &gt;99% specificity. Precision was defined as the fraction of correct calls. Results: Participants with and without cancer were similar in age. Specificity across all cancer types in the training and validation sets was 99.8% (95% CI, 99.2-99.9%) and 99.3% (98.3-99.8%), respectively, which reflects a consistent (P=0.292) false positive rate of &lt;1%. Across all cancer types, aggregate sensitivity was also consistent between the training and validation sets (55% [53-58%] and 55% [51-59], respectively; P=0.897). In the validation set, aggregate sensitivity in the prespecified group was 76% (73-78%); detection was 39% (27–52%) in stage I (n=62), 69% (56–80%) in stage II (n=62), 83% (75–90%) in stage III (n=102), and 92% (86–96%) in stage IV (n=130). Overall, the classifier assigned a TOO across &gt;20 cancer types in 96% of samples (344/359); of these, the TOO was correct in 93% (321/344) of cases, which was consistent with training set analyses. Conclusions: This multicancer early detection test detected cancer signal across &gt;20 cancer types with a single, fixed, low false positive rate and highly accurate TOO localization. Importantly, results in the independent validation set were consistent with the training set, demonstrating the robustness of machine learning classifier training, and confirms that data were not overfitted. These data support the feasibility of a single blood-based test that can detect multiple cancers, supporting further clinical development in preparation for the return of results. Citation Format: Anne-Renee Hartman, Geoffrey Oxnard, Eric Klein, Michael Seiden, Earl Hubbell, Oliver Venn, Arash Jamshidi, Nan Zhang, John Beausang, Samuel Gross, Kathryn Kurtzman, Eric Fung, Brian Allen, Alexander Fields, Hai Liu, Mikkael Sekeres, Donald Richards, Peter Yu, Alexander Aravanis, Minetta Liu. Multicancer detection of early-stage cancers with simultaneous tissue localization using a plasma cfDNA-based targeted methylation assay [abstract]. In: Proceedings of the AACR Special Conference on Advances in Liquid Biopsies; Jan 13-16, 2020; Miami, FL. Philadelphia (PA): AACR; Clin Cancer Res 2020;26(11_Suppl):Abstract nr IA02.

44 Background: A noninvasive cfDNA blood test detecting multiple cancers at earlier stages could decrease cancer mortality. In earlier discovery work, whole-genome bisulfite sequencing outperformed whole-genome and targeted sequencing approaches for multi-cancer detection across stages at high specificity. Here, multi-cancer detection and TOO localization using bisulfite sequencing of plasma cfDNA to identify methylomic signatures was evaluated in preparation for clinical validation, utility, and implementation studies. Methods: 2301 analyzable participants (1422 cancer [ > 20 tumor types, all stages], 879 non-cancer) were included in this prespecified substudy from the Circulating Cell-free Genome Atlas (CCGA) (NCT02889978) study - a prospective, multi-center, observational, case-control study with longitudinal follow-up. Plasma cfDNA was subjected to a targeted methylation sequencing assay using high-efficiency methylation chemistry to enrich for methylation targets, and a machine learning classifier determined cancer status and tissue of origin (TOO). Observed methylation fragments characteristic of cancer and TOO were combined across targeted regions and assigned a relative probability of cancer and of a specific TOO. Results: Performance is reported at 99% specificity (ie, a combined false positive rate across all cancer types of 1%), a level required for population-level screening. Across cancer types, sensitivity ranged from 59 to 86%. Combined cancer detection (sensitivity [95% CI]) was 34% (27-43%) in stage I (n = 151), 77% (70-83%) in stage II (n = 171), 84% (79-89%) in stage III (n = 204), and 92% (88-95%) in stage IV (n = 281). TOO was provided for 94% of all cancers detected; of these, TOO was correct in > 90% of cases. Conclusions: Detection of multiple deadly cancers across stages using methylation signatures in plasma cfDNA was achieved with a single, fixed, low false positive rate, and simultaneously provided accurate TOO localization. This targeted methylation assay is undergoing validation in preparation for prospective clinical investigation as a cancer detection diagnostic. Clinical trial information: NCT02889978.

5574 Background: A noninvasive cell-free DNA (cfDNA)-based cancer detection assay offers the hope of a blood test that might reduce morbidity and mortality of cancers, particularly those without recommended screening tests (eg, some gynecologic cancers). CCGA (NCT02889978) is a prospective, multi-center, longitudinal, case-control study evaluating models for discriminating cancer versus non-cancer. Here, we report F/U of control participants (pts) who demonstrated a cancer-signal in CCGA. Methods: Clinically evaluable samples (N = 2508) from pts enrolled without a cancer diagnosis (dx; NC) and treatment-naive pts with newly diagnosed cancer (C) were divided into training (n = 1564; 580 NC, 984 C) and test (n = 944; 368 NC, 576 C) sets. Classification performance (cancer/non-cancer) was assessed via 3 prototype assays: whole-genome bisulfite (WGBS), whole-genome (WGS), and targeted (507 gene) sequencing. Notable outlier NC pts were identified with cancer-like scores in either ≥2 assay classification results or by the presence of known cancer drivers with ≥1 assay classification result suggesting cancer. All pts are currently in F/U in accordance with study protocol (to date: 80% with > 10 mo and 15% with > 22 mo F/U). Results: Among training and test sets, 8 ( < 1%) NC pts were identified with a cancer-like signal. To-date, 2 have been diagnosed with a gynecologic malignancy: 1 stage IIIc clear cell endometrial carcinoma and 1 stage IIIc ovarian cancer, 3 and 2 months (mo) post-enrollment [PE], respectively. Among C pts in the study, sensitivity (at 98% specificity; WGBS) in these cancer types was: uterine/endometrial: 11% (n = 27 train) and 22% (n = 9 test); ovarian: 82% (n = 17) and 71% (n = 7). In addition, a third NC pt was diagnosed with a stage IV lung cancer 15 mo PE. Conclusions: This cfDNA-based assay detected a cancer-like signal that anticipated a clinical presentation of cancer in undiagnosed pts as early as 15 months prior to the actual dx. High specificity ( > 99%) requires accounting for undiagnosed cancers in study design and analysis. Together, these data suggest that this prototype assay may have high performance detecting a variety of gynecological and other cancers. Clinical trial information: NCT02889978.

3049 Background: For multi-cancer detection using cfDNA, TOO determination is critical to enable safe and efficient diagnostic follow-up. Previous array-based studies captured < 2% of genomic CpGs. Here, we report genome-wide fragment-level methylation patterns across 811 cancer cell methylomes representing 21 tumor types (97% of SEER cancer incidence), and define effects of this methylation database on TOO prediction within a machine learning framework. Methods: Genomic DNA from 655 formalin-fixed paraffin-embedded (FFPE) tumor tissues and 156 isolated cells from tumors was subjected to a prototype 30x whole-genome bisulfite sequencing (WGBS) assay, as previously reported in the Circulating Cell-free Genome Atlas (CCGA) study (NCT02889978). Two independent TOO models, one with and one without the methylation database, were fitted on training samples; each was used to predict on the test set. A WGBS classifier was used to detect cancer at 98% specificity; reported TOO results reflect percent agreement between predicted and true TOO among those detected cancers (166 cases: 81 stage I-III, 69 stage IV, 16 non-informative). Results: Genome-wide methylation data generated from this database allowed fragment-level analysis and coverage of ~30 million CpGs across the genome (~60-fold greater than array-based approaches). Incorrect TOO assignments decreased by 35% (20% to 13%) after incorporating methylation database information into TOO classification. Improvement was observed across all cancer types and was consistent in early-stage cancers (stage I-III). Respective performances in breast cancer (n = 23) were 87% vs 96%; in lung cancer (n = 32) were 85% vs 88%; in hepatobiliary (n = 10) were 70% vs 90%; and in pancreatic cancer (n = 17) were 94% vs 100%. Results using an optimized approach informed by these results in a large cohort of CCGA participants will be reported. Conclusions: Incorporating data from a large methylation database improved TOO performance in multiple cancer types. This supports feasibility of this methylation-based approach as an early cancer detection test across cancer types. Clinical trial information: NCT02889978.

1545 Background: Screening tests for early cancer detection are often criticized due to risk of overdiagnosis—detection of good prognosis cancers which may not require immediate treatment. We recently reported development of cfDNA sequencing approaches for cancer detection; longitudinal follow-up (F/U) data were utilized here to evaluate prognostic significance of cancer detection using cfDNA. Methods: Plasma cfDNA samples were subjected to whole-genome bisulfite sequencing (WGBS, 30X) as part of a previously-reported Circulating Cell-free Genome Atlas (CCGA; NCT02889978) substudy. This exploratory analysis evaluated the overall survival (OS) of training and test set participants (pts) with cancer (20 cancer types, any stage I-IV). Combining train and test set pts, univariate and multivariate analyses (Cox proportional hazards) assessed OS association with WGBS result (cancer detected vs not detected, set at 98% specificity), clinical stage (IV vs I-III), diagnostic method (symptom- vs screen-detected), sex, age, and histologic grade. Results: Of 827 pts from the training set with F/U (median 12.2 mo), 334 (40.4%) had WGBS-detected cancer. Among 127 (15.4%) pts with cancer that died during F/U, cancer was detected in 104 (81.9%). Results were similar in the test set. In univariate analyses all variables were associated with prognosis, including WGBS result (HR 7.7 p<0.001). In multivariate analyses accounting for other covariates, the three variables that most significantly remained prognostic were WGBS (HR 3.0, p<0.001), clinical stage (HR 3.3, p<0.001), and diagnostic method (HR 3.0, p<0.001). Validation of these findings is ongoing in an independent cohort of ~5,000 cancer pts from CCGA using an optimized assay; updated performance results will be reported. Conclusions: Cancers detected using WGBS of cfDNA had a worse prognosis than cancers not detected. WGBS cancer detection carried comparable prognostic significance as clinical stage. By preferentially detecting higher risk cancers, cancer detection using plasma cfDNA may avoid some of the overdiagnosis that has been seen with some existing cancer screening methods. Clinical trial information: NCT02889978.