Impact of Tumor-intrinsic Molecular Features on Survival and Acquired Tyrosine Kinase Inhibitor Resistance in ALK-positive NSCLC

While tyrosine kinase inhibitors (TKI) have shown remarkable efficacy in anaplastic lymphoma kinase (ALK) fusion-positive advanced non-small cell lung cancer (NSCLC), clinical outcomes vary and acquired resistance remains a significant challenge. We conducted a retrospective study of patients with ALK-positive NSCLC who had clinico-genomic data independently collected from two academic institutions (n = 309). This was paired with a large-scale genomic cohort of patients with ALK-positive NSCLC who underwent liquid biopsies (n = 1,118). Somatic co-mutations in TP53 and loss-of-function alterations in CDKN2A/B were most commonly identified (24.1% and 22.5%, respectively in the clinical cohort), each of which was independently associated with inferior overall survival (HR: 2.58; 95% confidence interval, CI: 1.62-4.09 and HR: 1.93; 95% CI: 1.17-3.17, respectively). Tumors harboring EML4-ALK variant 3 (v3) were not associated with specific co-alterations but were more likely to develop ALK resistance mutations, particularly G1202R and I1171N (OR: 4.11; P < 0.001 and OR: 2.94; P = 0.026, respectively), and had inferior progression-free survival on first-line TKI (HR: 1.52; 95% CI: 1.03-2.25). Non-v3 tumors were associated with L1196M resistance mutation (OR: 4.63; P < 0.001). EML4-ALK v3 and somatic co-alterations in TP53 and CDKN2A/B are associated with inferior clinical outcomes. v3 status is also associated with specific patterns of clinically important ALK resistance mutations. These tumor-intrinsic features may inform rational selection and optimization of first-line and consolidative therapy.

SIGNIFICANCE

In a large-scale, contemporary cohort of patients with advanced ALK-positive NSCLC, we evaluated molecular characteristics and their impact on acquired resistance mutations and clinical outcomes. Our findings that certain ALK variants and co-mutations are associated with differential survival and specific TKI-relevant resistance patterns highlight potential molecular underpinnings of the heterogenous response to ALK TKIs and nominate biomarkers that may inform patient selection for first-line and consolidative therapies.

Abstract 3128: Using Kmerizer, a germline and somatic genotyper for immune associated complex alleles in GuardantINFINITY, for immunotherapy response prediction using cfDNA

Background: HLA germline genotypes and somatic mutations show great promise as emerging biomarkers for immune checkpoint inhibitors (ICIs) and understanding patient prognosis. Multiple studies have shown that HLA homozygosity or loss-of-function somatic mutations negatively correlates ICI response rate. Here we present additional data on the algorithm Kmerizer, designed to perform HLA germline typing and somatic mutation detection from cfDNA input material, and we show how we use these in neoantigen prioritization for patient outcome prediction.

Methods: Kmerizer first leverages the high depth coverage of targeted sequencing to rapidly identify germline alleles by matching k-mers from the input reads to the k-mers of known HLA alleles. Careful realignment of reads ontothe called germlines is followed by proprietary somatic variant calling. MHC class1 germline allele calls are combined with patient mutation data to generate in silico TCR binding affinity predictions using net MHC-4.0. These predictions are compared across cohorts to assess how cancer type, TMB, and ICI response vary with the predicted neoantigens and TCR binding affinity.

Results: Of nineteen cell lines, twelve plasma samples and eight gDNA samples with confirmed HLA typing information, Kmerizer delivered 100% sensitivity on both MHC-I and II genes, with 99.5% and 98.7% specificity, respectively, based on GuardantINFINITY cfDNA sequencing data. For homozygous/heterozygous status, accuracy of 99.1% on class I and 97.7% on class II genes is achieved. HLA allele prevalence among our development samples is consistent with reference cohorts of similar geographic origin in MHC class I genes. The HLA somatic caller achieves &gt;99.99% specificity per base as computed on 48 normal samples, while achieves &gt;91% sensitivity for somatic events with expected allele frequency (AF) ~ 0.15% (AF range[0.08%,0.26%] for detected events) as evaluated through simulations. Additionally, we generated a total of 2,767 immunogenic (ic50&lt;500nM) class-I somatic neoantigens predictions across 112 samples from cancer patients with germline HLA typing results. We found average patient neoantigen TCR binding affinity was significantly associated with cancer type (χ2=86.08,p&lt;0.0001). Top predicted neoantigen binding affinity across patient HLA types were strongly inversely correlated with patient bTMB(rhospearman=-0.25, p&lt;0.0001).

Conclusions: The integration of Kmerizer into GuardantINFINITY enables accurate HLA germline and somatic detection along with neoantigen prediction, offering an enhanced and comprehensive biomarker profiling for ICI outcome prediction.

Citation Format: Sante Gnerre, Jun Zhao, Adrian Bubie, Yvonne Kim, Dustin Ma, Indira Wu, Bojan Losic, Tingting Jiang, Han-Yu Chuang. Using Kmerizer, a germline and somatic genotyper for immune associated complex alleles in GuardantINFINITY, for immunotherapy response prediction using cfDNA [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 3128.

Oncogene Overlap Analysis of Circulating Cell-free Tumor DNA to Explore the Appropriate Criteria for Defining MET Copy Number-Driven Lung Cancer

INTRODUCTION

Defining clinically relevant MET amplification levels in non-small cell lung cancer (NSCLC) remains challenging. We hypothesize that oncogene overlap and MET amplicon size decline with increase in MET plasma copy number (pCN), thus enriching for MET-dependent states.

PATIENTS AND METHODS

We interrogated cell-free DNA NGS results of 16,782 patients with newly diagnosed advanced NSCLC to identify those with MET amplification as reported using Guardant360. Co-occurring genomic mutations and copy number alterations within each sample were evaluated. An exploratory method of adjusting for tumor fraction was also performed and amplicon size for MET was analyzed when available.

RESULTS

MET amplification was detected in 207 (1.2%) of samples. pCN ranged from 2.1 to 52.9. Of these, 43 (20.8%) had an overlapping oncogenic driver, including 23 (11.1%) METex14 skipping or other MET mutations. The degree of (non-MET) oncogene overlap decreased with increases in pCN. Patients with MET pCN ≥ 2.7 had lower rates of overlapping drivers compared to those with MET pCN < 2.7 (6.1% vs. 16.3%, P = .033). None of the 7 patients with pCN > 6.7 had an overlapping driver. After adjusting for tumor fraction, adjusted pCN (ApCN) was also lower for those with overlapping drivers than those without (median ApCN 4.9 vs. 7.3, P =.024). There was an inverse relationship between amplicon size and pCN.

CONCLUSIONS

We propose that a high MET pCN and/or ApCN, together with the absence of overlapping oncogenic drivers and small MET amplicon size, will enrich for patients most likely to derive benefit from MET targeted therapy.

Variation in liquid biopsy cfDNA yield predicted by somatic mutation and clinical phenotypes across primary cancers

e13553

Background: Liquid biopsy provides a non-invasive alternative to tissue biopsy by profiling the genetic mutations and biomarker characteristics of cell-free DNA (cfDNA) derived from primary and metastatic tumors found in patients’ peripheral blood. Previous studies have demonstrated significant variation in cfDNA and circulating tumor DNA (ctDNA) yield based on patient age, sex, primary tumor type, disease stage, and treatment regimen. Somatic mutations are a primary determinant of tumor phenotype, characterizing therapeutic sensitivity and resistance, as well as growth and metastasis rates. However, there has not been extensive investigation as to how presence of somatic mutations affect downstream cfDNA yield and ctDNA fraction. Methods: We screened 140,000 samples analyzed using Guardant360 liquid biopsy test for patient age, diagnosis, treatment, mutations, and associated cfDNA yield (from assay input of 3-10ml peripheral blood, 1 streck tube). We selected and grouped somatic mutations based on prevalence across the cohort. Estimated ctDNA fractions of cfDNA yields were made using the maximum mutant allele frequency (MAF) from among somatic mutations. Samples were organized by phenotype and cfDNA yield and ctDNA fraction was assessed across groups using the Wilcoxon and Kruskal-Wallis for bi-variate and multivariate comparisons, respectively. To understand impact of cfDNA yield variation on assay performance, group comparisons were further evaluated on samples binned by input cfDNA (< 5ng, 5-15ng, 15-30ng, > 30ng) representing cfDNA input requirement ranges current commercial liquid biopsy tests advertise. Results: We found significant yield variation in sample cfDNA and estimated ctDNA fraction between primary disease types ( Myield diff = 16.35ng, p < 0.0001), therapy histories ( Myield diff = 1.49ng, p < 0.001), and patient age groups ( Myield diff = 2.24ng, p < 0.0001), although this finding did not extend to comparisons within input yield bins. We found significant differences in cfDNA and ctDNA fraction across several common somatic mutations, adjusting for test multiplicity, with the greatest difference between TERT splice + and - patients ( Myield diff = 8.57ng, p < 0.0001). Further, we found significant and strong correlations (𝝆 > 0.8, p < 0.001) between EGFR exon 19 deletions and EGFR L858R MAF and estimated ctDNA fraction, and moderate correlation between MAF and cfDNA (𝝆 > 0.25, p < 0.001). Conclusions: Our results indicate that input cfDNA and ctDNA fraction varies by patient age, diagnosis and treatment, as well as with the presence of several common somatic mutations. The demonstrated correlation between somatic mutation presence and cfDNA yield offers a potential explanation for the broad variation of cfDNA yields within tumor and patient phenotypes, and may help inform blood collection strategies for patients with known somatic variants or diagnoses.

Abstract 572: The liquid biopsy Guardant360 CDx has robust performance at low inputs allowing for high rate of returning patient results

Background

Liquid biopsy tests capable of tumor profiling with NGS have recently received FDA approval for companion diagnostic treatment selection and tumor profiling, demonstrating the clinical validity of the identification of druggable variants with such information. As the amount of cfDNA per volume of plasma varies greatly from patient to patient, and whole blood/plasma volumes from advanced cancer patients are often limited, it would be advantageous for a liquid biopsy to be able to operate with minimal input, providing robust results across a broad spectrum of blood sample volumes and resultant cfDNA input. In this study, we evaluate the performance of Guardant360 CDx with as little as 3 mL whole blood and 5 ng cfDNA input.

Methods

The primary concern with low input like 3 mL whole blood is excess preservative in a blood collection tube (BCT) relative to whole blood. In a simulated short draw experiment, the same advanced cancer patient provided whole blood in BCTs with standard, elevated (2X), and high (3.3X) preservatives. We evaluated the molecule coverage across the panel at key exons and positions of interest, across genes evaluated for copy number amplifications, and at frequent fusion gene breakpoint positions. Precision and limit of detection (LoD) have been demonstrated at 5 ng input (P200010B) the lowest allowable input. For clinical utility of a liquid biopsy with low cfDNA input, we also assessed the cfDNA input of 87,305 samples tested with Guardant360 CDx or LDT from a single 10mL BCT.

Results and Discussion

Differences in molecule coverage between the high preservative and standard preservative were found to fall within expected technical variation (the median of log10 differences in exon-level DNA molecule coverage between short draw and reference conditions fell within +/- 0.125), indicating no meaningful loss of sensitivity for detecting somatic variants. Among the 87,305 clinical samples tested by Guardant360 CDx or LDT, median cfDNA yield was 25ng, with 98.5% yielding at least 5 ng, 88.4% at least 10ng, 73.5% at least 15 ng, and 17.9% at least 60ng. Clinically actionable biomarkers are detectable in low input samples. Of the Guardant360 CDx samples yielding 5-15ng cfDNA input, an average of 0.64 clinically significant variants was reported per sample. With the robust analytical performance demonstrated at low inputs, Guardant360 CDx (a 5ng cutoff) may provide tumor profiling results to guide therapy decisions for about 34% more patients compared to some liquid biopsies that use 2 BCTs for minimal 30ng cfDNA yield (a 15 ng cutoff). Retrospective studies for pharma research are often performed with as little as 2mL of plasma, meaning Guardant360 CDx can return results on 88.4% of such samples whereas other technologies may only be able to address 17.9% (extrapolated from the cfDNA yield distribution for 10mL whole blood yielding 4mL plasma).

Citation Format: Aaron O. Richardson, Aaron Hardin, Soni Shukla, Jing Zhao, Catalin Barbacioru, Hilda Beas, Carin Espenschied, Adrian Bubie, Jamie Hutchins, Han-Yu Chuang, Tingting Jiang, Darya Chudova. The liquid biopsy Guardant360 CDx has robust performance at low inputs allowing for high rate of returning patient results [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 572.

Landscape of oncoviral genotype and co-infection via human papilloma and hepatitis B viral tumor in situ profiling

The role of oncoviral genotype and co-infection driving oncogenesis remains unclear. We have developed a scalable, high throughput tool for sensitive and precise oncoviral genotype deconvolution. Using tumor RNA sequencing data, we applied it to 537 virally infected liver, cervical, and head and neck tumors, providing the first comprehensive integrative landscape of tumor-viral gene expression, viral antigen immunogenicity, patient survival, and mutational profiling organized by tumor oncoviral genotype. We find that HBV and HPV genotype and co-infection serve as significant predictors of patient survival and immune activation. Finally, we demonstrate that HPV genotype is more associated with viral oncogene expression than cancer type, implying that expression may be similar across episomal and stochastic integration-based infections. While oncoviral infections are known risk factors for oncogenesis, viral genotype and co-infection are shown to strongly associate with disease progression, patient survival, mutational signatures, and putative tumor neoantigen immunogenicity, facilitating novel clinical associations with infections.

•

ViralMine parses oncoviral genotypes and co-infection from in situ tumor data

•

Oncoviral genotyping of TCGA CESC, HNSC, and LIHC cohorts

•

Tumor fitness, immunogenicity, and mutational signatures associate with oncoviral genotype

Abstract 1507: Regional DNA methylation profiling reveals novel epigenetic intra-tumoral heterogeneity signatures and aberrant molecular clocks in hepatocellular carcinoma

Introduction: There is limited understanding of epigenetic intra-tumoral heterogeneity (ITH) in hepatocellular carcinoma (HCC) and its potential role in driving cancer evolution. We aimed at deciphering the contribution of DNA methylation to molecular heterogeneity in HCC and identifying novel epi-drivers of cancer evolution.

Methods: We conducted reduced-representation bisulfite sequencing (RRBS) in 35 HCC and 12 adjacent non-tumoral regional samples across 9 patients with treatment-naive, early stage tumors. We computed regional differential DNA methylation ITH signatures across 17,276 well-covered sites within (hg19) promoter regions in CpG islands. We integrate these data with RNAseq, CNV, tumor-infiltrating lymphocyte (TIL) and hepatitis-B viral (HBV) expression on the same regional samples. We used linear models to regress out regional methylation signatures of TIL expression and clonality and HBV expression. Since RRBS significantly overlaps Horvath's clock CpG islands1 (237/354), we trained an expanded, aberrant HCC molecular clock function which quantifies clonal evolution in terms of a regionally varying apparent tumor molecular age and provides potential proxies for estimating tumor growth rates and transverse velocities.

Results: There was significant epigenetic ITH in 4/9 patients as determined by tumor regions for one patient clustering closer with other regions from a different patients than to the other tumor regions of the same patient. These epigenetic outliers were enriched in regions with differential TIL burdens (FDR&lt;0.05). We found ITH signatures that highlighted regional demethylation of MGAT1 (FDR &lt; 1e-3), a novel transcriptional target for the beta catenin signaling pathway, and tumor suppressive cyclin dependent kinase CDK3 (FDR &lt; 1e-2). We also found that while epigenetic tumoral age exceeds patient age (p &lt; 0.001), there exist well-defined conditions for highly regionally varying tumor age that implicate punctuated clonal evolution. Finally, using our regionally trained HCC clock, we leveraged the TCGA LIHC (HCC) cohort of 377 HCC single-biopsy patients profiled with Illumina 450k methylation array to predict the distribution of effective tumor ages. We found that our ITH methylation signatures and effective tumor age are significant predictors of survival under bootstrap validation, controlling for tumor stage.

Conclusions: Novel epigenetic ITH signatures in a subset of HCC tumors which further implicate beta catenin signaling and upregulation of key cyclin dependent kinases predict patient survival in the TCGA, and allow for a quantification of an accelerated aberrant HCC molecular time which estimates tumor age (tumor age ~ [1.5 - 10] x patient age) and growth history.

References:

1. Horvath, S. DNA methylation age of human tissues and cell types. Genome Biol. 14, 1-20 (2013).

Citation Format: Adrian Bubie, Paula Restrepo, Amanda Craig, Ismail Labgaa, Myron Schwartz, Swan Thung, Gustavo Stolovitzky, Bojan Losic, Augusto Villanueva. Regional DNA methylation profiling reveals novel epigenetic intra-tumoral heterogeneity signatures and aberrant molecular clocks in hepatocellular carcinoma [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 1507.

Tumor fitness, immune exhaustion and clinical outcomes: impact of immune checkpoint inhibitors

Recently proposed tumor fitness measures, based on profiling neoepitopes for reactive viral epitope similarity, have been proposed to predict response to immune checkpoint inhibitors in melanoma and small-cell lung cancer. Here we applied these checkpoint based fitness measures to the matched checkpoint treatment naive Cancer Genome Atlas (TCGA) samples where cytolytic activity (CYT) imparts a known survival benefit. We observed no significant survival predictive power beyond that of overall patient tumor mutation burden, and furthermore, found no association between checkpoint based fitness and tumor T-cell infiltration, cytolytic activity, and abundance (tumor infiltrating lymphocyte, TIL, burden). In addition, we investigated the key assumption of viral epitope similarity driving immune response in the hepatitis B virally infected liver cancer TCGA cohort, and uncovered suggestive evidence that tumor neoepitopes actually dominate viral epitopes in putative immunogenicity and plausibly drive immune response and recruitment.

Abstract 3415: Tumor fitness and immune exhaustion: Checkpoint vs endogenous settings

Recently proposed tumor fitness measures based on neoepitope profiling and viral epitope similarity have led to treatment efficacy and immune response prediction models in the checkpoint setting for melanoma (SKCM) and small-cell lung cancer (NSCLC). In this work we test if these checkpoint based fitness measures are associated with either tumor-infiltrating T-cell activity or abundance, and also overall patient survival, in the endogenous setting for the SKCM (n=337) and NSCLC (n = 307) TCGA cohorts. We also investigated if epitopes arising from tumor viral co-factors in Hepatitis B liver cancer (HBV+ HCC) drove T cell activity or response compared to neo-epitope burden.

We find no statistically significant correlation between immune activity and response (as measured by tumor RNA-seq profiling) and the proposed neo-epitope checkpoint-based fitness measures, and similarly no significant survival effect. Further, we found firm evidence that tumor neoepitopes dominate HBV viral epitopes in putative immunogenicity and drive immune response in TCGA HCC (n = 178), and confirmed this trend in multi-regional data (12 patients, 72 samples). These results suggest that tumor-immune response and recruitment in SKCM and NSCLC, and their interplay with viral-cofactor HBV in HCC, are fundamentally driven by different factors in the endogenous setting.

Finally, we propose that tumor fitness may be partially encoded by a simple T-cell exhaustion expression signature, which we demonstrate is a significant predictor of patient survival for SKCM and NSCLC and may translate more widely into other tumor types in the endogenous setting.

Citation Format: Adrian Bubie, Nicholas Akers, Augusto Villanueva, Bojan Losic. Tumor fitness and immune exhaustion: Checkpoint vs endogenous settings [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 3415.

Abstract B071: Validating sequence similarity-driven neoepitope fitness models via immunogenomics on TCGA and multiregional tumor data

Clonal fitness and survival models for checkpoint blockade response prediction have recently been proposed on the basis of ranking neoepitopes via an interaction between a proxy of T-cell recognition, a nonlinear function of neoepitope sequence similarity to known antigens, and neoantigen relative MHC binding affinity. In this work we examine the SKCM (n = 337) and NSCLC (n = 305) TCGA datasets to explicitly test if sequence similarity to known antigens from the IEDB is associated with tumor infiltrating lymphocyte (TIL) burden in patients, as measured by TCR sequencing and CDR reconstruction in RNA-seq data. We find that there is no statistically significant association between either the inferred clonality or magnitude of TIL response and neoepitope sequence similarity to known antigens. We do, however, find significant, moderate associations of TIL response to neoepitope burden. Further, we examined the LIHC (n = 193) and UCEC (n = 245) TCGA cohorts to explicitly derive tumor and virally derived epitopes (HepB, HPV respectively) and ranked their relative predicted MHC binding affinity profiles. We find a greater MHC binding affinity bias exists towards neoepitopes compared to virally derived peptides in a natural setting where both viral and tumor antigens are simultaneously present. Moreover, we find low and significant associations between TIL burden and overall neoepitope burden, but no association with overall viral epitope or expression burden. Finally, we used multiregionally sampled data (12 patients, 72 regions) from HepB-positive HCC liver cancer patients to confirm preferential MHC binding affinity and TIL response bias towards neoepitopes still holds and is significant. Our results suggest that neoepitopes dominate in their recruitment potential of, and association with, TIL burden compared to viral-cofactors. They also suggest that neoepitope sequence similarity to known antigens does not recapitulate patient TIL burden to first approximation.

Citation Format: Adrian Bubie, Nicholas Akers, Augusto Villanueva, Bojan Losic. Validating sequence similarity-driven neoepitope fitness models via immunogenomics on TCGA and multiregional tumor data [abstract]. In: Proceedings of the Fourth CRI-CIMT-EATI-AACR International Cancer Immunotherapy Conference: Translating Science into Survival; Sept 30-Oct 3, 2018; New York, NY. Philadelphia (PA): AACR; Cancer Immunol Res 2019;7(2 Suppl):Abstract nr B071.