Development and clinical application of an integrative genomic approach to personalized cancer therapy

Here we report clinical application of an integrative genomic approach for PCT. Overall, the integrative genomic approach applied in our study identified significantly more cancer-related somatic mutations as well as more actionable genetic and genomic alterations than would have been identified using any one of several commonly used commercial cancer panels (Table 5). In several patients where panel sequencing of mutational hotspots would not have identified any clinically actionable mutations, rare somatic mutations uncovered through our comprehensive genomic sequencing have the potential to impact treatment options and ultimately change clinical outcome for individual patients. Indeed, treatment decisions were influenced by our genomic studies in four cases. Other cases have been reported in which whole genome sequencing uncovered rare mutations that were missed by targeted sequencing and that had the potential to guide treatment but was discovered too late. For example, a rare activating BRAF mutation (p.L597R) was identified in an aggressive metastatic melanoma using whole genome sequencing that had previously been designated as “wild-type” for BRAF p.V600E/K mutations and several common KIT mutations using targeted sequencing [57]. The BRAF p.L597R mutation was demonstrated to be responsive to MEK inhibitors in metastatic melanomas [57], suggesting that with more comprehensive testing the patient in question may have benefitted from this treatment option [58]. While we identified actionable alterations in 91 % of the cases tested, we did not identify actionable results in four cases. More comprehensive approaches for profiling tumors that employ additional platforms such as proteomic and metabolic profiling may potentially provide clinically actionable information in these cases.

In comparison to commonly used targeted cancer panel sequencing, more comprehensive genomic profiling provides a number of advantages. First, cancer panels are generally designed to include well-characterized cancer-associated genes, whereas WES and RNA-Seq enable elucidation of the underlying cancer genetic drivers at the pathway level, given alterations in several components of the same pathway allow us to predict pathway dysregulation with higher confidence. For example, several genes such as DKK1/2, CSNK1A1, INPP5D, and INPPL1 in the cases we discussed in detail (Fig. 3) are not present on the cancer panels we examined, yet their functional roles in respective signaling pathways are well-documented [59, 60]. Many somatic mutations identified from WES are of unknown significance; however, as more WES data accumulate, retrospective analysis of these genomic data and association with clinical outcome and treatment response could inform novel actionable mutations. Second, whereas most clinically available cancer panel sequencing tests are designed to screen only tumor DNA, the more integrative profiling makes it possible to differentiate germline from somatic variants, especially in cases of novel or rare variants. Without knowledge of germline variants, accurately identifying somatic mutations becomes problematic as panel sizes increase [8]. For example, rare or private germline variants located in protein domains with hotspot oncogenic mutations, such as kinase domains, may be interpreted as somatic mutations by workflows using only tumor samples. However, knowledge that a variant is germline would be important because genetic counseling or increased surveillance may be indicated for a patient who is a carrier of a pathogenic allele, thus the cancer risk of the patient or their family may be substantially increased. For these reasons, despite the logistical challenges relating to sample collection and tissue banking, we favor sequencing of matched tumor-normal pairs. Sequencing of germline DNA provides the added benefit of identifying variants that may inform on drug metabolism or DNA repair pathways that are associated with response to chemotherapy, providing for the possibility of informing on the efficacy and toxicity of a given drug for a given individual. Such variants may be missed or inaccurately called from the tumor DNA. As illustrated in our follow-up patient survey, germline mutations associated with drug responses in patient P0025 would not have been identified with cancer panel sequencing of tumor samples only, and the rare EGFR somatic mutation in patient P0015 would not have been called somatic with high confidence without sequencing of both tumor and the matched normal control samples. In addition, germline variants that predispose to cancer may provide prognostic value and inform on treatment options, as we demonstrated in one of our cases (P0013) with the identification of a BRCA1 germline mutation that led to our recommendation for cisplatin chemotherapy.

We have further shown that RNA-Seq can significantly augment the utility of genetic testing for PCT. Across a number of cancer types, clinically relevant subclasses can be defined based on gene expression patterns. In breast cancers, both luminal A and luminal B subclasses are ER+. However, luminal A subtype is more responsive to hormonal therapy while luminal B subtype is more responsive to chemotherapy [42]. Therefore, sub-classification of breast cancers using RNA-Seq-derived signatures may have clear therapeutic implications. For example, in one of our cases (P0013), we identified a discrepancy in classification of the breast cancer in this patient between the pathology report (which classified the patient’s tumor as ER+ under the current ASCO/CAP guidelines [43]) and the RNA-Seq analysis results we carried out on this patient (resulting in a classification of basal like). A re-review of the pathology indicated that only 10 % of the tumor nuclei stained positive for ER, whereas the majority of the tumor cells are triple negative ER–/PR–/HER2–, suggesting in some cases the molecular profiling data may lead to a more accurate molecular characterization of the tumor. It is not uncommon that cancer driver pathways are activated by abnormal expression of key pathway components in the absence of genetic alterations. For example, high levels of expression of epiregulin and amphiregulin not only imply EGFR pathway activation, they have also demonstrated clinical utility as predictive biomarkers for response to anti-EGFR treatment [34, 35]. In patient P0009, a quadruple negative colon cancer case, both epiregulin and amphiregulin exhibited extraordinary over-expression based on RNA-Seq data analysis, making a case for cetuximab treatment. Identification of these types of gene expression biomarkers in the absence of genetic alterations would not be possible with DNA sequencing data alone. In addition, RNA-Seq data can be used to confirm somatic mutations identified in DNA or to infer driver CNAs when gene expression correlates with copy number changes. In genomics findings documents we generated, somatic mutations detected by both WES and RNA-Seq were denoted as “validated” to emphasize their significance. Finally, oncogenic fusion events cannot be reliably detected from WES, so that RNA-Seq offers a better tool for gene fusion analysis.

Although this study used a 50 % tumor purity cutoff (as determined by a pathologist review of HE sections adjacent to the tissue being sequenced), it is notable that many clinical specimens will fall below that threshold. Despite the 50 % cutoff, the estimate of tumor purity from WES and array data shows that 14 of the 46 tumor specimens may be falling below this threshold (Additional file 1: Table S11), highlighting known challenges in making tumor purity estimates. Importantly, analysis of WES and RNA-Seq identified actionable somatic alterations in samples containing as little as 25 % tumor cells (e.g. P0001, P0040) based on post-NGS purity estimation. As the price of sequencing continues to decline, it will be feasible to achieve higher sequencing depth in WES assays and resolve lower-purity tumors.

While our study demonstrates a clear benefit of WES and RNA-Seq over common targeted sequencing panels, the cost of WES and RNA-Seq remains an issue given their substantially higher price and the fact that today there is not a clear reimbursement mechanism for generating such data. Further, the data analyses and interpretations for the combined WES and RNA-Seq data take significantly more time to complete than data generated from targeted panel sequencing. Sample availability and quality also pose a barrier to performing genome-wide profiling. In addition, sequencing of gene panels, given they cover a small fraction of DNA compared to WES, is often performed at significantly higher sequencing depth, which allows for both increased sensitivity and specificity, given the heterogeneous mix of cells that comprise most tumor samples. However, the extent to which targeting of sub-clonal alterations can achieve clinical benefit is still under investigation.

Given the advantages and disadvantages to comprehensive sequencing, one could envision a staggered approach in which samples first undergo targeted sequencing and then progress to a deeper characterization if actionable alterations are not identified. Out of all DNA assays employed in our study, targeted panel sequencing had the highest data generation success rate (98/99 samples, 99.0 %; all samples were attempted regardless of available DNA mass), required the least input DNA (usually 30 ng, but lower input was accommodated), provided the fastest turnaround time, and produced the highest sequencing depth (mean 2587X), allowing detection of variants with allelic fractions as low as 5 % based on cell line dilution experiments (data not shown). Thus, a clinical pipeline should begin with a targeted panel (either a pan-cancer mutation hotspot panel as we employed, or a cancer-specific panel selected based on the patient’s diagnosis). A progression of increasingly comprehensive targeted panels is also possible. If no actionable alterations are identified, data generated from initial targeted panel sequencing may be informative to selecting parameters of follow-up assays, e.g. selecting the WES depth based on an initial tumor purity estimate from the panel (which would be based on non-actionable somatic variants). In our study, sufficient DNA mass was available for 41 out of 45 patients (91.1 %) to carry out WES. Although WES was successful for generating usable data for all 41 patients, eight of 41 patients (19.5 %) required multiple attempts, sometimes needing re-extractions of additional DNA from the tumor specimen, leading to delays (Additional file 2: Supplementary Results). Commercial kits for low-input, poor-DNA-quality WES library preparation are becoming increasingly available to ameliorate these issues and it is reasonable that more patients would be amenable to WES in the clinic in the future. Lastly, our concordance analysis of CNA from WES versus arrays data processed using the same CNA algorithm [61] (Additional file 2: Supplementary Results) highlights that the variability in CNA findings between platforms requires additional work to address. Thus, balancing the costs and benefits for different personalized genomics strategies is a rapidly evolving process.

Although cancer panel or exome sequencing based genetic testing is being broadly implemented, only a small portion of patients with actionable alterations followed the treatment recommendations through off-label use of drugs or enrollment into genotype-matched trials [11, 12, 14, 62]. In our study, according to the follow-up patient survey, only four out of the ten responders stated genomic findings altered treatment. The majority (21/31) of the patients we contacted did not even respond to the survey request. Previous studies described several major challenges in linking genomic findings to genotype-matched treatment [12, 14, 62], and these obstacles are present in our study as well. First, most of the patients were referred to Mount Sinai hospitals, and some of them did not return after genomic testing. Second, many patients in our study had gone through several lines of treatment at the time of genomic testing, and they were unlikely to be eligible for trials due to health deterioration and poor performance status. Third, genotype-matched trials may not be available, particularly for less common tumor types or less commonly mutated genes. Finally, long turnaround time of comprehensive genomic profiling in our study poses a significant barrier, similar to what has been reported by the Peds-MiOncoSeq consortium [14]. While many of the challenges are inherent to the overall design and the observational nature of current genomic-based PCT, we are taking measures for improvement such as reducing turnaround time in order to better realize the potential of genomics-driven individualized cancer treatment.