Validation of an NGS Panel Designed for Detection of Actionable Mutations in Tumors Common in Latin America

Next-generation sequencing (NGS) is progressively being used in clinical practice. However, several barriers preclude using this technology for precision oncology in most Latin American countries. To overcome some of these barriers, we have designed a 25-gene panel that contains predictive biomarkers for most current and near-future available therapies in Chile and Latin America. Library preparation was optimized to account for low DNA integrity observed in formalin-fixed paraffin-embedded tissue. The workflow includes an automated bioinformatic pipeline that accounts for the underrepresentation of Latin Americans in genome databases. The panel detected small insertions, deletions, and single nucleotide variants down to allelic frequencies of 0.05 with high sensitivity, specificity, and reproducibility. The workflow was validated in 272 clinical samples from several solid tumor types, including gallbladder (GBC). More than 50 biomarkers were detected in these samples, mainly in BRCA1/2, KRAS, and PIK3CA genes. In GBC, biomarkers for PARP, EGFR, PIK3CA, mTOR, and Hedgehog signaling inhibitors were found. Thus, this small NGS panel is an accurate and sensitive method that may constitute a more cost-efficient alternative to multiple non-NGS assays and costly, large NGS panels. This kind of streamlined assay with automated bioinformatics analysis may facilitate the implementation of precision medicine in Latin America.


Introduction
In recent decades, molecular pathology has advanced substantially thanks to the exponential growth of genetic sequencing technology. The introduction of next-generation sequencing (NGS) opened the doors to high-throughput, multi-gene, massive data collection. This tool's ability to sequence more, faster, and at a reduced cost has made it attractive for many clinical research applications. In cancer, using this technology to interrogate solid tumor samples has propelled a massive characterization of genes involved in the disease [1,2]. This rise in "oncogenomics" has been accompanied by an increase in cancer drug approval and development [3,4]. Identifying tumor-specific genetic signatures and correlating them to treatment outcome has evolved into a strategy termed "precision medicine", a new diagnostic and treat process based on approved genomic biomarkers [4].
In Latin America and the Caribbean, 1.4 million new cancer cases were estimated to occur in 2018, while mortality rates varied among and within the region [5,6]. The most common types of cancer with the highest incidence are prostate (age standardized rate (ASR) 60.4), breast (ASR 56.8), colorectal (ASR 18.6), cervix uteri (ASR 15.2), lung (ASR 13.1), and stomach (ASR 9.5) cancers [1]. Overall, estimated age-standardized cancer incidence rates in Latin America are lower than those reported in North America and some European countries; however, the region exhibits higher mortality rates [7]. This paradox reflects the disparities in early diagnosis and treatment opportunities in the region.
In high-and medium-income countries, precision medicine is making its way into standard cancer treatment, improving survival and investigational drug trial success for many patients. A combination of factors prevents this helpful tool from becoming accessible to most of the world's population. In Latin America, approved and available gene-based cancer screening assays are often solutions designed to meet first world standards. These large panels are great diagnostic tools in regions of abundant therapy options, but for Latin America and other regions, they are not cost effective, leaving behind a need for more comprehensive regional solutions. Additionally, the absence of automated clinician-ready reporting for many of these approved panels creates another major cost and obstacle to their widespread implementation. As a result, NGS-based oncology panels do not appear to be cost-effective solutions for many governments and are not being implemented in health and insurance systems despite local sequencing capabilities. This scenario creates an urgent need for customized validated solutions and data interpretation in a clinical environment [8].
In addition, an important caveat to interpreting Latin American cancer patient's genetic data is the under-representation of Latin American individuals in global resources characterizing the frequency of both germinal and cancer genome variants. Great cancer genomics efforts, such as TCGA and ICGC, are deprived of minorities (including subjects of Hispanic ethnicity [9]), limiting their capacity to describe somatic mutations with a prevalence below 10% and overcome the somatic background mutation frequency in specific ethnic groups [10]. For instance, the average Amerindian ancestry in cancer patients across all cohorts in TCGA is about 4% [9,11]. Additionally, the Latin American population is under-represented in germline variant repositories, which may induce a false categorization and overestimation of somatic variants [12][13][14]. Thus, an additional blood sample should accompany the tumor sample, increasing the sequencing costs.
To address these challenges, we designed, optimized, and validated a hybridizationbased target enrichment workflow with multiple automated analyses capable of detecting variants in 25 genes; 23 of them with an association to a drug's response supported by the FDA or well-powered studies with consensus from experts in the field. Although this panel was designed to meet current and near-future Chilean precision oncology needs, we expect the panel and workflow to be relevant to other countries in the region. This workflow was locally validated using breast, colorectal, gastric, ovarian, pancreatic, and gallbladder tumor tissue samples and we reported its ability to detect single nucleotide variants (SNVs) and small insertions and deletions with high sensitivity and specificity. Additionally, high reproducibility was obtained for non-synonymous variants between and within runs. Finally, to address the shortage of health professionals trained in bioinformatics, the entire workflow, including quality control of sequencing data and calling for somatic variants, was automated and made available.

Panel Design
The panel targets hotspots, selected exons, or complete coding regions of 25 genes and includes predictive biomarkers in solid tumors. We refer to this panel, plus its associated workflow and analysis, as TumorSec™. For selecting targeted regions, biomarker genes classified with evidence 1, 2, 3a, 3b, R1, and R2 were selected for solid tumors in the OncoKB database (www.oncokb.org, accessed on 1 June 2021) [15]. Next, biomarker mutations with level of clinical evidence A, B, and C were selected in the Clinical Interpretation of Variants in Cancer, CiVic database (https://civicdb.org/home, accessed on 1 October 2018) [16] and the Variant Interpretation for Cancer Consortium (VICC) meta-knowledgebase [17]. Biomarkers were selected based on their level of evidence and incidence in the targeted tumor in Latin America. TP53 and ARID1A complete coding regions were incorporated, as they contain prognosis and predictive chemotherapy biomarkers. The complete list of genes and drug associations is provided in Table 1.  1 Level of evidence according to AMP/ASCO/CAP Level A: FDA-approved therapy included in professional guidelines; B: well-powered studies with consensus from experts in the field; C: FDA-approved therapies for different tumor types or investigational therapies [18]. The highest level of evidence is shown. Some biomarkers may have additional indications with lower levels of evidence for different cancer types or protocols. * For these genes, all exons were targeted.
Synthesis of the soluble, biotinylated probe library was performed on the NimbleGen cleavable array platform (SeqCap EZ Choice (RUO); Roche/NimbleGen, Basel, Switzerland). The probe design was optimized using the NimbleDesign software utility (Nimble-Gen, Roche, Basel, Switzerland).

Sample Information
In total, 183 tumor tissue samples were sequenced for this study. In total, 19 were freshly frozen (FF): 13 colorectal and 6 breast; 164 were formalin-fixed paraffin-embedded (FFPE) blocks: 9 breast, 71 ovary, 1 gastric, 43 gallbladder, and 40 colorectal tumors. Additionally, DNA from 89 whole blood or buffy coat samples were sequenced: 7 from colorectal and 72 from breast cancer patients. Colorectal and gastric cancer samples were obtained from the "Biobanco de Tejidos y Fluidos de la Universidad de Chile." To capture real world heterogeneity in sample quality, breast, ovary, and gallbladder FFPE tissue samples were collected from the pathology services from several sites along the country (Fundación Arturo López Pérez, Clínica Dávila, Clínica Indisa, Red UC Christus, Biobanco de Tejidos y Fluidos, Hospital Padre Hurtado, Hospital Regional de Concepción, Hospital Regional de Talca, Hospital de Puerto Montt, Hospital San Juan de Dios, Hospital Santiago Oriente Doctor Luis Tisné Brousse, Instituto Nacional del Cáncer, Hospital del Salvador, Hospital Regional de Coquimbo, Hospital Regional de Arica, Hospital Clínico San Borja Arriarán).

DNA Extraction, Quantification, and Sample Quality Control
DNA from frozen tissues was extracted using the QIAamp DNA Mini Kit (Qiagen, Germantown, MD, USA). FFPE tissue DNA was extracted using GeneJet FFPE DNA Purification Kit and RecoverAll™ Total Nucleic Acid Isolation (Invitrogen, Thermo Fisher Scientific, Carslbad, CA, USA), following the manufacturer's instructions, with overnight lysis instead of the suggested 1-2 h for FFPE tissue. Germline DNA was purified from whole blood samples or buffy coat using the Wizard ® Genomic DNA Purification Kit (Promega, Madison, WI, USA), according to manufacturer's protocol.
Purified DNA was quantified using the Qubit(™) dsDNA HS Assay and Quant-IT(™) Picogreen ® dsDNA Reagent Kit (Invitrogen, Thermo Fisher Scientific, Carslbad, CA, USA). The purity of DNA was assessed by measuring the 260/280 nm absorbance ratio. For FFPE samples, fragment length and degradation were assessed using the HS Genomic DNA Analysis Kit (DNF-488) in a Fragment Analyzer (Agilent, formerly Advanced Analytical). DNA ranged from >1000 bp to 200 bp. Samples with <200 bp are not recommended for processing with the TumorSec workflow.

Library Preparation
Then, 100-150 ng of DNA (blood and frozen tissues) and 200 ng of DNA (FFPE) were used as input for sequencing library preparation. NGS libraries were prepared using KAPA HyperPlus Library Preparation Kit (Kapa Biosystems, Cape Town, South Africa). A double size selection was performed in libraries prepared with DNA from frozen tissue and blood and a single size selection for DNA from FFPE. Libraries were quantified using the QubitTM dsDNA HS Assay Kit (Invitrogen, Thermo Fisher Scientific, Carslbad, USA) and Quant-IT TM Picogreen ® dsDNA Reagent Kit (Invitrogen). The quality of the amplified library was checked by measuring the 260/280 absorbance ratio and fragment's length, using the HS NGS Analysis Kit (DNF-474) in a Fragment Analyzer (Agilent, formerly Advanced Analytical).

Target Enrichment
Prepared DNA libraries (1200 ng total mass) were captured by hybridization probes (Roche NimbleGen SeqCap EZ. Roche, Pleasanton, USA). The number of samples used for pre-capture multiplexing was based on sample type: six were pooled for FFPE and blood samples, while fresh frozen tumor samples were pooled in reactions of four. Captured libraries were assessed for concentration and size distribution to determine molarity.

Sequencing Run Set-Up
Libraries were diluted to a concentration of 4 nmol/L and processed for sequencing, according to the manufacturer's instructions (Illumina, San Diego, CA, USA). The final captured library concentration for sequencing was 9.4-9.5 pM. Libraries were sequenced in an Illumina ® MiSeq System using paired-end, 300 cycles (MiSeq Reagent Kits v2, Illumina ® Illumina, San Diego, CA, USA).

Bioinformatic Analyses
The bioinformatic pipeline is summarized in Supplementary Figure S1. Filtering of reads and base correction were performed with the fastp v0.19.11 tool. The filtered reads align with the reference genome GRCh37/hg19 using Burrows-Wheeler Alignment mem (BWA mem v0.7.12). MarkDuplicates tool of Picard v2.20.2-8 was applied to identify duplicates. To reduce the number of mismatches to the reference genome, the reads were realigned with RealignerTargetCreator and IndelRealigner from GATK v3.8 [19]. Finally, the quality scores were re-calibrated with the combination of GATK's BaseRecalibrator and PrintReads tools [19].

Variant Filtering and Sequence Quality Reporting
Variants with allele frequencies greater than 0.5 and with an altered allele depth ≥12 reads were selected. These thresholds were established as the limit of detection (LOD) for the NGS TumorSec panel following the recommendations of the Association for Molecular Pathology and College of American Pathologists [28]. Polymorphisms were eliminated, discarding alleles reported in 1000 Genomes, ESP6500, GnomAD, or ExAC [27] with a frequency greater than 0.01. Filtering was extended to include all under-represented populations that had information in GnomAD and ExAC. Additionally, a dataset containing genetic variants in Chilean individuals was used for further filtering.
The bioinformatics pipeline was executed automatically, creating pdf reports that allowed an easy view of quality metrics per sample. For this purpose, the programs FastQC v0.11.8, Qualimap v2.2.2a, Mosdepth v0.2.5, and MultiQC v1.8 were executed between the pre-processing of the bioinformatics workflow. The main metrics are the number of initial raw reads, the percentage of filtered reads, the duplication rate, the number of reads on target regions, the average depth on-target regions, the uniformity percentage, and the ratio of on-target regions with a minimum coverage of 100× to 500×. For variant calling and annotation, the coverage threshold for FFPE and FF was set at 300× in at least 80% of target regions.

Germline Variant Calling
Data were pre-processed following the pipeline shown in Supplementary Figure S1. Variant calls were made using the GATK HaplotypeCaller tool. A minimum confidence threshold of 30 was set for variant calling. Additionally, a coverage threshold was set at 200× in at least 80% of target regions. Finally, a variant calling hard-filter for SNPs and indels was applied separately following GATK recommendations [29].  Figure S2).

Panel Performance
The panel's performance was calculated using the reference HD200 (Horizon Discovery) standard FFPE sample containing characterized mutations in the following genes: BRAF, KIT, EGFR, KRAS, NRAS, PIK3CA, ARID1A, and BRCA2. As observed in Figure 1A, the assay captured all 13 positive variants. A 0.98 coefficient of correlation (r-squared) was extrapolated with a p-value of 3.221 × 10 −10 between expected variant allele frequencies (VAF) from the positive control and those reported by the assay. VAFs ranged from 24.5% to as low as 1%, showing the assay's high analytical sensitivity.
was extrapolated with a p-value of 3.221 × 10 −10 between expected variant allele frequencies (VAF) from the positive control and those reported by the assay. VAFs ranged from 24.5% to as low as 1%, showing the assay's high analytical sensitivity.
Additionally, the panel's performance for detecting BRCA1/2 germline mutations, which are predictive biomarkers for PARP inhibitor therapy in breast, ovarian, and prostate cancer, was tested (Table 1). Thus, references DNA HD793 and HD794 (Horizon Discovery), which contain known germline variants in BRCA1 and BRCA2 at VAF of 50 and 100%, were sequenced. Figure 1B shows that 11 out of 11 reported variants were detected at the expected VAF. The correlation coefficient between expected and reported VAF is 0.99 with a p-value of 2.2 × 10 −16 . Importantly, no mutations were detected in the 15 positions reported as "no-mutated" (true negatives), showing the assay's high specificity. For reproducibility assessment, three FFPE samples from different tumors (colorectal, ovary, and gallbladder) were used to prepare two separate libraries each. All samples passed the sequencing metrics threshold with ≥300× coverage in 99.6% of target regions and 89% uniformity. A 100% concordance among non-synonymous variants detected in the different libraries was observed (Figure 2A).
Inter-runs repeatability was assessed using four FFPE samples (three ovaries and one HD control). Different libraries were sequenced in different runs. Reproducibility of sequencing metrics (94% of target regions with ≥300× coverage and ≥87% uniformity) and concordance of detected variants were also observed ( Figure 2B). One ovarian FFPE sample was assessed in three different library preparations and three separate sequencing runs ( Figure 2C). A high correlation (r = 0.99) was observed between VAFs called in all the different settings ( Figure 2D). Additionally, the panel's performance for detecting BRCA1/2 germline mutations, which are predictive biomarkers for PARP inhibitor therapy in breast, ovarian, and prostate cancer, was tested (Table 1). Thus, references DNA HD793 and HD794 (Horizon Discovery), which contain known germline variants in BRCA1 and BRCA2 at VAF of 50 and 100%, were sequenced. Figure 1B shows that 11 out of 11 reported variants were detected at the expected VAF. The correlation coefficient between expected and reported VAF is 0.99 with a p-value of 2.2 × 10 −16 . Importantly, no mutations were detected in the 15 positions reported as "no-mutated" (true negatives), showing the assay's high specificity.
For reproducibility assessment, three FFPE samples from different tumors (colorectal, ovary, and gallbladder) were used to prepare two separate libraries each. All samples passed the sequencing metrics threshold with ≥300× coverage in 99.6% of target regions and 89% uniformity. A 100% concordance among non-synonymous variants detected in the different libraries was observed (Figure 2A).
Inter-runs repeatability was assessed using four FFPE samples (three ovaries and one HD control). Different libraries were sequenced in different runs. Reproducibility of sequencing metrics (94% of target regions with ≥300× coverage and ≥87% uniformity) and concordance of detected variants were also observed ( Figure 2B). One ovarian FFPE sample was assessed in three different library preparations and three separate sequencing runs ( Figure 2C). A high correlation (r = 0.99) was observed between VAFs called in all the different settings ( Figure 2D).

Comparison between FFPE, Fresh Frozen, and Blood gDNA
To assess whether the protocol and bioinformatic workflow for detecting somatic mutations discard FFPE-induced artifacts and germline variants, DNA from FFPE, fresh frozen (FF) tumor, and buffy coat samples from six ductal breast carcinoma subjects was sequenced.
Variants detected among all sample triads for each of the six subjects are outlined in Figure 3. Nine variants in the FF sample set were reported. Seven of these variants were also reported in the matching FFPE samples. It is worth noting that the two variants detected in an FF sample (FA6-005) were found in the FFPE sample but at frequencies <5% (the LOD established for the assay).

Comparison between FFPE, Fresh Frozen, and Blood gDNA
To assess whether the protocol and bioinformatic workflow for detecting somatic mutations discard FFPE-induced artifacts and germline variants, DNA from FFPE, fresh frozen (FF) tumor, and buffy coat samples from six ductal breast carcinoma subjects was sequenced.
Variants detected among all sample triads for each of the six subjects are outlined in Figure 3. Nine variants in the FF sample set were reported. Seven of these variants were also reported in the matching FFPE samples. It is worth noting that the two variants detected in an FF sample (FA6-005) were found in the FFPE sample but at frequencies <5% (the LOD established for the assay).
To further explore FF and FFPE samples' concordance, the allele frequencies of both synonymous and non-synonymous variants detected in both sample types were plotted ( Figure 4). Variants with AF < 5% display a low r-value (0.68, p-value of 2.479 × 10 −9 ). However, when all variants (73) are analyzed, correlation increases (r = 0.95, p-value of <2.2 × 10 −16 ). Importantly, no germline variants and no variants exclusive for FFPE samples were detected using the pipeline for somatic mutations. To further explore FF and FFPE samples' concordance, the allele frequencies of both synonymous and non-synonymous variants detected in both sample types were plotted ( Figure 4). Variants with AF < 5% display a low r-value (0.68, p-value of 2.479 × 10 −9 ). However, when all variants (73) are analyzed, correlation increases (r = 0.95, p-value of < 2.2 × 10 −16 ). Importantly, no germline variants and no variants exclusive for FFPE samples were detected using the pipeline for somatic mutations.   To further explore FF and FFPE samples' concordance, the allele frequencies of both synonymous and non-synonymous variants detected in both sample types were plotted ( Figure 4). Variants with AF < 5% display a low r-value (0.68, p-value of 2.479 × 10 −9 ). However, when all variants (73) are analyzed, correlation increases (r = 0.95, p-value of < 2.2 × 10 −16 ). Importantly, no germline variants and no variants exclusive for FFPE samples were detected using the pipeline for somatic mutations.

Validation of the Assay in Clinical Samples
To validate the assay and analysis capabilities in "real world" samples, 183 tumor biopsies from different clinical sites were processed. In total, 131 out of the 183 were successfully sequenced (116 FFPE and 15 FF): breast (14), ovary (69), gastric (1), gallbladder (31), and colorectal (16).
All variants with allelic frequency > 0.01 reported in at least one of the following four germline population variant's databases (PVDs): GnomAD, ESP6500, ExAC, and 1000 Genomes, were eliminated. However, given that the Latin American population is not well represented in these repositories, somatic mutations were initially overestimated due to their absence or low (<1%) VAF. Thus, the classification of remaining variants was performed using the algorithm depicted in Supplementary Figure S2. This algorithm was designed based on recommendations from Sukhai et al. (2019) [30] and using annotations in COSMIC, dbSNP, CLINVAR, and PVDs databases. Filtering was extended to include all under-represented populations that have information in GnomAD and ExAC. Additionally, a local genetic germline database from Chilean individuals was used. The resulting variants were classified as germline, somatic, putative germline, putative somatic, putative germline novel, or putative somatic novel.
A total of 256 protein-affecting variants were found in the 131 samples. Among these, 197 non-synonymous somatic and putative known and novel somatic variants were identified in 111 out of the 131 samples (85%) ( Table 2); moreover, 144 were unique variants. Figure 5 shows a breakdown of somatic mutations by gene and tumor type. Non-synonymous variants are shown according to mutation type (missense and nonsense mutations, frameshift causing deletions and insertions, and variants positioned in splice sites). Overall, TP53, BRCA2, PIK3CA, ARID1A, KRAS, TSC2, PTEN, and BRCA1 were the most frequently mutated genes. TP53 and PIK3CA were the most prevalent mutations in ovary cancer; TP53, BRCA2, and KRAS in colorectal cancer; and TP53 and PIK3CA in breast cancer. In 31 samples of GBC, we found 36 mutations in TP53 (16)

Identification of Biomarkers for Targeted Therapies
In total, 137 (69.5%) out of the 197 identified somatic variants are described as a biomarker for drug response, which is supported by different levels of evidence: FDA

Identification of Biomarkers for Targeted Therapies
In total, 137 (69.5%) out of the 197 identified somatic variants are described as a biomarker for drug response, which is supported by different levels of evidence: FDA guidelines (44) and clinical guidelines (9), late trials (37), early trials (119), case report (42), and (113) pre-clinical data. The affected gene and target drug associations with supporting evidence from "case reports" to "FDA guidelines" are depicted in Figure 6, which outlines: (1) the fraction of samples with reported genetic alterations; (2) the level of existing evidence for the biomarker; (3) the gene affected; and (4) the drug association (resistant or responsive). Table 3 contains a detailed description of the biomarker mutations supported by FDA and NCCN clinical guidelines found in this study in all samples.

Identification of Biomarkers for Targeted Therapies
In total, 137 (69.5%) out of the 197 identified somatic variants are described as a biomarker for drug response, which is supported by different levels of evidence: FDA guidelines (44) and clinical guidelines (9), late trials (37), early trials (119), case report (42), and (113) pre-clinical data. The affected gene and target drug associations with supporting evidence from "case reports" to "FDA guidelines" are depicted in Figure 6, which outlines: (1) the fraction of samples with reported genetic alterations; (2) the level of existing evidence for the biomarker; (3) the gene affected; and (4) the drug association (resistant or responsive). Table 3 contains a detailed description of the biomarker mutations supported by FDA and NCCN clinical guidelines found in this study in all samples. Figure 6. Predictive biomarkers for therapy response identified in the analyzed clinical tumor samples. The biomarker gene, the associated drugs, and the level of evidence supporting this association in solid tumors are depicted. Positive (drug-responsive) and negative (drug-resistant) associations are indicated by green and red dots, respectively. Information on these interactions was obtained from "The Cancer Genome Interpreter" [26]. Table 3. Biomarker mutations supported by FDA (n = 44) and NCCN guidelines (n = 9), found in the analyzed clinical samples. The associated drug and the mutation's predictive effect are indicated. In GBC samples, we found biomarker mutations in eight genes, with supporting evidence ranging from case reports to FDA and NCCN guidelines in different tumor types (Table 4). It is worth highlighting the presence of predictive biomarkers for drugs that are currently in use for treating different cancers, such as PARP, ERBB2, EGFR, PIK3CA, mTOR, and Hedgehog signaling inhibitors.

Discussion
As a result of its global adoption and implementation, the clinical utility of NGS in the field of oncology has a rapidly growing body of evidence. The ability to obtain massive amounts of genetic information from small amounts of tissue provides clear advantages for decision-making processes against cancer [2,4,31]. Nevertheless, a large portion of cancer patients around the world do not have this option readily available. This work attempts to favor the implementation of NGS in the Latin American health system by showcasing a locally developed assay, accompanied by an open-source automated analysis focused on the target population's needs.
A critical consideration for implementing NGS in low resource settings is finding a workflow compatible with low-quality, highly degraded samples. Although fresh frozen tissue is the gold standard for molecular analyses, its use in clinical practice is impractical because of its high cost and technical difficulties related to its obtainment, processing, and storing. The sample storage infrastructure found in the developed world, with dedicated −80 • C and −20 • C freezers, is too often not in the budget for many Latin American diagnostic laboratories. FFPE tissue samples are much more cost effective, as they can be stored at room temperature. However, tumor biopsies in this region are often fixed with formalin with different protocols and laboratory environments, producing varied DNA damage during and after the formalin fixation process (e.g., fragmentation, degradation, crosslinking [32,33]). DNA quality is affected by the type of formalin used for tissue fixation and the time since preservation [34], both of which vary highly in laboratories across Chile. A total of 48 FFPE samples failed to pass the DNA, library, or sequencing quality controls. Most of the failed samples were colorectal (33/40) and gallbladder (12/43) samples. Twenty-four failed colorectal samples had low on-target rates, suggesting issues with hybridization and/or capture, while the 12 gallbladder samples did not pass the library preparation quality metrics.
In general, FFPE samples showed a higher percentage of duplicates and off-target reads (Table S1). However, these characteristics do not affect sequencing results. FFPE samples have the highest mean on-target region's coverage compared to FF and BC. Removing duplicates is intended to reduce noise during the variant identification process and minimize false positives. These results suggest removing duplicates has little effect on this panel's performance at the sequencing depths of interest (~300×). As sequencing technologies continue to advance, PCR duplicate removal will become less of an issue [35].
Somatic mutation analysis and fusion detection are critical in cancer research [1,15]. In this project, we focused on creating a cost-effective regional tool with clinical relevance. As a result of our approach, a tailored DNA based panel, we lack the ability to detect relevant RNA gene alterations such as fusions. Although side-by-side DNA and RNA analysis would provide a more complete understanding of gene alteration profiles, it can be costly and time consuming. Future solutions such as total nucleic acid library preparation followed by target enrichment could provide a more comprehensive analysis in one workflow. In the meantime, this panel detects point mutations in ALK, MET, and ROS1, which have been associated with resistance to TKI.
Currently, there is no community consensus about the most appropriate variant caller for somatic mutations [36]. For this reason, the bioinformatic pipeline for variant calling used in this work incorporated six variant callers capable of producing highly accurate somatic mutation calls for both SNVs and small indels. Somatic variant callers discard germline variants by interrogating the reference genomes and population databases, such as 1000 Genomes, where Latin American genetic variation is not well represented. Thus, we implemented a more accurate bioinformatic pipeline that allowed variants' classification as somatic, germline, and putative somatic/germline. This variant calling process highlights the extra layer of difficulty Latin American researchers and clinical laboratories face due to the absence of reference genomes representative of our population in the major databases. The overestimation of somatic variants is a problem when facing the tumor of a patient from any region or ancestry without a reference genome informative of the genetic variation in that specific population. This issue is critical for therapy determinants, such as tumor mutational load, which should be carefully interpreted in these patients [37].
The addition of reference genomes representative of the population in combination with the genetic characterization of more cancers will help address questions involving the distribution of genetic alterations among different world populations.
The best approach for resolving the somatic vs. germline mutations issue is to include respective blood samples alongside biopsies. However, this raises the assay's cost per patient, which may delay the assay's implementation.
To achieve more accurate somatic variant calling, further efforts towards genetic characterization of the Latin American and other under-studied populations are needed. Building an inclusive tumor reference genome database will allow for the discovery of novel somatic mutations and non-explored correlations to the disease.
The assay was designed to meet the biomarker needs in countries with low participation in clinical trials and in which a limited number of drugs are available or currently in use in clinical practice. Among the 131 predictive biomarkers for therapies response detected by the assay, 52 are supported by evidence recognized by the FDA and NCCN clinical guidelines.
Although a small number of GBC samples was successfully sequenced, biomarker mutations in nine genes were identified. Importantly, these are predictive biomarkers for drugs that are currently in use for treating different cancers, such as PARP, ERBB2, EGFR, PIK3CA, mTOR, and Hedgehog signaling inhibitors. Since most of these drugs are available in Chile and LATAM, finding predictive biomarkers in GBC generates opportunities for specific and basket clinical trials including GBC patients from Chile and other regions in LATAM, where this disease has an unusually high prevalence.
Supplementary Materials: The following are available online at https://www.mdpi.com/article/10 .3390/jpm11090899/s1, Table S1: Sequencing metrics (median) for all samples that passed QC filters. Figure S1: General workflow of the bioinformatic pipeline for identifying somatic variants. Figure  S2: Sequencing metrics and mean coverage of targeted genes according to sample type. Figure   Institutional Review Board Statement: The study was conducted according to the guidelines of the Declaration of Helsinki, and approved by the "Comité de Ética Científico del Servicio Metropolitano de Salud Oriente", "Comité Ético Científico o de Investigación del Hospital Clínico de la Universidad de Chile" (approval number N • 17-18), and the "Comité de Ética de Investigación en Seres Humanos de la Facultad de Medicina de la Universidad de Chile".
Informed Consent Statement: Informed consent was obtained from all subjects involved in the study.