Whole Genome Clinical Pipeline for Cancer Predisposition

This white paper presents a comprehensive description of a clinical-grade whole genome sequencing (WGS) pipeline designed specifically for the identification and reporting of inherited cancer predisposition. The pipeline integrates high-throughput sequencing data processing, variant calling, clinical annotation, and structured interpretation, with a strong emphasis on reproducibility, scalability, and regulatory alignment.

The system has been architected to detect and classify multiple classes of genomic variation, including single nucleotide variants (SNVs), small insertions and deletions (indels), structural variants (SVs), and copy number variants (CNVs). Each variant class is processed through dedicated calling and interpretation logic, ensuring that the analytical breadth of the pipeline reflects the full continuum of clinically actionable inherited variation observed in cancer predisposition syndromes.

The pipeline leverages curated clinical databases, most notably ClinVar, alongside structured internal reference frameworks to produce clinically interpretable outputs aligned with ACMG/AMP guidelines. By combining authoritative external evidence with deterministic internal classification rules, the system delivers reports that are both scientifically defensible and consistently reproducible across samples, operators, and time.

The primary goal of the pipeline is to deliver a clinically robust, scalable, and reproducible system capable of identifying pathogenic variants associated with hereditary cancer syndromes. The design therefore prioritises analytical accuracy, deterministic behaviour, and traceability over flexibility or open-ended exploratory analysis, in keeping with the operational expectations of a regulated clinical laboratory.

Key objectives include the accurate detection of germline variants across the entire genome; the integration of curated clinical knowledge bases such as ClinVar and Ensembl; the consistent classification of variants using standardised frameworks; the clear separation of automated reporting and manual review pathways; the support of multiple inheritance models and disease groupings; and the transparent traceability of every decision made within the pipeline. Together these objectives define an analytical environment in which clinical conclusions can be reproduced, audited, and defended.

The system is explicitly designed to support clinical reporting in cancer predisposition panels, including Hereditary Breast and Ovarian Cancer (HBOC), DNA repair-associated ovarian cancer syndromes, moderate-risk cancer susceptibility genes, and prostate cancer predisposition. These disease definitions are formally encoded in the reference framework, enabling deterministic mapping between observed variants and the clinical conditions for which they are reportable.

3.1 Overview

The Whole Genome Clinical Pipeline is architected as a modular, deterministic system designed to transform raw sequencing data into clinically interpretable outputs. The architecture prioritises reproducibility, auditability, and scalability, ensuring that each processing stage produces traceable and verifiable outputs that can be independently inspected during validation, audit, or clinical review.

The pipeline consists of six core components: input data processing, alignment and pre-processing, variant calling, variant annotation, the clinical interpretation engine, and the reporting layer. Each component operates independently of the others but is tightly coupled through structured data formats and well-defined interfaces, enabling modular development, testing, and validation while preserving end-to-end determinism.

3.2 Data Ingestion and Input Handling

The pipeline begins with the ingestion of raw sequencing data, typically in the form of paired-end FASTQ files generated by high-throughput sequencing platforms. Inputs are characterised by paired-end reads (R1 and R2), high-coverage whole genome sequencing of typically thirty-fold or greater, and raw reads carrying base quality information that informs downstream alignment and variant calling.

Input validation steps include file integrity checks, read pairing validation, format verification, and basic quality assessment. These checks ensure that downstream processing is not silently compromised by malformed or incomplete input data, and provide an early opportunity to reject samples that would otherwise yield unreliable clinical conclusions.

3.3 Alignment and Pre-processing

Reads are aligned to the human reference genome (GRCh38) using a Burrows–Wheeler Transform-based aligner such as BWA-MEM, producing aligned reads in BAM format. The choice of GRCh38 as a fixed reference is itself a deliberate reproducibility decision, ensuring that genomic coordinates remain stable across samples and across time.

Following alignment, pre-processing steps standardise the alignment output and optimise it for variant calling. These steps include sorting by genomic coordinate, marking or removing duplicate reads, indexing the resulting BAM files, and optionally performing base quality recalibration. Together they produce alignment artefacts that are uniform in structure and suitable for the calling layer that follows.

3.4 Variant Calling Layer

The pipeline performs variant calling across multiple variant classes, namely SNVs, indels, structural variants, and copy number variants. This multi-class approach reflects the clinical reality that inherited cancer predisposition arises from a wide spectrum of genomic alterations, and that a pipeline restricted to a single variant class would systematically under-report clinically actionable findings.

Variant calling is typically performed using a combination of tools optimised for different variant types: haplotype-based callers for SNVs and indels, split-read and read-pair-based callers for structural variants, and read-depth and segmentation-based tools for copy number variants. Outputs are standardised into VCF files containing genomic coordinates, reference and alternate alleles, genotype information, quality metrics, and supporting evidence, providing a uniform substrate for downstream annotation and interpretation.

3.5 Annotation Engine

The annotation engine enriches raw variant calls with biological and clinical context, drawing on Ensembl for gene models, transcripts, and genomic structure; ClinVar for clinical significance; and the internal reference configuration for disease mappings and reporting rules. This combination of authoritative public sources with controlled internal configuration allows the pipeline to balance external evidence with laboratory-specific reporting policy.

Annotation outputs include gene mapping, transcript selection, variant consequence prediction, clinical classification where available, and population and functional context where integrated. The annotation layer is deterministic and driven by structured configuration files, ensuring that identical inputs reliably produce identical annotated outputs.

3.6 Reference Configuration Framework

A central component of the architecture is the reference configuration (for example, reference.json), which defines the reportable genes, disease groupings, interpretation rules, classification display logic, filtering behaviour, and reporting templates used throughout the pipeline. This configuration-driven approach ensures that pipeline behaviour is transparent and version-controlled, that updates to clinical knowledge can be incorporated without code changes, and that reporting logic remains consistent across all samples and across time.

3.7 Clinical Interpretation Engine

The interpretation engine applies structured rules to annotated variants in order to determine clinical significance, reporting status, risk classification, and disease associations. It operates using predefined classification hierarchies, gene–disease mappings, consequence-based interpretation logic, and ClinVar evidence prioritisation, producing categorical outputs spanning positive findings, carrier findings, variants of uncertain significance, manual review candidates, and negative findings. In doing so, the engine transforms genomic data into clinically actionable insights.

3.8 Reporting Layer

The reporting layer converts structured interpretation outputs into human-readable clinical reports. Each report contains a summary of findings, gene-specific results, variant-level details, clinical interpretation text, and a technical appendix. Reports are generated in a consistent format, ensuring standardised clinical communication, traceability of decisions, and alignment with regulatory expectations for genetic test reporting.

3.9 Data Flow Summary

The complete data flow can be summarised as the progression from FASTQ to BAM to VCF to annotated variants to clinical interpretation and finally to report. Each stage produces intermediate outputs that can be independently validated, enabling stepwise debugging, incremental validation, and the maintenance of full audit trails throughout the analytical lifecycle of a sample.

3.10 Scalability and Performance Design

The architecture is designed to scale across large datasets and high-throughput environments. Key design features include parallel processing at multiple stages, efficient indexing of BAM, FASTA, and VCF files, modular execution of variant classes, and a clear separation between compute-intensive and logic-intensive tasks. Together these features enable rapid processing of whole genome samples while maintaining the analytical accuracy required for clinical use.

3.11 Reproducibility and Auditability

The system enforces reproducibility through a fixed reference genome version (GRCh38), version-controlled reference configuration, deterministic processing logic, and structured logging at each stage. Auditability is achieved through full traceability of variant filtering decisions, preservation of intermediate files, logging of all pipeline steps, and explicit mapping from input data to final report outputs. These features are essential for clinical deployment and regulatory compliance.

4.1 Input Data Model

The pipeline is designed to begin with whole genome sequencing outputs derived from paired-end FASTQ files. These reads are aligned to the GRCh38 human reference genome and processed into standard genomic analysis formats, including BAM or CRAM for aligned reads, VCF or gVCF for small variant calls, dedicated VCFs for structural and copy number variants, and quality-control metrics covering coverage, read quality, mapping quality, genotype quality, and depth.

The purpose of the clinical pipeline is not merely to identify variants, but to convert raw genomic observations into clinically interpretable findings for cancer predisposition. The structured input model is therefore designed to ensure that every downstream interpretive decision can be traced back to a specific, well-characterised observation in the underlying data.

4.2 Small Variant Calling

Small variant processing focuses on SNVs and small indels and assesses these variants in the context of the genes defined in the cancer predisposition reference configuration. Each candidate small variant is characterised by chromosome, genomic position, reference and alternate alleles, gene symbol, transcript consequence, ClinVar clinical significance, genotype, depth of coverage, genotype quality, and allele balance. Clinically significant small variants are prioritised when they are located in a reportable gene and are classified as pathogenic or likely pathogenic.

4.3 ClinVar Annotation

ClinVar is used as a central evidence source for clinical significance. Variants are matched against ClinVar records using GRCh38 genomic coordinates and allele information, supporting the identification of pathogenic, likely pathogenic, uncertain, conflicting, benign, and likely benign records. The pipeline uses ClinVar evidence to distinguish reportable findings from variants that should be suppressed, filtered, or routed to manual review.

4.4 Transcript and Gene Context

Variant interpretation is performed in the context of preferred transcripts. The reference configuration specifies a preference for Ensembl canonical transcripts, with fallback to primary protein-coding transcripts where required. Transcript-aware annotation enables the pipeline to determine whether a variant is exonic, intronic, splice-site affecting, UTR-associated, non-coding, protein-altering, or protein-truncating — distinctions that are essential because the same genomic alteration may have very different biological implications depending on the transcript chosen.

4.5 Consequence Assignment

Each variant is mapped to a consequence category, and these categories are then translated into clinical interpretation phrases. Frameshift variants are interpreted as deleterious because they disrupt the reading frame; stop-gained variants are interpreted as deleterious because they introduce a premature termination codon; splice-acceptor and splice-donor variants are interpreted as likely disruptive because of their potential impact on RNA splicing; missense variants are interpreted as protein-altering and are assessed in clinical context; and synonymous variants are generally treated as non-protein-altering unless additional evidence exists. This consequence layer provides the biological bridge between variant detection and clinical reporting.

4.6 Structural Variant Calling

Structural variant processing evaluates larger genomic rearrangements that may disrupt cancer predisposition genes. Reportable structural event classes include deletions, duplications, copy-number losses and gains, partial exon deletions and duplications, and whole-gene deletions and duplications. The pipeline distinguishes between events that directly affect clinically relevant exons or genes and broad, non-specific events that lack sufficient specificity for automatic reporting, ensuring that structural calls entering the clinical report are anchored to interpretable genomic features.

4.7 CNV Interpretation

Copy-number variant interpretation is deliberately conservative. The system prioritises gene-specific or exon-specific copy-number changes over large, poorly resolved genomic segments, assessing each event against gene overlap, exon overlap, event size, event type, region specificity, and whether the event plausibly disrupts gene function. A CNV affecting all or most of a reportable tumour predisposition gene may be interpreted as a whole-gene deletion or duplication, while a CNV affecting one or more clinically relevant exons may be interpreted as a partial exon loss or gain.

4.8 Manual Review Candidate Detection

Certain findings are not automatically converted into positive clinical results; instead, they are retained as manual review candidates. These include exact ClinVar pathogenic matches in non-coding regions, ambiguous structural variants, variants with incomplete annotation, variants with conflicting evidence, and events that may be clinically relevant but fall outside automatic reporting thresholds. This manual review pathway is an important safety mechanism, helping to prevent both over-reporting and the inappropriate suppression of potentially meaningful findings.

4.9 Filtering Logic

Filtering is applied to reduce false positives and ensure that reported findings are clinically defensible. Common filtering categories include low quality, insufficient gene overlap, broad non-specific CNV, non-coding variant without an exact ClinVar match, benign or not actionable, and manual review only. Every filtered event remains traceable so that validation, audit, and clinical review processes can reconstruct precisely why a variant was excluded from the final report.

4.10 Output of the Annotation Layer

The output of the variant calling and annotation workflow is a structured set of candidate findings, each containing the gene, disease association, variant type, genomic coordinates, transcript consequence, clinical significance, quality metrics, classification, reporting status, and, where applicable, a filtering reason. This structured output is then passed into the reporting engine, where it is transformed into patient-facing and clinician-facing report sections.

5.1 Classification Model

Variant interpretation is governed by a structured classification system aligned with ACMG/AMP principles. Each detected variant is mapped into one of several categories: pathogenic, likely pathogenic, variant of uncertain significance, carrier, manual review, structural review, or negative. These categories are not merely classification labels; they are tied directly to reporting behaviour and patient-facing interpretation text, with explicit display semantics defined for each category, including severity, badge, and summary messaging.

For example, pathogenic variants trigger a positive status with high severity and a "High Risk" label, accompanied by interpretation text indicating an increased inherited disease risk, while variants of uncertain significance trigger a VUS status with interpretation text indicating that the available evidence is insufficient to establish clinical significance. This deterministic mapping ensures that classification outcomes are reproducible across all reports, regardless of operator or time of analysis.

5.2 Disease Mapping and Gene Panels

The pipeline defines disease entities as structured objects linking genes to clinical conditions, with each disease definition including a disease code, display name, gene list, clinical summary, and default interpretation phrases. Example disease mappings include HBOC_BRCA1 (BRCA1, hereditary breast and ovarian cancer), HBOC_BRCA2_PALB2 (BRCA2 and PALB2, multi-cancer predisposition), OVARIAN_DNA_REPAIR (BRIP1, RAD51C, RAD51D), and HEREDITARY_PROSTATE_CANCER (HOXB13, ATM, CHEK2). These mappings allow the pipeline to translate variant-level findings into clinically meaningful disease-level interpretations.

5.3 Gene-Level Metadata

Each gene is associated with structured metadata describing its functional role, inheritance model, clinical relevance, and associated conditions. BRCA2, for example, is annotated with a role in homologous recombination DNA repair and associations with breast, ovarian, pancreatic, and prostate cancers; PALB2 is annotated as the BRCA2 partner in DNA repair; and ATM is annotated as a DNA damage response kinase. This metadata is used to enrich reports and to ensure consistent clinical context wherever a gene is referenced.

The pipeline assigns biological meaning to variants through consequence classification. Each variant is mapped to a predefined consequence type with associated clinical and patient-facing descriptions, ensuring that the same biological event is reported in the same language regardless of when or in which sample it is observed.

Frameshift variants are classified as deleterious and interpreted as disrupting the reading frame; stop-gained variants are classified as deleterious and interpreted as causing premature protein truncation; missense variants are classified as protein-altering and interpreted as amino acid substitutions; and splice-site variants are classified as likely disruptive and interpreted as altering RNA splicing. Structural variants are interpreted in parallel terms, with copy losses interpreted as deletions and copy gains as duplications. These mappings are predefined and ensure consistent interpretation across all samples.

By centralising consequence interpretation in a single, version-controlled mapping, the pipeline avoids the well-documented risk of inconsistent terminology between samples, between operators, and between revisions of the underlying calling tools. The consequence layer is therefore both a biological abstraction and a quality control mechanism in its own right.

7.1 Structural Variant Classification

Structural variants are categorised using predefined templates spanning deletions, duplications, copy-number losses and gains, partial exon events, and whole-gene events. Each class has associated interpretation phrases for both clinical and patient-facing outputs, ensuring that reports communicate structural findings in language consistent with the rest of the report and appropriate to the audience.

7.2 CNV Interpretation Strategy

The pipeline adopts a deliberately conservative approach to CNV reporting. Large, broad CNVs are filtered unless they are clinically specific; gene overlap thresholds are enforced; partial exon and full gene impacts are distinguished from one another; and non-specific CNVs are flagged but not automatically reported. As stated within the reporting framework, broad CNV segments are handled conservatively while exact ClinVar pathogenic matches are retained for review, ensuring that automatic reporting is reserved for events with adequate biological resolution.

7.3 Manual Review Pathway

Variants are routed to manual review when non-coding regions contain exact ClinVar matches, when the available evidence is insufficient for automated classification, or when structural events are ambiguous in their genomic footprint. This pathway preserves clinical safety by ensuring that complex or borderline findings are evaluated by a qualified reviewer rather than being silently included or excluded by automated logic.

8.1 Variant Filtering

Variants are filtered on the basis of quality metrics such as genotype quality, depth of coverage, and allele balance, alongside clinical relevance, gene overlap, functional consequence, and the strength of supporting ClinVar evidence. The combination of technical and clinical filters ensures that the variants entering the interpretation layer are both analytically reliable and clinically meaningful.

8.2 Filtering Reasons

Each excluded variant is assigned a structured reason — for example, low quality, insufficient gene overlap, broad non-specific CNV, non-coding without ClinVar match, or benign or not actionable. The retention of explicit, machine-readable filtering reasons provides auditability and traceability for every filtering decision, allowing reviewers to reconstruct the precise rationale for excluding any given variant from the final report.

9.1 Transcript Selection

The pipeline prioritises transcripts using the Ensembl canonical transcript wherever available, with fallback to the primary protein-coding transcript when required. This deterministic transcript prioritisation ensures consistent variant interpretation across genes and across samples, and avoids the analytical drift that can occur when transcript selection is left to ad-hoc operator judgement.

9.2 Example: BRCA2 Gene Structure

The BRCA2 gene is represented within the pipeline with full genomic coordinates and exon structure, located on chromosome 13 between positions 32,315,508 and 32,400,268 on the positive strand, and containing multiple exons spanning coding and non-coding regions. This precise structural representation enables exon-level CNV detection, accurate splice-site interpretation, and the consistent assignment of functional consequence to variants that fall within or near defined functional domains.

10.1 Report Structure

Each clinical report includes a summary classification, gene-level findings, variant-level details, clinical interpretation, and a technical appendix. This structure is intentionally uniform across reports, providing clinicians with a predictable and easily navigated document and supporting downstream workflows such as multidisciplinary review and longitudinal patient management.

10.2 Report Sections

Standardised report sections include the clinical history analysis, additional findings such as variants of uncertain significance, variant classification explanations, and a technical appendix detailing the analytical provenance of the report. These standardised sections ensure consistency of presentation and support regulatory readiness by providing a complete and reproducible record of the analytical decisions underlying each clinical conclusion.

10.3 Output Status Logic

Final report statuses include positive, carrier, review, structural review, VUS, and negative. These statuses are ordered and prioritised according to predefined display rules, ensuring that the most clinically significant finding for a given sample is surfaced prominently and that the overall report status accurately reflects the highest-impact category of finding identified during analysis.

The pipeline is validated using a combination of synthetic and real-world datasets, including ClinVar-derived pathogenic variants, BAMSurgeon synthetic variant insertion, gene-level FASTQ generation, and truth-set comparison. This combination provides controlled ground truth for analytical validation while retaining the realism necessary to evaluate end-to-end performance against representative clinical samples.

Validation spans all major variant classes, including SNVs, indels, structural variants, and CNVs, with each generated dataset accompanied by an expected outcome that enables deterministic comparison between the pipeline's output and the predefined truth. The validation strategy is described in greater technical depth in the companion white paper on the Validation Tools framework, which forms an integral part of the analytical evidence base for this pipeline.

The pipeline is optimised for high-throughput processing, parallel execution, and large-scale genome analysis. These characteristics are essential for clinical environments in which sample volume, turnaround time, and analytical consistency must be reconciled within a single operational framework.

Key architectural characteristics include a distributed compute architecture, optimised input and output handling, and efficient indexing of alignment and reference resources such as BWA and FASTA indices. Together these features enable rapid turnaround for clinical-grade WGS while preserving the deterministic behaviour required for reproducible reporting.

The system is designed with explicit consideration for the Clinical Laboratory Improvement Amendments (CLIA), the College of American Pathologists (CAP) inspection framework, and the ACMG/AMP variant classification guidelines. Each of these frameworks imposes specific expectations regarding analytical validation, documentation, and reporting consistency, and the pipeline architecture has been shaped to satisfy those expectations as a matter of routine operation rather than retrospective effort.

Key compliance features include a full audit trail across every processing stage, deterministic outputs, documented filtering logic, validation traceability, and version-controlled reference data. These features collectively enable the laboratory to demonstrate, on inspection or on demand, that any clinical conclusion can be reconstructed from the underlying data and the documented analytical configuration in force at the time of reporting.

14.1 Strengths

The first principal strength of the pipeline is its fully deterministic interpretation framework. By encoding classification, disease mapping, and reporting logic in a version-controlled configuration and by avoiding undocumented heuristics, the system ensures that identical inputs reliably produce identical outputs, satisfying a foundational expectation of clinical analytical software.

A second strength lies in the strong integration with curated clinical databases, most notably ClinVar and Ensembl. By anchoring variant interpretation to authoritative external evidence, the pipeline ensures that clinical conclusions are grounded in the broader genomic community's understanding of pathogenicity rather than in isolated, laboratory-specific judgements.

A third strength is the robust handling of multiple variant classes, encompassing SNVs, indels, structural variants, and CNVs. This breadth of analytical coverage reflects the clinical reality of inherited cancer predisposition, in which clinically actionable variation spans the entire continuum from point mutations to large genomic rearrangements, and ensures that the pipeline does not systematically under-report any major variant class.

A fourth strength is the clear separation of automated decisions from those routed to manual review. By defining explicit criteria for manual review candidates, the system protects against both over-reporting of borderline findings and the silent suppression of potentially meaningful variants, providing a structured and defensible interface between automated analysis and expert clinical judgement.

A fifth strength is the highly scalable architecture, which supports parallel execution and efficient resource utilisation across large sample volumes. Scalability is a pragmatic as well as an analytical virtue: it ensures that the analytical rigour described elsewhere in this document can be sustained at the throughput required for routine clinical operation.

14.2 Limitations

A first limitation of the present system is that CNV modelling remains deliberately conservative. Although this conservatism is appropriate for clinical safety, it implies that certain low-resolution or genomically diffuse copy-number events may not be reported automatically, and that some clinically relevant CNVs will require manual evaluation rather than benefiting from end-to-end automation.

A second limitation is the continued reliance on manual review for complex non-coding variants. While the manual review pathway is a deliberate clinical safeguard, it nevertheless represents an analytical boundary beyond which the pipeline's automated logic does not extend, and the throughput of the system for such variants is therefore constrained by the availability of qualified human reviewers.

A third limitation is the dependence of the pipeline on the completeness and currency of external databases, particularly ClinVar. Although this dependence is unavoidable for any clinically anchored interpretation system, it does mean that the pipeline's interpretive reach is bounded by the present state of the curated clinical evidence base, and that variants lacking representation in those resources may be classified more conservatively than their underlying biology would ultimately warrant.

14.3 Future Enhancements

A first planned enhancement is the improvement of CNV modelling using more sophisticated read-depth methods. By incorporating refined depth-based segmentation and more granular event characterisation, the pipeline will be positioned to extend automated CNV reporting to a broader range of clinically relevant copy-number events while preserving the conservative posture currently adopted for ambiguous calls.

A second planned enhancement is the integration of population frequency datasets such as gnomAD. The systematic incorporation of population allele frequencies will support more nuanced filtering of common variation, improve the classification of variants of uncertain significance, and bring the interpretive logic of the pipeline into closer alignment with current ACMG/AMP recommendations on the use of population frequency evidence.

A third planned enhancement is the introduction of machine learning-based variant prioritisation. Such methods will be evaluated as adjuncts rather than replacements for the existing deterministic logic, with the explicit aim of improving the ranking of candidate variants for manual review while preserving the reproducibility and auditability that define the present analytical framework.

A fourth planned enhancement is the deployment of real-time pipeline monitoring dashboards. By exposing throughput, runtime, and quality metrics through a dedicated operational interface, the laboratory will gain continuous visibility into the analytical state of the pipeline, enabling earlier detection of performance regressions and providing an additional layer of operational evidence to support regulatory inspection.

This whole genome clinical pipeline represents a comprehensive, scalable, and clinically aligned system for the detection of inherited cancer predisposition. By combining structured clinical knowledge, high-performance genomics processing, and deterministic interpretation rules, the system achieves both technical accuracy and clinical utility within a single, coherent analytical framework.

The integration of Ensembl genomic context, ClinVar pathogenic evidence, and structured reporting frameworks ensures that findings are both scientifically robust and clinically actionable. Taken together, the architectural, interpretive, and operational features described in this paper position the pipeline as a defensible foundation for clinical genomic reporting in the cancer predisposition domain, and as a stable platform on which future analytical and interpretive enhancements can be incrementally deployed.

1. Richards, S. et al. (2015). Standards and guidelines for the interpretation of sequence variants. Genetics in Medicine.
2. Landrum, M.J. et al. (2020). ClinVar: improving access to variant interpretations. Nucleic Acids Research.
3. Yates, A.D. et al. (2020). Ensembl 2020. Nucleic Acids Research.
4. Li, H. & Durbin, R. (2009). Fast and accurate short read alignment with Burrows–Wheeler transform. Bioinformatics.
5. Ewing, A.D. et al. (2015). Combining tumor genome simulation with crowdsourcing to benchmark somatic mutation detection. Nature Methods.