paRDal: Bioinformatics Framework for a Parallel RNA and DNA NGS Analysis in Hereditary Cancer Diagnostics

Technical Summary: paRDal – A Bioinformatics Framework for Parallel RNA and DNA NGS Analysis in Hereditary Cancer Diagnostics

Problem Statement
Germline next-generation sequencing (NGS) of DNA is the gold standard for hereditary cancer diagnostics but faces significant limitations in clinical utility due to the high prevalence of Variants of Uncertain Significance (VUS). Approximately 40% of NGS-identified variants are VUS, many of which affect mRNA processing (splicing, nonsense-mediated decay, expression levels) in ways invisible to DNA analysis alone. While RNA NGS offers a complementary layer of evidence to resolve these variants, its clinical implementation is hindered by a lack of standardized, integrated frameworks. Existing workflows often fragment the process, separating raw data preprocessing, variant calling, and functional analysis, or relying on unidirectional approaches that fail to systematically link DNA variants to RNA outliers. Furthermore, the complexity of multi-layered data often renders it inaccessible to clinicians lacking specialized bioinformatic expertise.

Methodology
The authors present paRDal (parallel RNA/DNA analytical workflow), a containerized (Docker), modular framework designed to integrate matched germline DNA and RNA NGS data into a single diagnostic pipeline. The framework was validated on a cohort of 204 individuals of Czech and Slovak descent undergoing hereditary cancer testing, alongside external controls from the 1000 Genomes and Geuvadis projects.

Key methodological components include:

• Containerized Reproducibility: The entire workflow, from raw FASTQ processing to final visualization, is encapsulated in Docker to ensure software and dependency reproducibility.
• DNA and RNA Preprocessing: DNA data follows GATK Best Practices (BWA-MEM2, Picard, GATK HaplotypeCaller). RNA data utilizes a rigorous two-pass STAR alignment strategy. The first pass identifies novel splice junctions, which are then incorporated into the genome index for a second pass. Crucially, the second pass employs the WASP algorithm to suppress reference-mapping bias, a critical step for accurate Allele-Specific Expression (ASE) analysis.
• Variant Calling and Annotation: Both DNA and RNA variants are called and normalized. Annotation is performed using Ensembl VEP with a specialized plugin suite covering splicing predictions (SpliceAI, SpliceVault, SpliceRegion), deleteriousness scores (AlphaMissense, CADD, REVEL), and protein domain analysis.
• Functional Analysis Modules:
  • Splicing Outliers: Detected using the FRASER R-package, calculating Percent Spliced In (Ψ) with strict thresholds (adjusted p-value < 0.05, |ΔΨ| > 0.1).
  • Expression Outliers: Identified using the OUTRIDER R-package with a denoising autoencoder to correct for confounders.
  • Allele-Specific Expression (ASE): Quantified using GATK ASEReadCounter and statistically assessed via the MBASED R-package.
• Bidirectional Integration Engine: A core feature of paRDal is its bidirectionally reactive filtering system. Users can filter RNA outliers based on a selected DNA variant, or conversely, query the DNA variant pool based on an observed RNA outlier. This engine dynamically constructs biologically bounded genomic windows to link deep intronic or exonic variants to specific splicing events.
• Visualization Interface: An R/Shiny application consolidates multilayered evidence, allowing clinicians to toggle between "All Variants" (cohort-wide) and "Patient View" (individual). It integrates side-by-side DNA/RNA metrics, IGV visualization of BAM files, Sashimi plots for splicing events, and external GTEx data.

Key Results

• Performance: On a high-end server, paRDal processed 204 individuals (CZECANCA panel) in 9 hours (22.7 samples/hour). It also demonstrated robustness on standard workstations.
• Variant Resolution: From an initial 98,325 unique variants, filtering yielded 5,497 candidate causal variants.
• Splicing Landscape: In the 204-individual cohort, 292 variants mapped to splice regions. FRASER analysis identified 114 splicing outliers; 33% were linked to causal variants within the splice region, while others suggested causative variants in deep intronic sequences.
• Case Studies:
  • ATM:c.6197A>G: A VUS classified as a missense variant was re-evaluated. RNA analysis revealed severe exon skipping and intron retention, supported by significant ASE indicating transcript degradation, supporting a reclassification to pathogenic (PVS1(RNA)).
  • ATM:c.1899-123A>G: A deep intronic variant with high SpliceAI scores was confirmed to cause pseudo-exonization and exon elongation, leading to NMD.
  • APC:c.1627-8A>G: The two-pass STAR+WASP alignment was shown to rescue novel splice junctions missed in the first pass. In one case, it unified an alignment artifact (an 18-bp insertion) with a true biological event (7-bp exon elongation), increasing supporting reads ten-fold.
  • GAPPS Syndrome: The framework successfully linked a promoter variant to APC but also highlighted the limitations of automated ASE analysis when heterozygous SNP density is low, necessitating manual curation via the interface.
• Algorithmic Impact: The two-pass alignment strategy rescued spanning reads for 23% of splice junctions missed during the initial pass across four clinically actionable genes (APC, ATM, BRCA1, MSH3).

Significance and Claims
The paper claims that paRDal addresses the critical bottleneck of fragmented workflows in hereditary cancer diagnostics by providing a standardized, end-to-end solution that integrates DNA and RNA NGS. Its primary significance lies in:

1. Operationalizing ACMG/AMP Criteria: It facilitates the application of PVS1(RNA) (pathogenicity based on RNA evidence) and PP3 (computational evidence) criteria, directly aiding in the resolution of VUS.
2. Bidirectional Workflow: Unlike unidirectional tools, paRDal allows for both "variant-to-function" and "function-to-variant" analytical perspectives, enabling geneticists to anchor analysis on observed transcriptomic signals.
3. Accessibility: By consolidating complex multi-omics data into an interactive R/Shiny application, it bridges the gap between bioinformatic outputs and clinical decision-making without requiring specialized bioinformatic expertise.
4. Technical Robustness: The inclusion of two-pass alignment and WASP bias correction ensures high-fidelity detection of novel splice junctions and accurate ASE analysis, features often omitted in standard pipelines.

The authors maintain a modest tone regarding limitations, acknowledging that the workflow uses predefined parameters optimized for Mendelian diagnostics, that outlier detection requires a minimum cohort size (mitigated by public references), and that tissue-specific expression and NMD efficiency can limit detection sensitivity. They note future work will focus on integrating large indel detection (Pindel) and CNV analysis.