PaRDal links DNA and RNA sequencing to resolve cancer variant uncertainty

A containerized bioinformatics framework called paRDal can combine matched germline DNA and RNA sequencing data to help interpret variants of uncertain significance in hereditary cancer testing. In a study of 204 people of Czech and Slovak descent undergoing germline testing for cancer predisposition, the researchers found that RNA evidence supported pathogenicity for 12 candidate splicing variants, excluded a predicted splicing effect for 20 and left four uncertain. Ten evaluated variants were absent from ClinVar, while 15 helped resolve conflicting ClinVar entries. The findings show how transcript-level observations can complement DNA sequence analysis, although the authors emphasize that automated outputs remain evidence for classification rather than classifications themselves.

Germline DNA next-generation sequencing is central to hereditary cancer diagnostics, but its interpretation is often complicated by variants of uncertain significance, or VUS. The paper notes that VUS account for approximately 40% of variants identified by NGS and may be present in as many as 30% of patients. A VUS can indicate a sequence change whose biological effect and clinical significance are not yet established. This uncertainty can impede interpretation, particularly when DNA analysis does not reveal how a variant affects the messenger RNA produced from it. RNA sequencing offers a complementary view by capturing consequences at the transcript level, including abnormal splicing, changes in gene expression and patterns of allele-specific expression.

paRDal, short for parallel RNA/DNA analytical workflow, is designed as an end-to-end system rather than a new variant-classification algorithm. It incorporates established software including BWA-MEM2 for DNA alignment, STAR for transcriptome alignment, GATK for variant calling, VEP for annotation, OUTRIDER for expression outliers, FRASER for splicing outliers and MBASED for allele-specific expression. Raw DNA and RNA reads enter a modular pipeline implemented in Docker, with specified tools, reference files, dependencies and software versions intended to make analyses reproducible across institutional environments. The workflow used GRCh38 and Ensembl release 114, while the R/Shiny interface presented patient-level and cohort-level results in a common display.

The DNA component followed GATK Best Practices. Reads were aligned with BWA-MEM2, sorted and processed with Picard, and recalibrated using known variant sets. HaplotypeCaller generated per-sample variant files in GVCF mode before cohort-wide joint genotyping with GenotypeGVCFs. BCFtools and vt were used to decompose multiallelic variants and normalize indels. RNA reads underwent a more specialized STAR alignment procedure: a first pass identified splice junctions, cohort-wide junctions then informed updated genome indices, and a second pass mapped the data again. WASP was enabled during this second pass to reduce reference-allele bias, a step particularly important when quantifying unequal expression of alternative alleles.

After alignment, the framework searched for several complementary molecular signals. OUTRIDER identified expression outliers after filtering genes with fewer than 10 fragments per kilobase per million reads and applying a denoising autoencoder; outliers were called at an adjusted p-value below 0.05. FRASER detected aberrant splicing, requiring an adjusted p-value below 0.05, an absolute percent-spliced-in change greater than 0.1 and minimum junction coverage of five reads. Allele-specific expression was restricted to exonic heterozygous biallelic SNPs, with at least 10 reads at a site and at least five supporting each allele. MBASED evaluated gene-level imbalance with 105 simulations, and significant events required a Holm-adjusted p-value below 0.05, heterogeneity p-value of at least 0.05 and major-allele frequency below 0.4 or above 0.6.

The system also merged annotations from the two molecular layers and placed DNA and RNA metrics side by side. These included total depth, genotype quality, allelic depth and variant allele frequency, allowing users to distinguish technical artifacts from biological patterns such as monoallelic expression or transcript degradation. Its filtering engine operates in both directions: selecting a DNA or RNA variant retrieves linked transcriptomic outliers, while selecting an expression or splicing event searches for possible causal variants. For expression evidence, the interface links records through Ensembl gene identifiers and can use genomic coordinates to recover variants that lack gene annotations. For splicing events, it constructs a gene-bounded interval spanning the relevant upstream and downstream exons, allowing the search to include canonical splice sites, nearby exons and deep intronic positions without relying on a fixed distance buffer.

Performance tests showed that paRDal could process targeted DNA and RNA panels at different scales. On a server with an AMD EPYC 9654P 96-core processor and 768 GB of memory, the 204-person clinical cohort took nine hours, equivalent to 22.7 samples per hour. Processing 50 matched whole-genome and transcriptome pairs required 106.5 hours, or 0.47 samples per hour. On a workstation containing two Intel Xeon E5-2690 v2 processors and 128 GB of memory, 30 targeted-panel samples took 11 hours. These benchmarks describe processing times on the tested hardware and should not be treated as universal performance estimates. The unfiltered clinical dataset contained 98,325 unique variants across 839,163 variant-individual pairs, from which the researchers derived 5,497 candidate variants after applying frequency, annotation, clinical-significance and call-quality filters.

Several cases illustrated why joint analysis can reveal mechanisms missed by sequence prediction alone. An ATM variant classified as a VUS and predicted to cause a missense change was associated with exon 42 skipping, retention of two flanking introns and a marked imbalance between its DNA and RNA variant allele frequencies. The authors proposed that nonsense-mediated decay removed many aberrant transcripts, making the percent-spliced-in value appear milder than the underlying disruption. A deep-intronic ATM variant was linked to pseudo-exonization, alternative acceptor use and exon elongation, with allele-specific intron retention providing additional evidence. In an APC example, visual inspection showed that 46 apparent novel-junction reads were a mapping artifact caused by a common synonymous polymorphism. A second alignment pass consolidated the signal with five supporting reads into a single event supported by 51 reads.

Across APC, ATM, BRCA1 and MSH3, the two-pass alignment strategy recovered spanning reads for an average of 23% of splice junctions missed in the initial pass, while 1.7% of reads were lost. The analysis also produced 488,711 unannotated junctions, most with the canonical GT-AG intron motif and recurrence across multiple people. The cohort analysis detected 114 splicing outliers in 94 non-overlapping genes, which represented 67 distinct events. Manual review confirmed 28 as genuine aberrant splicing. The gene-bounded search retrieved a candidate DNA variant for 23 events, and the authors reported that most of the 32 events without a candidate were technical rather than biological. The study also illustrates the continuing need for expert review: an APC promoter variant associated with gastric adenocarcinoma and proximal polyposis was correctly linked to APC, but too few informative heterozygous sites were available for a reliable allele-specific expression estimate.

The framework has several limitations. Its predefined thresholds were optimized for Mendelian diagnostics and may need adjustment for other research settings, while OUTRIDER’s autoencoder performs best with cohorts of at least 60 samples. Tissue-specific expression can restrict interpretation, although integration with GTEx data helps identify signals absent from the tested tissue. Efficient nonsense-mediated decay may deplete aberrant isoforms beyond detection. Large insertions and deletions and copy-number variants are not comprehensively handled, and the authors identify dedicated Pindel and copy-number analysis as future additions. The paired individual-level data cannot be deposited publicly because they are identifiable and consent did not permit open sharing, while de-identified data are available to qualified researchers through a secure Charles University environment. Overall, paRDal provides a reproducible, technically detailed platform for combining DNA variants with measured RNA consequences, but its diagnostic value depends on cohort design, tissue selection, manual curation and laboratory interpretation.

Subject of Research: A containerized bioinformatics framework integrating germline DNA and RNA sequencing for hereditary cancer variant interpretation

Article Title: paRDal: bioinformatics framework for a parallel RNA and DNA NGS analysis in hereditary cancer diagnostics

Article References: Matejkova, K., Nehasil, P., Kleiblova, P., Ptackova, T., Cerna, M., Just, P., Vit, O., Zemankova, P., Giertlova, M., Janatova, M., Soukupova, J., Brorson, I. S., & Kleibl, Z. (2026). paRDal: bioinformatics framework for a parallel RNA and DNA NGS analysis in hereditary cancer diagnostics. Frontiers in Bioinformatics, 6, Article 1944540. https://doi.org/10.3389/fbinf.2026.1944540

Image Credits: AI Generated

DOI: 10.3389/fbinf.2026.1944540

Keywords: hereditary cancer diagnostics, germline DNA sequencing, RNA sequencing, variant classification, splice variants, multi-omics bioinformatics, clinical genomics, allele-specific expression, paRDal, bioinformatics, framework, parallel