Rapid minigene workflow for functional reclassification of splicing variants in hereditary cancer diagnostics

Abstract

Background

Next-generation sequencing of cancer predisposition genes is routinely used in hereditary cancer diagnostics. However, a substantial fraction of detected variants remains clinically unresolved. Using a customised 77-gene panel, we analysed 2142 individuals and identified 384 pathogenic or likely pathogenic variants across 54 genes, corresponding to a diagnostic yield of approximately 18%. Despite this, 17% of cases carried variants of uncertain significance, many of which were suspected to affect pre-mRNA splicing and are particularly challenging to interpret due to the limited reliability of in silico predictions and lack of experimental evidence.

Methods

To address this diagnostic gap, we developed a streamlined minigene-based workflow for rapid functional evaluation of splicing variants and applied it retrospectively. The approach relies on synthetic DNA and recombination-based cloning, eliminating the need for patient-derived RNA and enabling efficient construct generation within a clinically compatible timeframe. Computational prioritisation using AlphaGenome was integrated to support variant selection, while experimental assays provided direct evidence of splicing outcomes.

Results

Application of this strategy allowed the reclassification of previously unresolved variants and clarified cases with discordant computational evidence. Importantly, the workflow is designed for implementation in routine diagnostic settings, with a turnaround time aligned with clinical reporting requirements.

Conclusion

This approach provides a robust and scalable framework for functional interpretation of splicing variants, improving diagnostic resolution and supporting more informed clinical decision-making in hereditary cancer genetics.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Variants predicted to alter pre-mRNA splicing are a major source of uncertainty in hereditary cancer diagnostics. Although computational prediction tools have improved variant prioritisation, functional validation remains essential, and patient-derived RNA is often unavailable in routine clinical practice.

WHAT THIS STUDY ADDS

• We developed a rapid recombination-based minigene workflow using synthetic DNA that enables functional assessment of spliceogenic variants within clinically compatible turnaround times and supports evidence-based American College of Medical Genetics and Genomics/Association for Molecular Pathology variant reclassification.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• This workflow provides a practical and scalable strategy for integrating functional splicing assays into routine hereditary cancer diagnostics, reducing uncertainty in variant interpretation and improving clinical decision-making.

Introduction

Hereditary cancer syndromes arise from germline pathogenic variants in cancer predisposition genes (CPGs) involved in key cellular processes, including DNA repair, cell-cycle control and genome stability.^1–3 The widespread adoption of next-generation sequencing (NGS) has transformed hereditary cancer diagnostics by enabling simultaneous analysis of multiple CPGs and improving diagnostic yield through multigene panel testing.^4–6 Identifying hereditary cancer syndrome has important clinical implications. Germline pathogenic variants can guide treatment selection, surgical management and surveillance strategies and enable predictive testing of at-risk relatives.² Alterations in specific pathways also inform targeted therapies; for example, mismatch repair (MMR) deficiency predicts response to immune checkpoint inhibitors, while BRCA1/2 pathogenic variants guide PARP inhibitor use and risk-reducing interventions.^{7 8} Cascade testing further identifies carriers who benefit from intensified prevention and surveillance while reassuring non-carriers whose cancer risk approximates that of the general population.⁹ Despite these advances, the interpretation of NGS results remains a major challenge. A substantial proportion of detected variants are classified as variants of uncertain significance (VUS), limiting clinical utility and complicating patient management. Recent studies indicate that approximately one-third of tested individuals without a pathogenic or likely pathogenic (LP/P) finding harbour at least one VUS,¹⁰ while nearly 40% of hereditary cancer-associated variants currently deposited in ClinVar are classified as VUS.¹¹ Although VUS are common in hereditary cancer testing, the individual variants within this category are frequently rare or novel and often lack sufficient functional, segregation or literature evidence to support definitive classification.^12–14 This issue is especially pronounced for variants predicted to affect pre-mRNA splicing. Recent hereditary cancer studies have shown that approximately 5–15% of VUS are predicted to affect RNA splicing,^{15 16} representing a substantial source of diagnostic uncertainty in clinical genetic testing.^{16 17} Computational tools are routinely employed to predict the impact of splicing alterations, although their accuracy remains inconsistent, particularly for non-canonical splice-site variants and variants near exon–intron boundaries.^18–21 Widely used algorithms include MaxEntScan, NNSplice, Human Splicing Finder (HSF) and SpliceSiteFinder, whereas more recently developed deep learning-based models, such as SpliceAI and AlphaGenome, have substantially improved the prediction of splice-disrupting variants.^{19 22} Nevertheless, computational predictions alone remain insufficient for definitive variant classification and require complementary experimental validation.¹⁷ In addition, splice-disrupting variants frequently occur outside canonical splice sites, and exonic variants initially annotated as missense or synonymous may alter splicing regulatory elements. RNA analyses using patient-derived samples can provide direct evidence of splicing defects but are often limited in diagnostics by tissue availability, RNA quality and variable gene expression.^23–25 Minigene assays represent an established alternative for functional splicing analysis, enabling controlled evaluation of variant effects in a heterologous system.^{26 27} However, conventional workflows are often labour intensive and poorly compatible with diagnostic turnaround times. In this study, we present an optimised minigene approach that combines synthetic DNA with recombination-based cloning, enabling rapid functional assessment of splicing variants in hereditary cancer genes, without reliance on patient-derived RNA or specialised instrumentation. AlphaGenome²² was used as a complementary artificial intelligence (AI) tool to prioritise candidate variants for experimental testing. Applied retrospectively to a large diagnostic cohort analysed with a custom 77-gene cancer predisposition panel, this integrated approach provides a practical framework for resolving clinically relevant splicing variants with uncertain or conflicting interpretations.

Materials and methods

Detailed experimental procedures, including the complete NGS panel composition, variant annotation and classification, cell culture, transfection, RNA extraction and RT-PCR, are provided in the online supplemental materials and methods.

Study cohort and clinical data collection

Individuals referred for genetic testing due to suspected hereditary cancer predisposition were included. Evaluation followed routine clinical practice based on personal and/or family history suggestive of an inherited cancer syndrome. Clinical data, including tumour type, age at diagnosis and family history, were collected during genetic counselling and stored in dedicated databases. All participants provided written informed consent for genetic testing and use of anonymised data for research.

NGS panel design, sequencing and variant detection

Genomic DNA was extracted from peripheral blood using the Maxwell system (Promega, Madison, Wisconsin, USA) according to the manufacturer’s instructions. Targeted sequencing employed a custom panel of 77 CPGs with the SureSelect Agilent Hereditary Cancer Solution (Agilent Technologies, Santa Clara, California, USA) and sequenced on the Illumina MiSeq platform (Illumina, San Diego, California, USA). Library preparation followed manufacturer protocols, including enzymatic DNA fragmentation, adapter ligation, indexing, amplification and target enrichment by hybrid capture. Library quality and concentration were assessed using Agilent TapeStation 4150 (Agilent Technologies) and Qubit fluorometer (Thermo Fisher Scientific, Waltham, Massachusetts, USA). Equimolar libraries were pooled and loaded onto MiSeq at the recommended concentrations. FASTQ files were processed using SOPHiA Genetics DDM software (SOPHiA Genetics, Lausanne, Switzerland), aligned to the GRCh37/hg19 reference genome. The bioinformatic pipeline included read alignment, variant calling and annotation of single-nucleotide variants (SNVs) and small insertions/deletions.

Variant annotation, classification and confirmation

Variants were classified following American College of Medical Genetics and Genomics/Association for Molecular Pathology (ACMG/AMP) guidelines. Interpretation integrated population frequency (gnomAD), disease-specific and gene-specific databases (ClinVar, dbSNP, ClinGen, COSMIC, UCSC), biochemical parameters (Grantham distance, BLOSUM62), in silico predictions (Alamut Visual Plus, SOPHiA Genetics; Franklin, Genoox, Tel Aviv, Israel; VarSome, Saphetor, Lausanne, Switzerland), literature and functional studies. Variants classified as pathogenic (class 5), likely pathogenic (class 4) or of uncertain significance (class 3) and clinically relevant were confirmed by Sanger sequencing on independent DNA extractions.

In silico splicing prediction tools

Variants predicted to affect splicing were assessed using MaxEntScan, NNSplice, HSF, SpliceSiteFinder and SpliceAI. Predictions were qualitatively interpreted by comparing wild-type and variant scores and by evaluating concordance across tools.

AlphaGenome analysis

Selected variants were further analysed with AlphaGenome (DeepMind Technologies, London, UK), an AI framework predicting variant effects on gene expression and splicing.²² AlphaGenome integrates splice-site strength, usage and junction modelling to produce a composite splicing impact score. Composite scores and splicing-related signal tracks were interpreted in the context of other in silico predictions and functional data but were not used as standalone evidence for variant classification.

Minigene construct design and cloning strategy

Functional splicing analysis employed the pSpliceExpress vector (provided by Professor Stefan Stamm).²⁸ Genomic fragments covering the exon of interest and ~100–150 bp flanking intronic regions were PCR amplified from patient DNA using attB-containing primers or synthesised (Eurofins Genomics, Ebersberg, Germany). Inserts were cloned into pSpliceExpress via Gateway recombination (Thermo Fisher Scientific). All constructs were verified by Sanger sequencing before functional assays.

Results

NGS analysis of CPGs and identification of candidate splicing variants

Individuals with suspected hereditary cancer predisposition were analysed using a consolidated NGS diagnostic workflow based on a custom panel of 77 CPGs, organised into specific subpanels. This design enables the detection of SNVs, small insertions/deletions and CNVs in genes associated with established or emerging hereditary cancer syndromes. Between 2020 and 2025, a total of 2142 patients were analysed using this panel. Genes were grouped according to clinical indication, including gastrointestinal tumours, breast-ovarian-pancreatic cancer, endocrine tumours, RASopathies, familial melanoma, renal tumours, PTEN-related syndromes, brain tumours, Li-Fraumeni syndrome and others (figure 1A). An LP/P variant was identified in 18% of cases (figure 1B), corresponding to 384 LP/P variants distributed across 54 genes. The most frequently affected genes are shown in figure 1C. Identification of LP/P variants had a direct clinical impact, informing cancer risk assessment, surveillance strategies and therapeutic management. Despite this, VUS were detected in 17% of individuals. Notably, approximately 26% of annotated variants classified as pathogenic, likely pathogenic or VUS were predicted to affect pre-mRNA splicing (figure 1D). Of the 196 candidate spliceogenic variants identified, 88 remained unresolved after routine diagnostic interpretation despite integration of clinical information, segregation analysis (when available), database and literature review and multiple in silico prediction tools. These unresolved variants constituted the primary pool for functional assessment. From this group, eight variants were prioritised for minigene analysis based on their clinical relevance, the uncertainty of computational predictions and the feasibility of construct generation (online supplemental figure 1).

Figure 1. Overview of the NGS diagnostic cohort and identification of candidate splicing variants. (A) Distribution of patients tested between 2020 and 2025 according to the clinical indication that guided selection of the corresponding virtual diagnostic subpanel within the hereditary cancer testing workflow. The categories shown represent clinical indications and hereditary cancer syndromes rather than individual tumour types. (B) Classification of variants identified in 2142 NGS tests, showing proportions of LP/P, VUS and LB/B variants. (C) Distribution of LP/P variants across the 20 most frequently affected CPGs within the panel. (D) Distribution of variants classified as LP/P or VUS according to their predicted molecular consequences. Variants predicted to affect pre-mRNA splicing are shown separately from missense, nonsense and other variant categories; overlaps indicate variants with multiple predicted functional effects. CNS, central nervous system tumours; CPGs, cancer predisposition genes; ENDO, endocrine tumours; FM, familial melanoma; GI, gastrointestinal cancer predisposition; HBOC, breast, ovarian and pancreatic cancer predisposition; LB/B, benign or likely benign; LFS, Li-Fraumeni syndrome; LP/P, pathogenic or likely pathogenic; NGS, next-generation sequencing; OTH, other indications; PTEN, PTEN-related syndromes; RAS, RASopathies; REN, renal tumours; VUS, variants of uncertain significance.

Rapid functional assessment of splicing variants using an optimised recombination-based minigene workflow

To address the diagnostic uncertainty associated with candidate splicing variants, we performed a retrospective functional analysis using a rapid and optimised minigene-based workflow (online supplemental figure 2). This strategy enables splicing assessment independently of tissue availability or RNA quality. As summarised in figure 2, selected variants were prioritised for functional testing when classification remained unresolved after clinical, bioinformatic and segregation analyses. For each variant, a genomic fragment containing the exon of interest and its flanking intronic regions was generated either by PCR amplification from patient genomic DNA using attB-containing primers or by synthesis as a custom DNA fragment. The attB-flanked inserts were cloned into the pSpliceExpress splicing reporter vector via Gateway recombination, enabling rapid single-step recombination with the vector attP sites.²⁸ In the resulting constructs, the test fragment is positioned between two constitutively spliced rat insulin exons, providing a standardised context that allows detection of exon inclusion, exon skipping or activation of alternative splice sites. After bacterial transformation and plasmid amplification, constructs were sequence verified and transiently transfected into HEK-293T cells. Total RNA was extracted from transfected cells, and splicing outcomes were analysed by RT-PCR using primers annealing to the vector exons. Amplification products were visualised by agarose gel electrophoresis and characterised by Sanger sequencing to define the splicing isoforms generated by each construct. For assay validation, the system was tested using the known pathogenic splice-site variant PTEN:c.209+1G>T, which served as a positive control. The observed splicing pattern matched previously reported data,²⁹ confirming the robustness of the approach (online supplemental figure 3). A summary of clinical data, computational predictions, functional results and final interpretations is provided in table 1. Overall, functional analysis resulted in the reclassification of two variants, confirmed the predicted spliceogenic mechanism for three additional variants and excluded a splice-altering mechanism for three variants. Three representative cases are described below to illustrate the diagnostic impact of rapid functional splicing analysis.

Figure 2. Minigene-based assay for functional assessment of splicing variants. (A) Schematic representation of minigene construct generation. Genomic fragments encompassing the exon of interest and flanking intronic regions are obtained by PCR amplification from genomic DNA or by synthetic DNA production, flanked by attB recombination sites and cloned into the pSpliceExpress splicing reporter vector via Gateway BP recombination. The insert is positioned between two constitutively spliced rat insulin exons (Ins2 and Ins3). The red asterisk indicates the position of the variant under investigation. (B) Experimental workflow of the minigene assay. Following bacterial transformation, positive clones are selected, amplified and verified by restriction digestion and Sanger sequencing. Wild-type (WT) and mutant (MUT) constructs are transiently transfected into HEK-293T cells. Total RNA is extracted, and splicing outcomes are analysed by RT-PCR using vector-specific primers, followed by agarose gel electrophoresis and Sanger sequencing to define exon inclusion or skipping events.

Table 1. Summary of splicing variants analysed by the minigene assay and comparison with in silico predictions.

		Case 1	Case 2	Case 3	Case 4	Case 5	Case 6	Case 7	Case 8
	Diagnostic suspect	Adenomatous or hamartomatous polyposis	Endocrine tumours	Endocrine tumours	Gastrointestinal cancer predisposition	Gastrointestinal cancer predisposition	Gastrointestinal cancer predisposition	Gastrointestinal cancer predisposition	Hereditary diffuse gastric cancer (CDH1 related)
	Gene	PTEN	SDHA	SDHB	POLD1	MLH1	PALB2	PALB2	PALB2
	Annotation	NM_000314.6	NM_004168.3	NM_003000.2	NM_002691.3	NM_000249.4	NM_024675.3	NM_024675.3	NM_024675.3
	Variant	c.209+1G>T	c.1795-1G>T	c.286+2T>C	c.1241A>G	c.117-6T>A	c.3350G>A	c.109C>T	c.3350+2C>T
	Exon	3	12	2	10	2	12	3	12
gnomAD	Alt freq	–	0.0008512%	–	–	–	–	0.0003977%	–
	Alt/ref freq	–	0.0017024%	–	–	–	–	0.0007954%	–
	ClinVar	LP(1)/P(2)	LP(3)/P(1)	LP(1)	VUS(1)	VUS(1)/LP(1)	VUS(6)/LP(1)/P(2)	VUS(8)/LB(1)/B(1)	–
	dbSNP	LP/P	LP/P	–	VUS	–	VUS	VUS	–
	ACMG/AMP classification	LP (PVS1_Strong; PM2_supporting; PS3_supporting)	LP (PVS1; PM2_supporting)	LP (PVS1; PM2_supporting)	VUS (PM2_supporting; PP3)	VUS (PM2_supporting; PP3)	VUS (PM2_supporting; PM5_supporting)	VUS (PM2_supporting)	VUS (PM2_supporting)
	CADD	–	35	–	29.3	–	32	32	–
	PhyloP	9.16	8.33	8.19	4.86	2.55	8.87	5.46	–
	PolyPhen2	–	0.04	–	0.81	–	0.413	1	–
	REVEL	–	0.51	–	0.13	–	0.254	0.28	–
	AlphaMissense	N/A	N/A	N/A	0.081	N/A	0.287	0.662	N/A
	Grantham distance	N/A	N/A	N/A	26	N/A	26	180	N/A
SpliceAI	Acceptor loss	0.9	1	0.79	0.03	0.9	0.43	0.02	0
	Donor loss	1	0.06	0.94	0.01	0.03	0.14	0	0
	Acceptor gain	0	0.86	0	0.09	0.99	0	0	0.53
	Donor gain	0	0	0.26	0	0	0	0	0.85
Pangolin	Splice loss	0.87	0.89	0.77	0.05	0.65	0.5	0	0.01
	Splice gain	0.27	0.74	0.13	0.08	0.84	0	0	0.5
	MaxEntScan	−100%	−100%	−100%	−22.10%	−89.20%	–	−2.50%	+100%
	NNSplice	−100%	−100%	−100%	−1.30%	−71.40%	–	−0.30%	+100%
	SSF	−100%	−100%	−10.6%	−9.90%	−100%	−12.80%	−0.60%	0%
AlphaGenome	Splice site	0.99	1.00	0.08	0.30	0.99	0.077	0.03	0.008
	Splice-site usage	0.97	1.00	0.03	0.06	0.94	0.019	0.03	0.008
	Splice junction	9.55	9.02	0.93	0.98	6.16	1.034	0.68	0.190
	Composite score	3.87	3.80	0.29	0.56	3.18	0.300	0.20	0.054
	Minigene	Exon skipping	Broken acceptor site; activation of cryptic splice site	Broken donor site	No splicing alteration detected	Activation of cryptic splice site; exon skipping	Exon skipping	No splicing alteration detected	No splicing alteration detected
	Evidence added by this study	None; used as positive control	PS3_supporting	PS3_supporting	BS3_supporting	PVS1; PS3_supporting. PP3 not retained after functional confirmation	PVS1(RNA). PP3 not retained after functional confirmation	BS3_supporting	BS3_supporting
	Final interpretation	LP, unchanged	LP, unchanged	LP, unchanged	VUS, unchanged	LP	P	VUS, unchanged	VUS, unchanged

The table reports the set of variants functionally evaluated using the pSpliceExpress minigene assay. For each variant, results from multiple splicing prediction tools (MaxEntScan, NNSplice, SpliceSiteFinder, SpliceAI and Pangolin) are shown. The ACMG/AMP classification reported reflects both the interpretation before minigene testing and the final interpretation, incorporating the evidence added by this study. ACMG/AMP evidence codes applied include PVS1 (pathogenic very strong null/loss-of-function variant), PM2 (pathogenic moderate variant, absent or rare in population databases), PP3 (pathogenic supporting computational evidence of a deleterious effect) and PM5 (pathogenic moderate novel missense change at a known pathogenic residue). ClinGen recommendation where indicated, when available. Colour shading indicates increasing predicted splice impact from green (low impact) to red (high impact).

ACMG, American College of Medical Genetics and Genomics; Alt freq, alternate allele frequency; Alt/ref freq, alternate/reference genotype frequency; AMP, Association for Molecular Pathology; B, benign; LB, likely benign; LP, likely pathogenic; P, pathogenic; VUS, variants of uncertain significance.

Functional characterisation of the MLH1:c.117-6T>A variant

Clinical description

One individual in the cohort was diagnosed at 58 years of age with synchronous adenocarcinomas of the caecum and rectum. Tumour analysis showed high microsatellite instability (MSI), indicating a deficiency of the DNA MMR system. Immunohistochemistry revealed loss of MLH1 and PMS2 protein expression. The absence of a BRAF V600E mutation argued against sporadic MLH1 promoter hypermethylation,³⁰ raising suspicion of Lynch syndrome. No genetic information from affected relatives was available. Germline NGS identified the rare heterozygous intronic variant MLH1:c.117-6T>A (NM_000249.4), located six nucleotides upstream of exon 2.

Variant interpretation and in silico analysis

The MLH1:c.117-6T>A variant is absent from population databases, including gnomAD, indicating an extremely low allele frequency, and is reported in dbSNP and ClinVar, with conflicting clinical interpretations. According to ACMG/AMP criteria, the variant was initially classified as VUS, fulfilling PM2_supporting (absence from population databases) and PP3 (computational evidence supporting a splicing effect). Multiple splicing prediction tools, including MaxEntScan, NNSplice and SpliceSiteFinder, predicted marked weakening of the canonical acceptor splice site of exon 2, while SpliceAI suggested a high probability of aberrant splicing (table 1). However, in the absence of segregation or functional data, these computational findings were insufficient for variant reclassification. To further evaluate the potential splicing impact, AlphaGenome analysis was performed.²² This AI-based framework integrates multiple parameters, including splice-site strength and splice-junction modelling. For MLH1:c.117-6T>A, AlphaGenome predicted a substantial reduction in canonical splice acceptor strength together with the emergence of a novel upstream intronic splice site. Mutant-specific signal peaks across splicing-related tracks were consistent with activation of an alternative splice site and loss of normal splice recognition. These alterations were summarised by a composite AlphaGenome score of 3.2, indicating a high probability of aberrant splicing (figure 3A).

Figure 3. Functional characterisation of the MLH1:c.117-6T>A variant affecting exon 2 splicing. (A) AlphaGenome prediction of the splicing impact of the MLH1:c.117-6T>A variant. Splicing-related tracks comparing wild-type (WT) and mutant (MUT) sequences show reduced signal at the canonical splice acceptor site of exon 2 and the emergence of MUT-specific peaks consistent with activation of an alternative intronic splice site upstream of exon 2. AlphaGenome predictions are based on total RNA-seq data from colon tissue (UBERON:0001157) and include RNA expression and splice-site usage tracks. (B) RT-PCR analysis of transcripts generated by the pSpliceExpress minigene assay, showing distinct splicing patterns for WT and MLH1:c.117-6T>A constructs. (C) Sanger sequencing of RT-PCR products confirming exon inclusion in the WT construct and aberrant splicing in the MUT construct, including an alternative intronic splice site and complete exon 2 skipping. (D) Schematic representation of the splicing mechanisms identified by the minigene assay. The WT construct shows normal exon 2 inclusion, whereas the MLH1:c.117-6T>A variant leads to activation of an alternative intronic splice site or complete skipping of exon 2, both predicted to disrupt the reading frame. The red asterisk indicates the position of the variant under investigation.

Functional splicing analysis

Minigene splicing assays demonstrated that MLH1:c.117-6T>A disrupts normal splicing. RT-PCR revealed two aberrant transcripts in the mutant construct compared with the wild-type control (figure 3B). Sanger sequencing showed that the variant activates a cryptic splice site upstream of the canonical acceptor site, leading to inclusion of four intronic nucleotides at the 5′ end of exon 2 [r.(116_117insCCAG), p.(Leu40Glnfs*14)]. In addition, complete skipping of exon 2 was observed [r.(117_207del), p.(Cys39*)] (figure 3C,D). Because exon 2 is normally in frame, both events introduce a frameshift and generate premature termination codons. These transcripts are predicted to undergo nonsense-mediated mRNA decay or produce a truncated non-functional MLH1 protein.

Diagnostic interpretation

The minigene assay provides direct experimental evidence that MLH1:c.117-6T>A has a deleterious effect on pre-mRNA splicing. These findings support reclassification of the variant from VUS to likely pathogenic, fulfilling ACMG/AMP criteria PVS1, PS3_supporting and PM2_supporting. Importantly, the observed splicing defects are consistent with the patient’s tumour phenotype characterised by MSI and loss of MLH1/PMS2 expression. This case highlights the diagnostic value of functional splicing assays for resolving clinically relevant variants in hereditary cancer syndromes.

Functional assessment of the PALB2:c.3350+2C>T variant

Clinical description

One individual included in the study was diagnosed with desmoid sarcoma of the shoulder at 36 years of age and infiltrating lobular breast carcinoma at 60. Family history included cervical cancer in a sister, lung cancer in both parents and gastric cancer in the mother. Initial testing for BRCA1 and BRCA2 mutations was negative. Germline sequencing using the 77-gene panel identified the heterozygous variant PALB2:c.3350+2C>T (NM_024675.4) at the canonical donor splice site of exon 12. Given PALB2’s role in hereditary breast, ovarian and pancreatic cancer susceptibility,³¹ the variant was selected for further investigation.

Variant interpretation and in silico analysis

PALB2:c.3350+2C>T is a rare intronic substitution absent from population databases. The affected nucleotide shows moderate evolutionary conservation (PhyloP 1.87; GERP 4.17). Splicing prediction tools produced discordant results. HSF and MaxEntScan predicted weakening of the donor splice site, whereas BDGP, ESEfinder, SpliceSiteFinder and SpliceAI did not indicate a significant effect (table 1). The variant has not been previously reported and is classified as VUS in ClinVar. Based on available evidence, it was initially classified as ACMG/AMP class 3, fulfilling criterion PM2_supporting due to its absence from population databases. AlphaGenome analysis predicted only a localised signal alteration without activation of alternative splice sites or major junction rearrangements. The composite AlphaGenome score was 0.05, suggesting preserved splice-site recognition and a neutral effect on splicing (online supplemental figure 4A).

Functional splicing analysis

Minigene analysis showed that PALB2:c.3350+2C>T does not alter pre-mRNA splicing. RT-PCR revealed identical banding patterns for wild-type and mutant constructs, consistent with correct exon 12 inclusion (online supplemental figure 4B). Sanger sequencing confirmed precise exon inclusion without exon skipping, intron retention or alternative splice-site usage (online supplemental figure 4C). The schematic representation of splicing outcomes confirmed preservation of the canonical pattern (online supplemental figure 4D).

Diagnostic interpretation

The functional assay demonstrated that PALB2:c.3350+2C>T does not alter RNA splicing despite affecting the highly conserved +2 position of the canonical donor splice site, providing BS3_supporting evidence against a splice-altering mechanism. These findings highlight the importance of experimental validation in preventing overinterpretation of splice-site variants based solely on sequence context or computational predictions.

Functional characterisation of the PALB2:c.3350G>A variant

Clinical description

One individual in the cohort was diagnosed with bilateral prostatic adenocarcinoma at 59 years of age. Family history was notable for melanoma, perithyroidal tumour, colorectal cancer and prostate cancer in his father, gastric cancer in a paternal uncle, leukaemia in another uncle and metastatic prostate cancer in a cousin. Given this clustering of tumours, germline sequencing using the 77-gene panel was performed. Analysis identified the heterozygous variant PALB2:c.3350G>A.

In silico predictions and variant interpretation

PALB2:c.3350G>A is a rare variant absent from population databases including gnomAD, 1000 Genomes and ESP5400. The variant causes a missense substitution p.(Arg1117Lys) affecting a conserved residue. Protein-level prediction tools yielded conflicting results, with some suggesting deleterious effects and others indicating a benign impact (table 1). Importantly, the substitution affects the last nucleotide of exon 12, a highly conserved position contributing to recognition of the canonical donor splice site. Splicing prediction algorithms were discordant: SpliceSiteFinder and HSF predicted disruption of the donor site, whereas MaxEntScan, NNSplice and SpliceAI predicted minimal impact. AlphaGenome analysis revealed a modest but consistent signal of splice-site perturbation with a composite score of 0.3, suggesting only a modest perturbation of splice-site recognition (figure 4A). According to ACMG/AMP, the variant was initially classified as class 3, fulfilling PM2_supporting and PM5_supporting.

Figure 4. Functional characterisation of the PALB2:c.3350G>A variant affecting exon 12 splicing. (A) AlphaGenome prediction of the splicing impact of the PALB2:c.3350G>A variant. Splicing-related tracks show concordant mutant (MUT)-specific signal changes at the exon 12 donor splice site, consistent with disruption of canonical splice-site recognition and a predicted splice-altering effect. AlphaGenome predictions are based on total RNA-seq data from prostate tissue (UBERON:0002367) and include RNA expression and splice-site usage tracks. (B) RT-PCR analysis of transcripts generated by the pSpliceExpress minigene assay, showing normal exon 12 inclusion in the wild-type (WT) construct and complete exon 12 skipping in the PALB2:c.3350G>A MUT construct. (C) Sanger sequencing of RT-PCR products confirming correct exon 12 inclusion in the WT construct and exon 12 skipping in the MUT construct. (D) Schematic representation of the splicing outcomes identified by the minigene assay. The WT construct shows canonical exon 12 inclusion, whereas the PALB2:c.3350G>A variant results in complete exon 12 skipping. The red asterisk indicates the position of the variant under investigation.

Functional splicing analysis

The minigene assay demonstrated complete exon 12 skipping (figure 4B). Sanger sequencing confirmed exon 12 skipping, resulting in a frameshift and introduction of a premature termination codon [r.(3202_3350del), p.(Gly1068ValfsTer5)] (figure 4C). The schematic representation of splicing outcomes is shown in figure 4D.

Diagnostic interpretation

Exon 12 encodes a region required for interaction with BRCA2 and for the integrity of the PALB2–BRCA1–BRCA2 complex involved in homologous recombination DNA repair. Disruption of exon 12 splicing is therefore expected to severely impair protein function. The minigene assay provides direct functional evidence that PALB2:c.3350G>A disrupts RNA splicing, supporting the application of the PVS1(RNA) criterion, according to the PALB2-specific ClinGen VCEP recommendations (version 1.2.0). Notably, despite its pathogenic classification, both protein-level and splicing predictions for this variant were highly discordant, highlighting the limited reliability of computational predictions alone and the importance of functional assays for accurate variant interpretation.

Discussion

VUS remain a major barrier to the full clinical utility of hereditary cancer multigene testing. In our cohort of 2142 patients, ~17% of individuals carried VUS, consistent with large studies showing that VUS frequency depends on panel size, gene content and population background.^{6 12 13} Variants predicted to affect RNA splicing are particularly challenging, as their functional consequences are difficult to infer from sequence alone.^{16 17} However, not all putative splice-altering variants present the same level of interpretative difficulty. Variants affecting highly conserved splice-site positions are generally expected to disrupt normal splicing and are often supported by strong computational evidence. In contrast, variants located in adjacent intronic regions or exonic positions contributing to splice-site recognition may have less predictable effects on RNA processing. In this context, the principal value of the minigene assay lies in confirming predicted splice defects and in directly defining the molecular mechanism underlying aberrant splicing, thereby providing functional evidence for clinically relevant variants that remain unresolved after routine computational and clinical evaluation. In silico splicing predictors are essential for variant triage but show variable performance, especially for non-canonical splice sites or complex exon–intron contexts.^19–21 Deep learning models have improved mis-splicing prediction, yet benchmarking studies reveal persistent uncertainty, highlighting the need for experimental validation. Accordingly, ClinGen recommends that computational evidence alone is insufficient for definitive classification of suspected splice-altering variants.¹⁷ RNA-based analyses provide direct transcript-level evidence and can resolve splicing-related uncertainty.^{16 23} However, routine RNA testing is limited by tissue accessibility, RNA quality and variable expression, emphasising the need for complementary functional assays that do not rely on patient RNA.

To fill this gap, we developed an optimised recombination-based minigene workflow for rapid functional assessment of splicing variants. Minigene assays offer a controlled system to dissect spliceogenic effects and are well established in variant interpretation.^{26 32} The combination of synthetic DNA fragments with Gateway recombination substantially streamlines construct generation by eliminating primer design, PCR optimisation, colony screening and repeated sequence verification while maintaining high cloning efficiency.^{28 33} As summarised in online supplemental figure 2, this strategy reduces the estimated turnaround time to 9–10 days, making the assay compatible with routine clinical reporting timelines. Importantly, splicing patterns of wild-type and mutant constructs can be directly compared in the same cellular context while requiring only standard molecular biology equipment. When synthetic DNA is used, the assay can be performed without relying on any patient-derived biological material. As an internal validation of the assay, the workflow was first tested using the known pathogenic splice-site variant PTEN:c.209+1G>T, for which the observed splicing pattern was fully concordant with previously reported experimental data.²⁹ Although patient-derived RNA remains the preferred material for evaluating spliceogenic variants, suitable samples are frequently unavailable in routine diagnostic practice. Previous studies have demonstrated a high concordance between minigene assays and patient-derived RNA analyses, with complete concordance reported for BRCA1/2 and MMR gene variants and an overall good agreement across complementary experimental approaches.^{25 34 35} These observations support the use of minigene assays as a reliable complementary approach when patient-derived RNA is unavailable or inconclusive.

Application to unresolved variants demonstrated its value across complementary clinical scenarios. The assay proved informative both for confirming predicted spliceogenic mechanisms and for excluding splice defects in variants initially suspected to alter RNA processing, thereby refining their clinical interpretation. MLH1:c.117-6T>A induced cryptic splice-site activation and exon skipping, supporting reclassification towards likely pathogenic, consistent with tumour phenotype. In contrast, PALB2:c.3350+2C>T maintained exon inclusion despite its canonical donor site location, illustrating that splice-site proximity alone does not imply pathogenicity and highlighting the risk of overinterpretation without functional evidence.^{14 17} PALB2:c.3350G>A underscores the importance of functional validation even for variants deemed pathogenic. Despite meeting PALB2-specific criteria based on RNA data, in silico predictions were discordant. Minigene confirmation of exon skipping provided direct mechanistic evidence, reinforcing the role of functional assays when computational results are inconclusive. Beyond these cases, the optimised workflow is broadly applicable to splicing variants in diverse genes and disease contexts for both germline and somatic variants. The recombination-based design and synthetic DNA enable rapid adaptation and scalability. AlphaGenome can be used to prioritise candidates for functional testing, improving resource allocation without replacing functional evidence. In the present series, AlphaGenome predictions were useful for variant prioritisation, but concordance with experimental results was not absolute. In several cases, variants associated with clear spliceogenic effects in the minigene assay showed only modest predicted splicing impact, highlighting the limitations of relying exclusively on AI-based prediction tools for variant interpretation. These observations suggest that current AI-based predictors may not fully capture the impact of all splice-altering variants and reinforce the need for experimental validation before clinical classification. Minigene assays do not fully recapitulate tissue-specific splicing or complete genomic context, including cis-acting variants or regulatory polymorphisms. Nonetheless, isolating the variant effect under standardised conditions yields robust evidence of spliceogenic potential. When combined with clinical data, segregation and in silico predictions, minigene results represent a valuable component of a multilayered variant interpretation strategy.

In conclusion, this study demonstrates that an optimised recombination-based minigene workflow is a reliable, rapid and clinically relevant method for functional characterisation and reclassification of suspected splice-altering variants in hereditary cancer genes. Integrating such assays into diagnostic pipelines can reduce unresolved splicing uncertainty, enhance NGS utility and support informed clinical decision-making.

The funders had no role in the design of the study, data collection, data analysis, interpretation of the results, manuscript preparation or the decision to publish.

Data availability statement

All data relevant to the study are included in the article or uploaded as supplementary information.