Protein-truncating variants (PTVs) introduce premature termination codons (PTCs) into mRNA (PTC-variants). Typically, these PTC-variant bearing transcripts are degraded by nonsense-mediated decay (NMD), resulting in loss-of-function (LoF) alleles. However, some PTC-variant bearing transcripts escape NMD. These transcripts may lead to truncated or altered proteins with potential LoF, gain-of-function (GoF), or neutral effects. Accurately predicting whether a PTC variant will undergo NMD remains challenging. Existing models, such as the canonical rule based on the exon junction complex (EJC), explain only about 50% of NMD outcomes for PTC variants, highlighting the need for additional rules, possibly incorporating alternative signals like the EJC-independent model of NMD, to improve prediction accuracy.

We propose identifying novel NMD prediction rules by analyzing large-scale RNA sequencing (RNA-Seq) and whole genome sequencing (WGS) datasets. We extracted ~13K high-quality heterozygous germline PTC-variants from 7,617 samples available from the NHLBI Trans-Omics for Precision Medicine (TOPMed) freeze 1.1 RNA-Seq dataset coupled with WGS and mapped them to the GENCODE v26 canonical transcript set. For NMD-triggering variants, the expected proportion of wild-type allele to total expression is ~1, while for NMD-escape variants, it should be ~0.5. We quantified the NMD efficiency of these variants based on their extracted allele-specific expression (ASE) values and analyzed these values against known and potential novel NMD rules. These rules included canonical rule of NMD, 3'UTR length, PTC distance to start codon/stop codon/nearest exon junction, RNA-binding motifs near PTC-variants, potential EJCs, and within 3’UTR sequences.

Our analysis confirmed the canonical rule, showing higher NMD efficiency for PTC-variants upstream of the last EJC (p-value=4.4e-12). Other significant rules included increased NMD efficiency for PTC-variants with longer 3'UTRs (p-value=1.9e-3) and decreased efficiency for PTC-variants near the start codon (first 200 bp) (p-value=1.6e-4) and within the first 31% of a transcript (p-value=8.7e-5). PTC-variants with three downstream exon-exon junctions (p-value=7e-4) had increased NMD efficiency, whereas those further from the nearest exon-exon junction (≥162 bp) had decreased efficiency (p-value=7.2e-3). Using these rules, we built a random forest classification model to predict NMD outcomes for PTC-variants on a training set of high-quality germline PTC-variants from ~80% of the samples (N=6,094) in the dataset. We tested the model on the remaining samples (N=1,523), achieving ~71% prediction accuracy. With this model, we classified ~42.6% of all of the 165.3K high-quality PTC-variants available from the 160.8K samples in the TOPMed freeze 9 dataset as NMD-escape and the remaining 57.4% as NMD-triggering.

In sum, identifying novel NMD prediction rules and developing a more accurate NMD prediction model can better inform genetic and clinical research, potentially leading to more precise diagnosis and treatment strategies for diseases caused by such PTC-variants.