Add NMD prediction support

[mhoang22]

Add Nonsense Mediated Decay (NMD) prediction support for pVACsplice (might work for pVACseq as well, for indels).

Contenders: (tl,dr: NMDj is top tool candidate based on benchmark result and accessibility)

• aenmd:

  • R tool. Paper in 2023. Predict NMD-escape from 5 well-known rules

• NMDClassifier (2nd best candidate?):

  • perl tool. Paper in 2017 .
  • Explore only 1 rule: 50-NT rule. (Reason: benchmark shows that 3/5 other rules arent equally reliable, the 55-NT rule is equally good, yet largest MCC value occurred at 51 NT, very close to 50NT -> prioritize 50-NT rule - see section 'Detection of NMDTs').
  • How to install: download the .tar on biotools: https://bio.tools/nmd_classifier . Have no github repo.
  • inputs: transcripts.gtf , annotation gtf file (ensembl/ncbi reference gtf ), genomic sequence (fa?) , annotation mode (ensembl or ncbi) , optional filtering parameter.

• NMDj (top candidate?): python tool

  • python tool. Paper in 2025.
  • Principle: also use 50-NT rule and compare (novel?/input) transcript to reference transcript, just as NMD Classifier. However, NMD Classifier uses best partner transcript as reference, whereas NMDj uses (1) MANE-select transcript as default or (2) transcript with best expression level , in user-input mode. Transcript with best expression level is determined by psi value, calculated using split read counts reported by pyIPSA.
  • Rationale behind option number 2:

the probability of NMDT being derived from a protein- coding transcript via AS depends not only on the similarity in their exon-intron architectures but also on their expression levels. The coding transcript with the highest expression level is more likely to be the source of NMDT [14]. Furthermore, NMDT may be derived from different transcripts with comparable expression levels, which calls into question the validity of the approach based on the selection of only one matching transcript partner.

• repo: https://github.com/zavilev/NMDj/

• installation: git clone https://github.com/zavilev/NMDj.git

• inputs: input trancripts.gtf , OPTIONAL inputs: annotation.gtf (reference gtf?), genome.fa (reference fa), transcripts.txt (user-input reference transcript ids), file.txt (File with paths to ipsa files containing counts of RNA-Seq split-reads aligned to junctions), ...

• benchmark with NMDclassifier. Also benchmark MANE-select vs best expression transcript. Best expression transcript wins.

• NMDEP: an AI/ML model. Manuscript on arxiv in 2025.

  • repo/installation note: none, as of Mar 2026.
  • no benchmark with previous tools. but the model also have PCT as top predictor, so pretty much agree with literature.

• factR/predictNMD: R function. https://rdrr.io/github/fursham-h/factR/man/predictNMD.html