Benchmarking deep learning splice prediction tools using functional splice assays
Problem
Predicting whether variants in noncanonical splice sites and deep intronic regions disrupt splicing is difficult, complicating diagnosis of hereditary disorders.
Why it matters
Misjudged splice variants can lead to missed or incorrect genetic diagnoses.
What the AI did
Benchmarked ten splice prediction tools, including deep learning models, against functional midigene and minigene assays on gold-standard variant sets in ABCA4 and MYBPC3.
Results
SpliceRover best for ABCA4 noncanonical splice-site variants; SpliceAI best for a truncated variant class; full rankings not stated in source.
Impact
Independent functional-assay benchmarking identifies which AI splice predictors are most reliable for diagnostics.
Record
| Venue | Human Mutation |
|---|---|
| Published | 2021 |
| Work type | article |
| Authors | Tabea Riepe, Mubeen Khan, Susanne Roosing, Frans P.M. Cremers, Peter A.C. ’t Hoen |
| Citations | 100 (OpenAlex, at import time) |
| References | 64 |
| Model or method | SpliceAI, SpliceRover, MMSplice among benchmarked tools |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1002/humu.24212 |
| Open access | hybrid |
| Keywords | splice prediction, deep learning, SpliceAI, benchmarking, functional assays, genetic variants |
Importance rating
real but narrower advance
Abstract states hereditary disorders are frequently caused by splicing-affecting variants and that prediction of noncanonical/deep intronic variants is difficult; benchmarking aids diagnosis, no scale given.