Heng Li's minisplice: a 7k-parameter CNN halves minimap2 splice junction error rate

anshulkundaje · x · 2026-09-24

Heng Li's team published "Improving spliced alignment by modeling splice sites with deep learning" in Algorithms for Molecular Biology.

minisplice learns splice signals with a 1D-CNN of just 7,026 parameters, trained for vertebrate and insect genomes
The tiny model captures conserved splice signals across phyla and reveals GC-rich introns specific to mammals and birds
It assigns empirical splicing probabilities to every GT/AG site, and minimap2 and miniprot now optionally use these scores to improve spliced alignment
Evaluation on human long-read RNA-seq and cross-species protein datasets shows greatly improved junction accuracy, especially for noisy long reads and distant-homology proteins

Open source: github.com/lh3/minisplice. The author says he now uses the new feature routinely.

More from Research