Quick Start Guide: predict

This guide provides a brief walkthrough for using the predict subcommand in OpenSpliceAI to predict splice sites from DNA sequences. If you haven’t done so already, please see the Installation page for details on installing and configuring OpenSpliceAI.

Before You Begin

• Install OpenSpliceAI: Ensure you have installed OpenSpliceAI and its dependencies as described in the Installation page.

• Check Example Scripts: We provide an example script examples/predict/predict_cmd.sh

One-liner Start

OpenSpliceAI can predict splice sites from DNA sequences with a default 10,000 nt flanking region. You need:

1. A reference genome (FASTA) : chr22.fa

2. A reference annotation (GTF) : chr22.gff

3. A pre-trained OpenSpliceAI model or directory of models:

   • GitHub (models/openspliceai-mane/10000nt/) or

   • FTP site (OSAI-MANE/10000nt/)

Run the following commands (adapt or replace filenames as needed):

openspliceai predict \
    -m 10000nt/ \
    -o results \
    -f 10000 \
    -i data/chr22.fa \
    -a data/chr22.gff \
    -t 0.9

This command will generate prediction results in the specified output directory (results/). The predictions will be based on the input FASTA file (chr22.fa) and the annotation file (chr22.gff). The results will include a GFF file with predicted splice sites and their scores.

Important

Strand and flanking context matter. OpenSpliceAI scores the sequence on the strand you provide, and the deep models need flanking_size/2 bp of real sequence on each side of every position they score (5,000 bp per side for the 10,000 nt model). The simplest way to satisfy both is to pass the whole chromosome/genome FASTA together with a GFF via -a/--annotation (as above): OpenSpliceAI then reverse-complements minus-strand genes for you and scores each gene in its genomic context. If you instead feed a short, standalone gene FASTA, the sequence is padded with N and minus-strand genes are scored on the wrong strand — both of which collapse the scores to near zero.

Predicting a single gene of interest

If you only care about one gene, do not extract just the gene body into a FASTA and run it directly — that loses the flanking context and (for minus-strand genes) the correct strand. Instead, either:

• Recommended: keep the full chromosome FASTA and provide a GFF for the gene(s) you want, with -a. OpenSpliceAI extracts each gene with flanking_size/2 bp of real flanking on each side (configurable via --gene-flank) and reverse-complements minus-strand genes automatically.

• Or extract the gene ± 5,000 bp of surrounding genome yourself; if the gene is on the minus strand, reverse-complement that window before running predict without -a.

A correctly run, well-annotated gene should produce donor/acceptor scores close to 1.0 at its true splice junctions. If your scores look uniformly low, see the troubleshooting Q&A (“My predicted donor/acceptor scores are unexpectedly low”).

Next Steps

• Explore ``predict`` Options: Dive into the predict documentation to learn more about the available options for predicting splice sites.

• Begin Variant Prediction: Check out the Quick Start Guide: variant guide to predict the impact of genomic variants on splice sites.

We hope this quick start guide helps you get up and running with OpenSpliceAI. Happy predicting!