OpenSpliceAI

SpliceAI, published in 2019, remains one of the most accurate methods for reading splicing signals directly out of genomic DNA, and its delta scores are now routine in clinical variant interpretation. But the released model was frozen in place: trained once on GRCh37, distributed as TensorFlow/Keras weights, and shipped without a usable training pipeline. Researchers who wanted to update it to a current human annotation, or apply it to a non-human genome, had no supported route to do so.

OpenSpliceAI is a ground-up reimplementation of that system in PyTorch, released by Kuan-Hao Chao, Alan Mao, Anqi Liu, Steven Salzberg, and Mihaela Pertea at Johns Hopkins University. It preserves SpliceAI's dilated residual convolutional architecture and scoring semantics while replacing the frozen artifact with a complete, modular toolchain: dataset construction, training from scratch, transfer learning, probability calibration, prediction, and VCF variant annotation, each exposed as a subcommand.

Critically, OpenSpliceAI is usable without any retraining. The project ships pretrained checkpoints for human — OSAI-MANE, retrained on GRCh38 with the current MANE annotation — alongside models trained from scratch on mouse, zebrafish, honey bee, and Arabidopsis thaliana. Retraining is an available capability, not a prerequisite. The work appeared as a bioRxiv preprint in March 2025 and was published in eLife later that year.

#Key Features

• Faithful but improved human model: On held-out chromosomes, the 10,000 nt OSAI-MANE model exceeds the original SpliceAI-Keras model by roughly 1.3% (donor) and 1.6% (acceptor) in top-1 accuracy, with AUPRC gains of about 1.9% and 1.8%.
• Cross-species pretrained checkpoints: Species-specific models for mouse, zebrafish, honey bee, and Arabidopsis remove the human-centric bias that degrades SpliceAI on distant genomes, with average gains over the human-trained model widening from roughly 2% on mouse to over 50% on honey bee and Arabidopsis.
• Transfer learning: Fine-tuning the human model onto a new species reaches near-optimal accuracy after a single epoch, where training from scratch needs roughly ten.
• Variant impact scoring: The variant subcommand annotates VCFs with delta scores and positions for SNPs and indels, computing the maximum splicing-score change within a configurable window (50 nt by default) on either side of the mutation.
• Calibrated probabilities: A dedicated calibration step applies temperature scaling so that output scores behave more like true probabilities, which matters when thresholds drive triage.
• Efficiency at genome scale: Dynamic graphs and on-demand GPU allocation cut runtime and peak memory relative to the Keras original, making chromosome-scale prediction feasible on a single GPU.

#Technical Details

Each model is a deep residual CNN that takes one-hot encoded DNA and emits a three-class softmax per nucleotide — neither, acceptor, or donor. Four flanking-context configurations are provided: 80, 400, 2,000, and 10,000 nt, built from 4, 8, 12, and 16 residual units respectively, with dilation rates escalating through 1, 4, 10, and 25 and kernel widths through 11, 21, and 41 to reach the full receptive field. The 10,000 nt configuration is the recommended default. Five independently seeded replicates are released per configuration and averaged as an ensemble at inference, matching SpliceAI's practice.

OSAI-MANE was trained on roughly 20,000 protein-coding genes from RefSeq MANE v1.3 on GRCh38, holding out chromosomes 1, 3, 5, 7, and 9 for testing and using the remainder for training, with paralogous test sequences filtered out. Agreement with the original model was assessed directly rather than assumed: in silico mutagenesis across 200 splice sites gave a median Pearson correlation of 0.857 between the two models' attributions, and the learned motif logos matched at 0.996 (donor) and 0.997 (acceptor) similarity.

#Applications

The most immediate use is clinical and research variant interpretation in human genomes, where delta scores flag mutations that create or destroy splice sites — the paper walks through cryptic splicing in MYBPC3 and pseudoexon inclusion in OPA1. Beyond human, the cross-species checkpoints and retraining pipeline open splice-site annotation and variant scoring to model organisms, crops, and newly sequenced genomes, where accurate splice prediction directly improves gene structure annotation. The package installs from PyPI and Bioconda, so it drops into existing genomics pipelines alongside tools like AbSplice2 for tissue-specific splicing effects.

#Impact

OpenSpliceAI's contribution is less a new architecture than the removal of a bottleneck: a widely trusted splicing model that could not be updated or ported is now a maintained, retrainable open-source package under GPL-3.0, with documented checkpoints and a demonstrated path to new species. That matters even as broader sequence-to-function models such as AlphaGenome and splicing-focused language models like SpliceBERT expand the space, since a cheap, well-calibrated single-task convolutional model remains practical for genome-wide scans. The pretrained models do inherit SpliceAI's core limitation: they score splice-site strength from local sequence context and do not model tissue- or condition-specific isoform usage.