ConvSplice

The dinucleotides that mark an intron's edges — GT at the donor end, AG at the acceptor — are far too common in genomic sequence to identify a splice site on their own. Whether a given occurrence is used depends on context that can sit thousands of bases away: the lengths of the flanking exon and intron, competing sites elsewhere in the gene, branch points, and nearby splicing enhancers and silencers. How much sequence a model sees is therefore the design constraint that matters most in splice site prediction.

ConvSplice answers that constraint with receptive field. It takes a 20 kb one-hot encoded DNA window — 10 kb of flanking sequence on each side of the region scored — and passes it through a stack of dilated residual convolutions, emitting a three-way softmax at every position: splice acceptor, splice donor, or neither. Dilation is what makes the arithmetic work, spacing each convolution's taps so that twenty blocks reach across the full window without the parameter cost of twenty wide layers. The architecture builds on the dilated-residual framework introduced by SpliceAI, doubling its sequence context.

ConvSplice was developed at Princeton University and the Flatiron Institute in Olga Troyanskaya's group, released alongside Otari, the graph attention framework it feeds. Otari builds each isoform into a graph whose nodes carry a 23,600-dimensional attribute vector drawn from three frozen sequence models: Sei supplies chromatin features, Seqweaver supplies RNA-binding protein affinities, and ConvSplice supplies the splicing channel.

#Key Features

• 20 kb receptive field: Each prediction sees 10 kb of sequence to either side, enough to take in whole introns and the competing sites that decide which boundaries a transcript uses.
• Per-position scoring: The model labels every base of the input rather than classifying pre-selected candidates, so it runs as a scanner across a whole transcript.
• Five-seed ensemble: Five checkpoints trained from independent initializations are averaged into the final score, damping the seed-to-seed variance a deep convolutional stack shows on a rare positive class.
• Frozen at inference: The released predictor runs all five checkpoints under model.eval() inside torch.no_grad(), fitting nothing at run time — which is what lets a downstream model consume its outputs as fixed features.

#Technical Details

The released implementation lifts the input to 32 channels with a 1×1 convolution, then applies twenty residual blocks of two dilated Conv1d layers each — forty convolutions — with batch normalization and SiLU activations and a 1×1 skip projection summed in after every fourth block. Kernel width and dilation widen together down the stack: eight blocks at kernel 11 with dilation 1 then 4, four at kernel 21 with dilation 10, and eight at kernels 41 and 51 with dilation 25. A 1×1 convolution to 128 channels and a second to three produce the per-position softmax, for roughly 1.12 million parameters per checkpoint and 5.6 million across the ensemble. Training used GENCODE v40 principal transcripts annotated by APPRIS, taking annotated 5′ and 3′ splice sites as positives and every other transcript position as a negative; chromosomes 1, 3, 5, 7, and 9 were held out as the test set and the remainder split 80/20 for training and validation, optimized with Adam at learning rate 1e-4 under a ReduceLROnPlateau schedule. On the held-out chromosomes the ensemble reaches precision-recall AUC 0.98 for both donor and acceptor sites against SpliceAI's 0.97, and top-k accuracy 0.94 for both site types against SpliceAI's 0.93 and 0.92.

#Applications

The direct use is scoring splice site strength anywhere in the human genome, including under mutation: substitute a variant into the 20 kb window, re-score, and the change in donor or acceptor probability quantifies how much that variant weakens or creates a boundary. Inside Otari, that difference carries the variant's effect through the transcript graph into isoform-level abundance predictions across 30 tissues. The wrapper shipped with the framework is built for that use — it expects an Otari transcript object and returns zeros for positions not already annotated as splice sites — so scanning with ConvSplice means calling the model module directly, which takes any one-hot 20 kb window.

#Impact

ConvSplice demonstrates the frozen-backbone pattern now standard in regulatory genomics: trained once on annotation-derived labels, then reused untouched as a feature extractor for a task it was never fit to. Its scores are a small fraction of an Otari node vector, but they carry splicing signal the chromatin and RBP channels cannot supply, and the paper's ablations mask each channel in turn to check that contribution. Scope is bounded: the model is human-only, its labels come from annotation rather than measured splicing, and it scores individual sites rather than isoform-level outcomes. So is availability — architecture and inference wrapper sit in the unlicensed Otari repository, whose training scripts cover the graph network rather than ConvSplice, and the five checkpoints ship inside a 5.1 GB Zenodo archive under CC BY 4.0 rather than as standalone downloads.