Programs that find splice sites, the exact positions where a gene's intervening sequence is cut out, perform well in the species they were trained on. They lose much of that accuracy when moved to a distant species. The obvious remedy is to carry known human sites across to the new genome through a sequence alignment and train on those copied labels. That remedy breaks exactly where it is most needed. After enough evolutionary time, small insertions and deletions slide each site a few bases away from where the alignment puts it, and in the most divergent regions the alignment fails altogether. Copied labels then pin the "true" site on the wrong nucleotide, which teaches the model to trust nearby decoys: the far more numerous GT and AG letter pairs that look like splice sites but are not.
We propose to use conservation as a tolerant training signal instead of as an exact label. Each annotated human site is embedded in its full surrounding sequence. The model is then told only that the matching site in the other species lies somewhere within a few bases of where the alignment points, with nearer offsets weighted more heavily. Every other GT/AG candidate in the same orthologous locus (the corresponding stretch of the other genome) is treated as a hard negative. The loss combines three existing ideas: offset weights from shift-tolerant string kernels, the "neighboring sentences in the same passage are the hardest negatives" trick from contextual sentence retrieval, and the treatment of related sequences as imperfect translations of each other. The question is whether this lets a splice model trained with human labels alone reach high precision in a mouse or zebrafish genome that has no annotation. A secondary question is whether the gain comes specifically from sites that alignment displaced.
Core Claim
A contrastive loss that accepts any candidate site within a weighted offset window as the positive, while treating the other candidates in the same locus as negatives, will raise area under the precision-recall curve (auPRC) and recall at a fixed calibrated false-discovery rate in an unannotated target species. The baseline for this comparison is training on alignment-projected labels, which is the same method with the window shrunk to zero offset. The gain should grow with evolutionary distance from human and should appear almost entirely at sites whose true position differs from the alignment-projected one.
The motivation comes from comparative genomics. Footer showed that binary conservation filters fail when functional sites sit in fast-changing sequence. Kheradpour et al. showed across 12 Drosophila genomes that regulatory motifs frequently shift coordinates, so strict alignment produces many false negatives. Chan et al. found vertebrate tissue expression conserved over 450 million years despite regulatory DNA rearranged beyond what linear aligners can recover. Together these say that conservation is real but positionally loose. Existing pipelines use it either as a hard filter, which costs sensitivity, or as an exact label, which costs precision near decoys. We use it as neither.
What Is Borrowed and What Is New
The positional looseness has a well-understood classical treatment. The weighted degree kernel with shifts in Accurate splice site prediction using support vector machines scores substring matches across offsets $\delta$ with a decay weight $c(\delta)$. That kernel operates within one species on labeled windows of a few hundred bases. Large scale genomic sequence SVM classifiers scaled it to a million C. elegans sequences but notes that wide shift windows and mismatches make its index structures blow up in memory. We keep only the $c(\delta)$ idea. We move it from the similarity function into the definition of which training pair counts as positive, where it costs nothing in memory.
The contrastive structure comes from text retrieval. Sentence-aware Contrastive Learning for Open-Domain Passage Retrieval (DCSR) reads out one vector per sentence from an unchunked passage and uses the other sentences of the same passage as hard negatives. This raised SQuAD top-20 retrieval from 52.8% to 63.7%. CroCoAlign contextualizes each sentence with its neighbors to align books across languages when translations merge, split, and reorder sentences. It beat the dynamic-programming aligner Vecalign by 1.6 F1 on average. Our analogy is that an orthologous locus is a noisy translation of the human locus and candidate dinucleotides are its sentences. Neither paper deals with sequences lacking natural boundaries, labels that exist on only one side of the translation, or thousand-to-one decoy imbalance. Those three features are what our loss is built for.
The backbone, calibration procedure, and benchmark come from OpenSpliceAI. That work retrains a dilated convolutional splice model with a context of up to 10,000 bases and calibrates it with class-wise temperature scaling. It also filters paralogous genes to stop train/test leakage, and it shows large gains when models are retrained on each species' own annotations. OpenSpliceAI assumes target annotations exist. We ask what can be done when they do not.
We should be candid that the most tempting "contextual sentence encoding for DNA" proposal adds little here. That proposal would insert boundary tokens at candidate sites and contrast true sites with in-window decoys within one species. A per-nucleotide classifier with 10,000 bases of context already treats every decoy in its window as a negative. The new ingredient is therefore not in-sequence contrast by itself. It is contrast across species, with a positive set that tolerates shifts.
Research Program
The model is an OpenSpliceAI backbone with a small projection head. The head reads out an embedding at every GT (donor) and AG (acceptor) candidate position, playing the role of DCSR's sentence tokens. The ordinary per-nucleotide cross-entropy loss is kept on human annotations.
For each human annotated site $i$, we locate the orthologous target locus using whole-genome alignment chains only. Ortholog calls derived from target gene annotations are not used, because that would leak labels. Let $p_i$ be the alignment-projected target position, $W(i)$ the target candidates within $\pm\Delta$ of $p_i$, and $C(i)$ all same-type candidates in the orthologous locus plus in-batch candidates from other loci. With $h_i$ the human anchor embedding, $g_j$ a target candidate embedding, $s$ cosine similarity, and $\tau$ a temperature, the added loss is
$$\mathcal{L}_i = -\log \frac{\sum_{j \in W(i)} c(\delta_j)\, e^{s(h_i, g_j)/\tau}}{\sum_{j \in C(i)} e^{s(h_i, g_j)/\tau}}.$$
Here $\delta_j = |j - p_i|$ and $c(\delta)$ is a decaying weight in the spirit of the shift kernel. This is a multiple-instance loss: it rewards putting weight on some candidate in the window without saying which one. Setting $\Delta = 0$ turns it into ordinary contrastive training on projected labels, which is the key ablation.
Target predictions come from the shared classifier head applied to target sequence, which the contrastive term has shaped. A lightweight variant adds one round of self-training. In that round, the highest-similarity candidate in each window becomes a pseudo-label, provided its calibrated score clears a threshold. All scores are calibrated with OpenSpliceAI's class-wise temperature scaling on a human validation set before any fixed-false-discovery-rate comparison. Otherwise precision differences could simply reflect miscalibration.
The analysis divides target test sites into three strata. "Aligned-exact" sites are those where the annotated site equals $p_i$. "Aligned-shifted" sites are those within the alignment but offset from $p_i$. "Unaligned" sites are those in loci the chains do not cover. The mechanism predicts gains in the second stratum, no loss in the first, and at most modest spillover to the third through the shared encoder.
Baselines
The do-nothing reference is the human-trained OpenSpliceAI model applied directly to the target genome, after the same temperature calibration. It measures how much accuracy is lost to divergence. OpenSpliceAI's reported numbers show the size of this gap. Retraining on species-specific data improved accuracy over the human-centric SpliceAI by 57% in Arabidopsis, 54% in honeybee, and 19% in zebrafish (OpenSpliceAI). We expect a smaller but clear gap for mouse.
The strongest existing competitor we can build without target annotations is projected-label training. Human annotations are lifted through the same alignment chains, and OpenSpliceAI is trained on human labels plus the lifted target labels. It is the strongest competitor because it uses exactly the same information as our method: human labels, target genome, and alignment. It is also the standard practice in comparative annotation. It is run with the same backbone, data, and number of training steps, and it gets the same hyperparameter search budget as our method. Our method must beat it on target auPRC and on recall at a calibrated 10% false-discovery rate, with a confidence interval computed by bootstrapping over genes. The advantage must show up in the aligned-shifted stratum.
The key ablation is our loss with $\Delta = 0$. It differs from the projected-label baseline only in using a contrastive rather than a cross-entropy form. Comparing it with the full loss isolates the effect of shift tolerance. A second ablation removes in-locus negatives from the denominator and keeps only in-batch negatives from other loci. This tests whether DCSR's same-passage negatives matter when decoys outnumber true sites about a thousand to one.
A hard-conservation-filter baseline keeps human-model target predictions only where they fall at an aligned human site. It represents the precision-first extreme that Kheradpour et al. warn loses sensitivity. We expect it to have high precision and poor recall, and we will report its full curve.
A classical reference is the shifted weighted degree kernel from Accurate splice site prediction using support vector machines, trained on human plus projected target labels. On the human benchmark that paper reports auPRC of roughly 54–57%, against roughly 16–25% for Markov chains. We found no reported number on Lacuna for it on mouse or zebrafish.
The upper bound is OpenSpliceAI retrained on the target species' own annotations. We report the fraction of the gap between the do-nothing model and this upper bound that each method closes.
Why not just…?
Why not just train on the target species' own annotations? When good annotations exist, that is the right answer, and it serves as our upper bound. The proposal is aimed at the many sequenced genomes whose annotations are missing or were themselves projected from human. If our method closes only a small fraction of the gap to the upper bound, the practical case weakens, and we will report that fraction directly.
Why not just lift human annotations through the alignment and train on those? This is the real threat. For mouse, where most sites are aligned exactly, we expect it to perform about as well as our method. Our claim holds only if the advantage grows from mouse to zebrafish and is concentrated among shifted sites. The $\Delta = 0$ ablation and the stratified analysis settle the question. If the shifted stratum shows no gain in zebrafish, the idea is wrong, and projected labels are sufficient.
Why not just filter human-model predictions by conservation? A filter can only remove calls, so it trades recall for precision. The motif-drift results in Kheradpour et al. and Footer indicate that it will discard displaced true sites. The same pattern appears in variant calling, where requiring agreement among tools raises specificity but inflates false negatives. Our comparison at a matched calibrated false-discovery rate will show whether the filter's precision comes at a recall cost our method avoids.
Why not just use shifted string kernels, which already tolerate offsets? They tolerate offsets only between labeled sequences of the same species, within windows of a few hundred bases. Large scale genomic sequence SVM classifiers documents their memory blowup with wider shifts and mismatches. They also cannot use unlabeled target sequence. We include them as a reference rather than dismissing them.
Proposed Experiments
We use OpenSpliceAI's multi-species setting with human MANE v1.3 as the only labeled source. Mouse and zebrafish serve as unannotated targets. Target annotations are used only for evaluation, and paralogs are filtered with minimap2 as in OpenSpliceAI. The metrics are target auPRC for donors and acceptors, OpenSpliceAI's top-1 accuracy, and recall at a calibrated 10% false-discovery rate, all reported per stratum.
The first decisive check is the interaction between species and stratum. The method must beat projected-label training in the zebrafish aligned-shifted stratum without losing in the aligned-exact stratum. The second check sweeps $\Delta$ over 0, 2, 5, 10, and 25 bases. Performance should rise and then plateau near the empirical distribution of true offsets, which we measure from held-out annotations. A monotone decline would mean the window only adds label noise. The third check removes in-locus negatives. If precision at fixed false-discovery rate is unchanged, the DCSR-style negatives contribute nothing, and we will say so.
As a stress test, we extend the approach to honeybee and Arabidopsis, which cannot be aligned to human at the genome level. There, orthologous loci are defined by protein-level gene homology and the window spans whole candidate regions. We expect this regime to be hard and treat any positive result there as exploratory.
References
- Footer: A quantitative comparative genomics method for efficient recognition ofcis-regulatory elements
- Reliable prediction of regulator targets using 12 Drosophila genomes
- Conservation of core gene expression in vertebrate tissues
- Accurate splice site prediction using support vector machines
- Large scale genomic sequence SVM classifiers
- Sentence-aware Contrastive Learning for Open-Domain Passage Retrieval
- CroCoAlign: A Cross-Lingual, Context-Aware and Fully-Neural Sentence Alignment System for Long Texts
- OpenSpliceAI provides an efficient modular implementation of SpliceAI enabling easy retraining across nonhuman species