Pangolin

Summary¶

Pangolin predicts, for each position, both the probability that it is a splice site and its usage — the proportion of a gene's transcripts that use it — separately in heart, liver, brain and testis.[1] Two design choices distinguish it from SpliceAI, whose architecture it otherwise closely mirrors: it trains on quantitative splice-site usage measurements rather than binary annotations, and on RNA-seq from four species rather than human alone, both shown by ablation to improve accuracy independently.[1] It is the model that beats both AlphaGenome and Borzoi on specific splicing benchmarks, which is why it matters disproportionately to its citation count.

Provenance Note¶

Pangolin comes from the same group as LeafCutter and LeafCutter2 (Yang I. Li, University of Chicago) — the prediction and quantification sides of splicing developed together.

Architecture and training¶

16 stacked residual blocks (batch normalisation, ReLU, dilated convolutions) with skip connections feeding the penultimate layer from before blocks 1, 5, 9 and 13. Receptive field ±5,000 bp — the same reach as SpliceAI-10k.[1]
The architectural difference from SpliceAI is confined to the last two layers: Pangolin's final layer has multiple outputs, producing a probability (softmax) and a usage value (sigmoid) per tissue, where SpliceAI emits a single tissue-invariant donor/acceptor/neither call.[1]
Training labels come from SpliSER, which computes a per-tissue Splice-Site Strength Estimate (SSE) from split and non-split reads, applied to RNA-seq from up to 8 samples per species per tissue across human, rhesus macaque, mouse and rat.[1]
Split: test = human chromosomes 1, 3, 5, 7, 9; train = the remaining human chromosomes plus non-ortholog, non-paralog genes from the other three species. Orthologs of human test genes were excluded from the non-human training data to prevent leakage — and the authors verified that improvements held for genes with low identity to cross-species training genes.[1]

Ablation result worth carrying forward: three separable factors each improved AUPRC — (i) training on multiple tissues rather than one, (ii) multiple species rather than human alone, and (iii) quantitative usage measurements alongside binary classifications.[1] This is direct evidence that label quality and diversity, not just architecture, drove the gain over SpliceAI, whose backbone Pangolin essentially reuses.

Performance¶

Splice site identification (test chromosomes, averaged across tissues):[1]

Model	Top-1 accuracy	Top-0.5 accuracy	AUPRC
Pangolin	79%	94%	0.85
SpliceAI	75%	87%	0.77
MMSplice, HAL, MaxEntScan	<37%	—	<0.30

Rare splice-disrupting variants (MFASS). On 27,733 ExAC variants tested by minigene Sort-seq, of which 3.8% strongly disrupt exon usage, Pangolin reached AUPRC 0.56 versus SpliceAI's 0.47.[1] At 80% precision it recalls 29% of splice-disrupting variants. Performance is strongly distance-dependent: AUPRC 0.75 for variants 0–9 bases from a splice site, but below 0.35 beyond 9 bases.[1]

Disputed: Pangolin's own paper reports MFASS AUPRC 0.56 for Pangolin and 0.47 for SpliceAI [1]. The AlphaGenome evaluation of the same benchmark reports 0.54 for Pangolin and 0.49 for SpliceAI (with AlphaGenome at 0.51 and DeltaSplice at 0.49) [3]. The rank order of Pangolin > SpliceAI is preserved, but the absolute values differ, and neither paper documents the variant filtering or score-aggregation choices in enough detail to reconcile them from the texts alone. Treat small MFASS auPRC differences between papers as within reimplementation noise.

In vivo splicing efficiency (MAPSy). Pearson correlation with measured effects: Pangolin 0.61, SpliceAI 0.50, MMSplice 0.37.[1]

Epistasis. On FAS exon 6 with 189 single substitutions and 3,059 combinations of substitutions, predicted-versus-experimental PSI Spearman was 0.79 for singles and 0.80 for combinations — where a linear model built from the individual variants' PSIs achieved only 0.48.[1] The model captures non-additive interactions between variants, which single-variant scoring cannot.

Tissue specificity. Predicting each tissue's deviation from mean usage gave Spearman r 0.35–0.50 (median 0.43), compared with MTSplice's 0.09–0.40 (median 0.22).[1] The authors' own verdict is candid: "predicting differential splicing across tissues from sequence alone is possible but remains a considerable challenge."[1] Motifs enriched near correctly predicted brain- and testis-specific splice sites did not match any well-studied splice factor.

Applications¶

Clinical variant interpretation — the strongest result. On 3,893 BRCA1 single-nucleotide variants with functional status measured by saturation genome editing:[1]

Loss-of-function variants are strongly enriched among those Pangolin predicts to affect splicing (χ² p = 6.3 × 10⁻¹¹⁹ at cutoff 0.2).
Excluding missense and nonsense variants, Pangolin distinguishes LOF from functional variants with AUPRC 0.95 (1,591 variants), and 0.90 for the 861 variants in the extended splice region (±15 bp, excluding canonical splice sites).
For missense variants alone, AUPRC is only 0.31 — as expected, since most missense LOF acts through protein sequence, not splicing. Even so, 5–10% of LOF missense variants appear to act through splicing effects rather than amino acid change.

ClinVar. For variants in extended splice regions, Pangolin separated 842 pathogenic from 11,256 benign SNVs at AUPRC 0.90 versus SpliceAI's 0.87; including annotated splice sites raised this to 0.99.[1] Applied to 21,363 variants of uncertain significance in splice regions, 5,766 were predicted likely to impact splicing and therefore likely pathogenic (cutoff 0.2).[1]

sQTL causal-variant nomination. For the top 500 DGN sQTLs called with LeafCutter, SNPs nominated as causal by Pangolin had smaller sQTL p-values than those nominated by SpliceAI, which in turn beat random SNPs.[1]

Comparative genomics. Pangolin, untrained on chimpanzee, recovered 35% (550/1,560) of human–chimp brain splice sites differing in usage by ≥0.5, at a ~5% false sign rate — implying sequence divergence within 5 kb of the splice site explains a large fraction of inter-species splicing differences. Of the predicted differences, 47% were attributable to a single variant.[1]

In silico mutagenesis recovered known splicing grammar de novo, including polypyrimidine tract strength at 3′ splice sites and the U2AF1 preference for C or T at the −3 position. Notably, far fewer mutations are predicted to increase usage than decrease it, implying most exons already sit near optimal splicing accuracy.[1]

Where Pangolin fits¶

Synthesis: Pangolin is the counter-example to a clean "newer model wins" narrative. It has a 10 kb receptive field — smaller than Borzoi (524 kb) and AlphaGenome (1 Mb) — yet beats Borzoi on fine-mapped sQTLs within 10 kb of a splice site [2] and beats AlphaGenome on MFASS [3]. Reading its ablations against those results suggests why, though no cited paper states this: Pangolin's advantage comes from training targets (quantitative, multi-tissue, multi-species usage measurements) rather than context length [1], and splice-site strength is a fundamentally local property, so a larger receptive field buys little on this specific task. The corollary is that context length and label quality are separate axes of model improvement, and the field's recent progress has been concentrated on the first.

Its complementarity with coverage-based models is distance-structured: Pangolin wins on distant de novo splice gains, where a new site simply appears and local scoring suffices; Borzoi wins within 200 bp of a junction, where the question is competitive usage; and the average of both models' ranks beats either alone.[2]

Citations¶

[1] Zeng, T. & Li, Y. I. (2022). Predicting RNA splicing from DNA sequence using Pangolin. Genome Biology 23:103. Supports: architecture and the two-layer difference from SpliceAI; SpliSER-based quantitative multi-species training and the train/test split; all top-1/top-0.5/AUPRC, MFASS, MAPSy, epistasis and tissue-specificity figures; the three-factor ablation; BRCA1, ClinVar, VUS, DGN sQTL, human–chimp and ISM results. Location: Main text (all sections); Methods ("Deep neural network architecture"; "Generating training and test sets"); Figs. 1–2. Source paper: s13059-022-02664-4.pdf

[2] Linder, J. et al. (2025). Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation. Nature Genetics 57:949–961. Supports: the distance-stratified Pangolin-versus-Borzoi sQTL result (ΔAUPRC 0.01 at ≤10 kb favouring Pangolin; 0.02 at ≤200 bp favouring Borzoi) and the >0.02 gain from averaging both models' ranks. Location: Results ("Functional splicing variant interpretation"); Figs. 7d–e. Source paper: s41588-024-02053-6.pdf

[3] Avsec, Ž. et al. (2026). Advancing regulatory variant effect prediction with AlphaGenome. Nature 649:1206–1217. Supports: the MFASS auPRC values reported in the Disputed note, and Pangolin's position as best comparator on the ClinVar splicing benchmarks. Location: "Improved splicing variant predictions"; Figs. 3h, 3i. Source paper: s41586-025-10014-0.pdf