Table 2 Transformer-45k and SpliceAI-10k performance on GTEx V8 splice junctions

From: Transformers significantly improve splice site prediction

 

PR-AUC [95% CI]

Top-k accuracy [95% CI]

Transformer-45k fine-tuned on GTEx V8

0.849 [0.848, 0.851]

0.758 (\(\frac{137,139}{180,872}\)) [0.757, 0.760]

SpliceAI-10k fine-tuned on GTEx V8

0.842 [0.841, 0.844]

0.751 (\(\frac{135,829}{180,872}\)) [0.750, 0.753]

SpliceAI-10k pre-trained weights

0.836 [0.835, 0.838]

0.747 (\(\frac{135,112}{180,872}\)) [0.746, 0.749]

SpliceAI-10k trained on ENSEMBL

0.775 [0.773, 0.777]

0.706 (\(\frac{127,737}{180,872}\)) [0.705 0.709]

Transformer-45k trained on ENSEMBL

0.772 [0.770, 0.774]

0.711 (\(\frac{128,651}{180,872}\)) [0.710, 0.713]

  1. These results are only for chromosomes left out during training, here the combined number of splice sites is 180,872. The performance metrics are PR-AUC and top-k accuracy. 95% confidence intervals (CIs) are shown in brackets and the best score is displayed in bold.
Back to article page