Additional analysis on SpliceBERT. The original repository is available at SpliceBERT.
We fine-tuned SpliceBERT on SpliceAI's GTEx dataset with R-Drop regularization for 5 times using different random seeds (model weights: Google Drive). The average AP scores of SpliceBERT (900nt) is comparable (donor) or slightly superior (acceptor) to SpliceAI-10K, while the ensemble model (averaging the predictions of 5 models) underperforms that of SpliceAI-10K, which is likely because that SpliceBERT models were fine-tuned based on the same pre-trained model and thus lack sufficient diversity.
The source codes are available in benchmark_spliceai-gtex.
model | receptive field size | AP (donor) | AP (acceptor) |
---|---|---|---|
SpliceBERT | 900 | 0.8547 |
0.8458 |
SpliceAI-10k | 10001 | 0.8547 |
0.8434 |
SpliceAI-2k | 2001 | 0.8369 |
0.8270 |
SpliceAI-400 | 401 | 0.7961 |
0.7873 |
SpliceAI-80 | 81 | 0.5216 |
0.4449 |
model (ensemble) | receptive field size | AP (donor) | AP (acceptor) |
---|---|---|---|
SpliceAI-10k (ensemble) | 10001 | 0.8735 | 0.8644 |
SpliceBERT (ensemble) | 900 | 0.8608 | 0.8524 |
Though SpliceBERT was pre-trained on primary RNA sequences, it can also be applied to DNA sequences. We finetuned SpliceBERT on DeepSTARR's dataset (https://zenodo.org/records/5502060) to identify sequences with potential enhancer activity. SpliceBERT outperformed DeepSTARR (convolution model) and Nucleotide Transformer (DNA language model). The results are available at benchmark_deepstarr.
model | Developmental | Housekeeping |
---|---|---|
SpliceBERT | 0.70 | 0.78 |
DeepSTARR | 0.68 | 0.74 |
Nucleotide Transformer (multi-species) | 0.64 | 0.75 |
SpliceBERT_on_DeepSTARR (show 20% points)
For any questions, contact chenkenbio_[at]_gmail.com
@article{chen2024self_bbae163,
title={Self-supervised learning on millions of primary RNA sequences from 72 vertebrates improves sequence-based RNA splicing prediction},
author={Chen, Ken and Zhou, Yue and Ding, Maolin and Wang, Yu and Ren, Zhixiang and Yang, Yuedong},
journal={Briefings in Bioinformatics},
volume={25},
number={3},
pages={bbae163},
year={2024},
publisher={Oxford University Press}
}