Genomics · InstaDeepAI
nucleotide_transformer_downstream_tasks
A collection of 18 genomic classification benchmark tasks covering promoters, enhancers, splice sites, and histone marks.
What nucleotide_transformer_downstream_tasks is
nucleotide_transformer_downstream_tasks is a benchmark suite created by InstaDeepAI to evaluate foundation models on molecular biology tasks. Introduced alongside the Nucleotide Transformer, it gathers 18 sequence classification tasks across regulatory, transcription, splicing, and epigenetic features in human and multi-species genomes.
The dataset addresses the lack of standardised, comprehensive benchmarks for evaluating genomic machine learning models. By consolidating tasks from multiple genomics publications into fixed training and evaluation splits, it provides a consistent framework to test sequence understanding across varied sequence lengths and biological mechanisms.
Researchers can use this benchmark to test the predictive performance of representations from pre-trained language models and neural architectures on tasks including promoter detection, enhancer strength classification, splice site identification, and yeast histone mark prediction.
How it was made
The benchmark collates data from five distinct published studies into binary and multi-class classification setups. Promoter data, sourced from DeePromoter, contains 300-nucleotide sequences covering TATA and non-TATA promoters paired with negative sequences created by random sub-sampling. Enhancer tasks, taken from a deep learning framework study, include 200-nucleotide sequences with original and generatively augmented examples evaluated as binary enhancers or three-class enhancer types (strong, weak, or non-enhancer).
Splicing tasks comprise 400-nucleotide sequences from SpliceFinder (three-class classification across donor, acceptor, and non-splice sites) and 600-nucleotide binary classification sets for donor and acceptor sites from Spliceator. Finally, epigenetic tasks drawn from ChIP-chip yeast experiments provide ten binary classification datasets of 500-nucleotide sequences assessing nucleosome occupancy across acetylation and methylation marks: H3, H4, H3K9ac, H3K14ac, H4ac, H3K4me1, H3K4me2, H3K4me3, H3K36me3, and H3K79me3.
Key facts
- Release date
- 16 June 2023
- Number of tasks
- 18
- Sequence lengths
- 200, 300, 400, 500, and 600 nucleotides
- Data format
- Parquet
- Task types
- Binary and multi-class classification
What you can use it for
- Promoter Detection
- Train and evaluate models on 300 bp DNA sequences to identify active promoter regions, split into overall, TATA-box, and non-TATA promoter datasets.
- Enhancer Identification and Classification
- Classify 200 bp sequences as enhancers versus non-enhancers, or stratify them into non-enhancer, weak enhancer, and strong enhancer classes.
- Splice Site Recognition
- Predict RNA splice sites from 400 bp sequences across three classes (donor, acceptor, non-splice) or 600 bp binary sets specifically for donor or acceptor recognition.
Explore the data
No licence declared
The authors have not published a licence, which means no permission to redistribute the data has been granted to anyone. The rows are not served here; read them on the original dataset page.How the columns are distributed
Over 461,850 rows of default/train
label
int- min
- 0
- median
- 1
- mean
- 0.56
- max
- 2
name
string_text- min
- 5
- median
- 20
- mean
- 18.27
- max
- 30
sequence
string_text- min
- 199
- median
- 500
- mean
- 437.12
- max
- 600
task
string_label- promoter_all12%
- promoter_no_tata10%
- H3K4me37%
Strengths and limits
Good at
- Provides a standardised collection of 18 diverse genomic tasks spanning promoters, enhancers, splicing, and epigenetics.
- Supplies pre-defined training and test splits for all 18 tasks to facilitate reproducible model comparison.
- Covers multiple sequence context lengths ranging from 200 to 600 base pairs.
Watch out for
- Superseded by a revised version (nucleotide_transformer_downstream_tasks_revised) developed during peer review to improve benchmark robustness.
- Enhancer training splits rely partly on synthetic sequences produced by a generative model rather than purely biological sequences.
- Epigenetic and histone mark prediction tasks are limited to yeast ChIP-chip experimental data.
Questions
Can I look inside nucleotide_transformer_downstream_tasks without downloading it?
Not here: it declares no licence, so no permission to redistribute its rows has been granted. Its description and statistics are still shown.
What licence is nucleotide_transformer_downstream_tasks under?
Its authors have not declared a licence, which means no permission to reuse or redistribute it has been granted.
What tasks are included in this benchmark?
The benchmark includes 18 tasks: three promoter tasks (promoter_all, promoter_tata, promoter_no_tata), two enhancer tasks (enhancers, enhancers_types), three splicing tasks (splice_sites_all, splice_sites_acceptor, splice_sites_donor), and ten yeast histone mark tasks (H3, H4, H3K9ac, H3K14ac, H4ac, H3K4me1, H3K4me2, H3K4me3, H3K36me3, H3K79me3).
What is the sequence length for each benchmark task?
Enhancer sequences are 200 nucleotides; promoter sequences are 300 nucleotides; splice_sites_all sequences are 400 nucleotides; histone mark sequences are 500 nucleotides; and donor/acceptor splice site sequences are 600 nucleotides.
Is this the latest version of the Nucleotide Transformer benchmark?
No. The developers revised and improved the benchmark during peer review and recommend using the newer repository, nucleotide_transformer_downstream_tasks_revised, which is considered more robust.
How to cite it
The Nucleotide Transformer: Building and Evaluating Robust Foundation Models for Human Genomics, bioRxiv (2023), doi: 10.1101/2023.01.11.523679
Written from nucleotide_transformer_downstream_tasks's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.