Genomics · InstaDeepAI

NTv3_benchmark_dataset

A multi-species genomic benchmark dataset providing functional tracks and structural genome annotations.

What NTv3_benchmark_dataset is

The NTv3 benchmark dataset is a genomic resource developed by InstaDeepAI for evaluating machine learning models on functional genomics tasks. It brings together biological signals and genomic features across seven diverse species: human, chicken, cattle, arabidopsis, maize, rice, and tomato.

The resource addresses the need for standardised sequence-to-signal and regression benchmarks in computational biology. By offering curated experimental assays alongside structural genomic annotations, it allows researchers to benchmark fine-tuned models on tasks ranging from chromatin accessibility to translational profiling.

How it was made

The dataset is organised per species into two distinct track directories: functional tracks and genome annotations. Functional genomics experiments are stored as BigWig files, a binary format for dense continuous genomic data, indexed by source repository run IDs (such as ENCSR, ERX, or SRR identifiers) and split by strand orientation where relevant.

Genome annotations are provided in BED format and specify coordinates for structural elements including exons, introns, splice acceptors, and start codons. Within each species evaluation setting, all functional signals and structural annotations share a common coordinate system.

Key facts

Data format: BigWig (.bigwig), BED (.bed), CSV (.csv)
Coverage: 7 species (human, chicken, cattle, arabidopsis, maize, rice, tomato)
Modalities: Tabular, continuous genomic signals, text
Task types: Sequence-to-signal, regression, fine-tuning

What you can use it for

Benchmarking sequence-to-signal models: Evaluate genomic foundation models on predicting continuous molecular assay profiles across different tissues, cell types, and species.
Chromatin and transcription regulation profiling: Train and test models on chromatin accessibility, PRO-cap transcription initiation, and histone modification tracks (H3K4me1, H3K4me3, H3K27ac).
RNA biology and post-transcriptional prediction: Assess prediction accuracy for protein-RNA interactions via eCLIP binding sites, gene expression levels, and mRNA translation using Ribo-seq tracks.
Genome feature identification: Test models on recognising structural annotations including splice acceptors, start codons, exons, and introns.

Explore the data

This dataset is gated

Its authors require their terms to be accepted before the data can be read. Accept them on the original dataset page. The platform does not accept terms on your behalf, so the rows are not shown here.

Strengths and limits

Good at

Spans seven distinct plant and animal species, enabling cross-species benchmark comparisons.
Covers multiple complementary functional genomics modalities, including accessibility, histone marks, transcription, translation, and RNA binding.
Uses standardised, widely supported bioinformatics file formats (BigWig and BED).

Watch out for

Track types are unevenly distributed across species; for example, cattle only contains genome annotations, while arabidopsis, maize, and rice only provide Ribo-seq functional tracks.
Files are indexed primarily by raw database IDs, requiring external benchmark documentation to map IDs to detailed sample descriptions.

Questions

Can I look inside NTv3_benchmark_dataset without downloading it?

Not here: its authors gate it behind terms each user must accept on the original page. Its description and statistics are still shown.

What licence is NTv3_benchmark_dataset under?

Its authors declare cc-by-nc-4.0. Read the terms on the original page before reusing the data.

Which species are included in the dataset?

The dataset contains data for seven species: human, chicken, cattle, arabidopsis, maize, rice, and tomato.

What types of functional genomics experiments are covered?

The functional tracks include chromatin accessibility, histone modifications (H3K4me1, H3K4me3, H3K27ac), transcription initiation (PRO-cap), RNA binding sites (eCLIP), gene expression, and mRNA translation (Ribo-seq).

Which structural genome annotations are available?

Genome annotation BED files cover exons, introns, splice acceptors, and start codons for available species.

How are the BigWig and BED files typically parsed in Python?

BigWig files can be parsed using libraries such as pyBigWig, whereas BED annotation files can be loaded using pyranges or pandas.

Written from NTv3_benchmark_dataset'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.