Plant Genomics · InstaDeepAI
plant-genomic-benchmark
A multi-species plant genomics benchmark evaluating sequence models across eight regulatory and expression tasks.
What plant-genomic-benchmark is
The plant genomic benchmark is an evaluation suite created by InstaDeepAI to assess machine learning models on plant genomic sequence data. Introduced alongside the AgroNT foundational large language model for edible plant genomes, it spans eight distinct functional genomics tasks across both model plant organisms and key agricultural crops.
The suite evaluates models across single-variable regression, multi-variable regression, binary classification, and multi-label classification. It includes datasets covering polyadenylation site detection, splice site recognition, long non-coding RNA identification, promoter strength, terminator strength, chromatin accessibility, gene expression levels, and intergenic enhancer regions.
Beyond benchmark evaluation splits, the resource includes results from in silico saturation mutagenesis across the cassava (Manihot esculenta) genome. This dataset quantifies the predicted impact of more than 10 million sequence mutations on cassava gene expression levels and intergenic enhancer elements.
How it was made
The benchmark compiles genomic sequences of varying lengths corresponding to specific regulatory tasks. Sequence lengths range from 170 base pairs for promoter and terminator activity assays, to 398 or 400 base pairs for splice site and polyadenylation prediction, up to 1,000 base pairs for chromatin accessibility and enhancer regions, and 6,000 base pairs for gene expression regression and long non-coding RNA classification.
Individual tasks are formatted for specific learning objectives. Chromatin accessibility is structured as a multi-label classification problem across thousands of base pairs, while gene expression across tissues is structured as multi-variable regression. For cassava in silico saturation mutagenesis, tabulated genomic coordinates, reference alleles, and alternative alleles are provided with calculated log2 fold changes for enhancer probabilities and tissue-specific expression.
Key facts
- Number of evaluation tasks
- 8
- Task formulations
- Binary classification, multi-label classification, single-variable regression, multi-variable regression
- Input sequence lengths
- 170 to 6000 base pairs
- Mutagenesis scale
- Over 10 million in silico mutations in Manihot esculenta
- Covered organisms
- Arabidopsis thaliana, Oryza sativa, Zea mays, Glycine max, Manihot esculenta, Sorghum bicolor, Setaria italica, Brachypodium distachyon, Medicago truncatula, Trifolium pratense, Solanum lycopersicum, Triticum aestivum, and Chlamydomonas reinhardtii
What you can use it for
- Benchmarking genomic foundation models
- Evaluate DNA sequence models across diverse regulatory tasks including splice site recognition, polyadenylation site classification, and promoter strength prediction.
- Predicting multi-tissue gene expression
- Train and test regression models to predict expression levels across tissues from 6,000 base pair upstream sequences in crops such as maize, rice, soybean, and tomato.
- Profiling chromatin accessibility
- Perform multi-label classification to determine open chromatin regions across seven plant species and varieties, including rice, maize, and sorghum.
- Analysing variant consequences in cassava
- Examine log2 fold changes from over 10 million precomputed in silico mutations to evaluate impacts on cassava gene expression and intergenic enhancer function.
Explore the data
No queryable preview
This dataset is not available in a queryable form — usually an unusual file format or a repository too large to convert. The original source may still be readable.Strengths and limits
Good at
- Spans eight distinct genomic prediction problems covering diverse sequence lengths from 170 to 6,000 base pairs.
- Incorporates genomic data across both model plants and major agricultural crops including maize, rice, wheat, soybean, cassava, and tomato.
- Supplies extensive in silico saturation mutagenesis data covering more than 10 million variant effects in Manihot esculenta.
Watch out for
- Several tasks, including polyadenylation, splicing, and long non-coding RNA prediction, lack predefined validation partitions.
- Long non-coding RNA datasets for individual species are small, requiring joint training across species to improve model learning.
- Enhancer classification and saturation mutagenesis resources are restricted to a single species, Manihot esculenta.
Questions
Can I look inside plant-genomic-benchmark without downloading it?
Not here: it is not available in a queryable form. Its description and statistics are still shown.
What licence is plant-genomic-benchmark under?
Its authors declare cc-by-nc-sa-4.0. Read the terms on the original page before reusing the data.
Which plant species are included in the benchmark?
The suite includes datasets from Arabidopsis thaliana, Oryza sativa (indica and japonica groups), Zea mays, Glycine max, Manihot esculenta, Sorghum bicolor, Setaria italica, Brachypodium distachyon, Medicago truncatula, Trifolium pratense, Solanum lycopersicum, Triticum aestivum, and Chlamydomonas reinhardtii.
Why are long non-coding RNA datasets combined during training?
The individual species datasets for long non-coding RNA classification are relatively small, so datasets across species are combined into a single training set to support effective model learning.
What information is included in the cassava saturation mutagenesis tables?
The tables provide genomic coordinates, reference and alternative alleles, affected gene IDs or intergenic regions, tissue identifiers, and the calculated log2 fold changes in expression or enhancer probability resulting from each mutation.
How to cite it
Mendoza-Revilla & Trop et al., 2024. AgroNT (A Foundational Large Language Model for Edible Plant Genomes). DOI: 10.57967/hf/2464
Written from plant-genomic-benchmark'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.