InstaDeepAI/plant-genomic-benchmark

Dataset Overview

This dataset features the 8 evaluation tasks presented in the AgroNT (A Foundational Large Language Model for Edible Plant Genomes) paper. The tasks cover single output regression, multi output regression, binary classification, and multi-label classification which aim to provide a comprehensive plant genomics benchmark. Additionally, we provide results from in silico saturation mutagenesis analysis of sequences from the cassava genome, assessing the impact of >10 million mutations on gene expression levels and enhancer elements. See the ISM section below for details regarding the data from this analysis.

Name	# of Datasets(Species)	Task Type	Sequence Length (base pair)
Polyadenylation	6	Binary Classification	400
Splice Site	2	Binary Classification	398
LncRNA	6	Binary Classification	101-6000
Promoter Strength	2	Single Variable Regression	170
Terminator Strength	2	Single Variable Regression	170
Chromatin Accessibility	7	Multi-label Classification	1000
Gene Expression	6	Multi-Variable Regression	6000
Enhancer Region	1	Binary Classification	1000

Dataset Sizes

Task Name	# Train Samples	# Validation Samples	# Test Samples
poly_a.arabidopsis_thaliana	170835	---	30384
poly_a.oryza_sativa_indica_group	98139	---	16776
poly_a.trifolium_pratense	111138	---	13746
poly_a.medicago_truncatula	47277	---	8850
poly_a.chlamydomonas_reinhardtii	90378	---	10542
poly_a.oryza_sativa_japonica_group	120621	---	20232
splicing.arabidopsis_thaliana_donor	2588034	---	377873
splicing.arabidopsis_thaliana_acceptor	1704844	---	250084
lncrna.m_esculenta	4934	---	360
lncrna.z_mays	8423	---	1629
lncrna.g_max	11430	---	490
lncrna.s_lycopersicum	7274	---	1072
lncrna.t_aestivum	11252	---	1810
lncrna.s_bicolor	8654	---	734
promoter_strength.leaf	58179	6825	7154
promoter_strength.protoplast	61051	7162	7595
terminator_strength.leaf	43294	5309	4806
terminator_strength.protoplast	43289	5309	4811
gene_exp.glycine_max	47136	4803	4803
gene_exp.oryza_sativa	31244	3702	3702
gene_exp.solanum_lycopersicum	27321	3827	3827
gene_exp.zea_mays	34493	4483	4483
gene_exp.arabidopsis_thaliana	25731	3401	3402
chromatin_access.oryza_sativa_MH63_RS2	5120000	14848	14848
chromatin_access.setaria_italica	5120000	19968	19968
chromatin_access.oryza_sativa_ZS97_RS2	5120000	14848	14848
chromatin_access.arabidopis_thaliana	5120000	9984	9984
chromatin_access.brachypodium_distachyon	5120000	14848	14848
chromatin_access.sorghum_bicolor	5120000	29952	29952
chromatin_access.zea_mays	6400000	79872	79872
pro_seq.m_esculenta	16852	1229	812

*** It is important to note that fine-tuning for lncrna was carried out using all datasets in a single training. The reason for this is that the datasets are small and combining them helped to improve learning.

Example Usage

from datasets import load_dataset

task_name='terminator_strength.protoplast' # one of the task names from the above table

dataset = load_dataset("InstaDeepAI/plant-genomic-benchmark",task_name=task_name)

In Silico Saturation Mutagensis

File structure for: ISM_Tables/Mesculenta_305_v6_PROseq_ISM_LOG2FC.txt.gz

Intergenic enhancer regions based on Lozano et al. 2021 (https://pubmed.ncbi.nlm.nih.gov/34499719/)
Genome version: Manihot esculenta reference genome v6.1 from Phytozome
CHR: Chromosome
POS: Physical position (bp)
REF: Reference allele
ALT: Alternative allele
LOG2FC: Log fold change in Intergenic enhancer probability (log2(p_mutated_sequence / p_original_sequence))

File structure for: ISM_Tables/Mesculenta_v6_GeneExpression_ISM_LOG2FC.txt.gz

Gene expression prediction based on: Wilson et al. 2016 (https://pubmed.ncbi.nlm.nih.gov/28116755/)
Genome version: Manihot esculenta reference genome v6 from Ensembl 56
CHR: Chromosome
POS: Physical position (bp)
REF: Reference allele
ALT: Alternative allele
GENE: Gene ID
STRAND: Gene strand
TISSUE: Tissue type (Acronyms detailed in Figure 1 of Wilson et al.)
LOG2FC: Gene expression log fold change (log2(gene_exp_mutated_sequence / gene_exp_original_sequence))