README.md · InstaDeepAI/genomics-long-range-benchmark at ea6154d266f23abce923a09961af9cdc2164e41b

metadata

license: cc-by-nc-sa-4.0
language:
  - en
tags:
  - Genomics
  - Benchmarks
  - Language Models
  - DNA
pretty_name: Genomics Long-Range Benchmark
viewer: false

Summary

The motivation of the genomics long-range benchmark (LRB) is to compile a set of biologically relevant genomic tasks requiring long-range dependencies which will act as a robust evaluation tool for genomic language models. While serving as a strong basis of evaluation, the benchmark must also be efficient and user-friendly. To achieve this we strike a balance between task complexity and computational cost through strategic decisions, such as down-sampling or combining datasets.

Benchmark Tasks

The Genomics LRB is a collection of nine tasks which can be loaded by passing in the corresponding task_name into the load_dataset function. All of the following datasets allow the user to specify an arbitrarily long sequence length, giving more context to the task, by passing the sequence_length kwarg to load_dataset. Additional task specific kwargs, if applicable, are mentioned in the sections below.
*Note that as you increase the context length to very large numbers you may start to reduce the size of the dataset since a large context size may cause indexing outside the boundaries of chromosomes.

Task	`task_name`	Sample Output	ML Task Type	# Outputs	# Train Seqs	# Test Seqs	Data Source
Variant Effect Causal eQTL	`variant_effect_causal_eqtl`	{ref sequence, alt sequence, label, tissue, chromosome,position, distance to nearest TSS}	SNP Classification	1	88717	8846	GTEx (via Enformer)
Variant Effect Pathogenic ClinVar	`variant_effect_pathogenic_clinvar`	{ref sequence, alt sequence, label, chromosome, position}	SNP Classification	1	38634	1018	ClinVar, gnomAD (via GPN-MSA)
Variant Effect Pathogenic OMIM	`variant_effect_pathogenic_omim`	{ref sequence, alt sequence, label,chromosome, position}	SNP Classification	1	-	2321473	OMIM, gnomAD (via GPN-MSA)
CAGE Prediction	`cage_prediction`	{sequence, labels, chromosome,label_start_position,label_stop_position}	Binned Regression	50 per bin	33891	1922	FANTOM5 (via Basenji)
Bulk RNA Expression	`bulk_rna_expression`	{sequence, labels, chromosome,position}	Seq-wise Regression	218	22827	990	GTEx, FANTOM5 (via ExPecto)
Chromatin Features Histone_Marks	`chromatin_features_histone_marks`	{sequence, labels,chromosome, position, label_start_position,label_stop_position}	Seq-wise Classification	20	2203689	227456	ENCODE, Roadmap Epigenomics (via DeepSea
Chromatin Features DNA_Accessibility	`chromatin_features_dna_accessibility`	{sequence, labels,chromosome, position, label_start_position,label_stop_position}	Seq-wise Classification	20	2203689	227456	ENCODE, Roadmap Epigenomics (via DeepSea)
Regulatory Elements Promoter	`regulatory_element_promoter`	{sequence, label,chromosome, start, stop, label_start_position,label_stop_position}	Seq-wise Classification	1	953376	96240	SCREEN
Regulatory Elements Enhancer	`regulatory_element_enhancer`	{sequence, label,chromosome, start, stop, label_start_position,label_stop_position}	Seq-wise Classification	1	1914575	192201	SCREEN

Usage Example

from datasets import load_dataset

# Use this parameter to download sequences of arbitrary length (see docs below for edge cases)
sequence_length=2048

# One of:
# ["variant_effect_causal_eqtl","variant_effect_pathogenic_clinvar",
# "variant_effect_pathogenic_omim","cage_prediction", "bulk_rna_expression",
# "chromatin_features_histone_marks","chromatin_features_dna_accessibility",
# "regulatory_element_promoter","regulatory_element_enhancer"] 

task_name = "variant_effect_causal_eqtl"

dataset = load_dataset(
    "InstaDeepAI/genomics-long-range-benchmark",
    task_name=task_name,
    sequence_length=sequence_length,
)

1. Variant Effect Causal eQTL

Predicting the effects of genetic variants, particularly expression quantitative trait loci (eQTLs), is essential for understanding the molecular basis of several diseases. eQTLs are genomic loci that are associated with variations in mRNA expression levels among individuals. By linking genetic variants to causal changes in mRNA expression, researchers can uncover how certain variants contribute to disease development.

Source

Original data comes from GTEx. Processed data in the form of vcf files for positive and negative variants across 49 different tissue types were obtained from the Enformer paper located here. Sequence data originates from the GRCh38 genome assembly.

Data Processing

Fine-mapped GTEx eQTLs originate from Wang et al, while the negative matched set of variants comes from Avsec et al . The statistical fine-mapping tool SuSiE was used to label variants. Variants from the fine-mapped eQTL set were selected and given positive labels if their posterior inclusion probability was > 0.9, as assigned by SuSiE. Variants from the matched negative set were given negative labels if their posterior inclusion probability was < 0.01.

Task Structure

Type: Binary classification

Task Args:
sequence_length: an integer type, the desired final sequence length

Input: a genomic nucleotide sequence centered on the SNP with the reference allele at the SNP location, a genomic nucleotide sequence centered on the SNP with the alternative allele at the SNP location, and tissue type
Output: a binary value referring to whether the variant has a causal effect on gene expression

Splits

Train: chromosomes 1-8, 11-22, X, Y
Test: chromosomes 9,10

2. Variant Effect Pathogenic ClinVar

A coding variant refers to a genetic alteration that occurs within the protein-coding regions of the genome, also known as exons. Such alterations can impact protein structure, function, stability, and interactions with other molecules, ultimately influencing cellular processes and potentially contributing to the development of genetic diseases. Predicting variant pathogenicity is crucial for guiding research into disease mechanisms and personalized treatment strategies, enhancing our ability to understand and manage genetic disorders effectively.

Source

Original data comes from ClinVar and gnomAD. However, we use processed data files from the GPN-MSA paper located here. Sequence data originates from the GRCh38 genome assembly.

Data Processing

Positive labels correspond to pathogenic variants originating from ClinVar whose review status was described as having at least a single submitted record with a classification but without assertion criteria. The negative set are variants that are defined as common from gnomAD. gnomAD version 3.1.2 was downloaded and filtered to variants with allele number of at least 25,000. Common variants were defined as those with MAF > 5%.

Task Structure

Type: Binary classification

Task Args:
sequence_length: an integer type, the desired final sequence length

Input: a genomic nucleotide sequence centered on the SNP with the reference allele at the SNP location, a genomic nucleotide sequence centered on the SNP with the alternative allele at the SNP location
Output: a binary value referring to whether the variant is pathogenic or not

Splits

Train: chromosomes 1-7, 9-22, X, Y
Test: chromosomes 8

3. Variant Effect Pathogenic OMIM

Predicting the effects of regulatory variants on pathogenicity is crucial for understanding disease mechanisms. Elements that regulate gene expression are often located in non-coding regions, and variants in these areas can disrupt normal cellular function, leading to disease. Accurate predictions can identify biomarkers and therapeutic targets, enhancing personalized medicine and genetic risk assessment.

Source

Original data comes from the Online Mendelian Inheritance in Man (OMIM) and gnomAD databases. However, we use processed data files from the GPN-MSA paper located here. Sequence data originates from the GRCh38 genome assembly.

Data Processing

Positive labeled data originates from a curated set of pathogenic variants located in the Online Mendelian Inheritance in Man (OMIM) catalog. The negative set is composed of variants that are defined as common from gnomAD. gnomAD version 3.1.2 was downloaded and filtered to variants with allele number of at least 25,000. Common variants were defined as those with minor allele frequency (MAF) > 5%.

Task Structure

Type: Binary classification

Task Args:
sequence_length: an integer type, the desired final sequence length
subset: a boolean type, whether to use the full dataset or a subset of the dataset (we provide this option as the full dataset has millions of samples)

Input: a genomic nucleotide sequence centered on the SNP with the reference allele at the SNP location, a genomic nucleotide sequence centered on the SNP with the alternative allele at the SNP location
Output: a binary value referring to whether the variant is pathogenic or not

Splits

Test: all chromosomes

4. CAGE Prediction

CAGE provides accurate high-throughput measurements of RNA expression by mapping TSSs at a nucleotide-level resolution. This is vital for detailed mapping of TSSs, understanding gene regulation mechanisms, and obtaining quantitative expression data to study gene activity comprehensively.

Source

Original CAGE data comes from FANTOM5. We used processed labeled data obtained from the Basenji paper which also used to train Enformer and is located here. Sequence data originates from the GRCh38 genome assembly.

Data Processing

The original dataset from the Basenji paper includes labels for 638 CAGE total tracks over 896 bins (each bin corresponding to 128 base pairs) totaling over ~70 GB. In the interest of dataset size and user-friendliness, only a subset of the labels are selected. From the 638 CAGE tracks, 50 of these tracks are selected with the following criteria:

Only select one cell line
Only keep mock treated and remove other treatments
Only select one donor

The [896 bins, 50 tracks] labels total in at ~7 GB. A description of the 50 included CAGE tracks can be found here cage_prediction/label_mapping.csv. *Note the data in this repository for this task has not already been log(1+x) normalized.

Task Structure

Type: Multi-variable regression
Because this task involves predicting expression levels for 128bp bins and there are 896 total bins in the dataset, there are in essence labels for 896 * 128 = 114,688 basepair sequences. If you request a sequence length smaller than 114,688 bps than the labels will be subsetted.

Task Args:
sequence_length: an integer type, the desired final sequence length, *must be a multiple of 128 given the binned nature of labels

Input: a genomic nucleotide sequence
Output: a variable length vector depending on the requested sequence length [requested_sequence_length / 128, 50]

Splits

Train/Test splits were maintained from Basenji and Enformer where randomly sampling was used to generate the splits. Note that for this dataset a validation set is also returned. In practice we merged the validation set with the train set and use cross validation to select a new train and validation set from this combined set.

5. Bulk RNA Expression

Gene expression involves the process by which information encoded in a gene directs the synthesis of a functional gene product, typically a protein, through transcription and translation. Transcriptional regulation determines the amount of mRNA produced, which is then translated into proteins. Developing a model that can predict RNA expression levels solely from sequence data is crucial for advancing our understanding of gene regulation, elucidating disease mechanisms, and identifying functional sequence variants.

Source

Original data comes from GTEx. We use processed data files from the ExPecto paper found here. Sequence data originates from the GRCh37/hg19 genome assembly.

Data Processing

The authors of ExPecto determined representative TSS for Pol II transcribed genes based on quantification of CAGE reads from the FANTOM5 project. The specific procedure they used is as follows, a CAGE peak was associated to a GENCODE gene if it was withing 1000 bps from a GENCODE v24 annotated TSS. The most abundant CAGE peak for each gene was then selected as the representative TSS. When no CAGE peak could be assigned to a gene, the annotated gene start position was used as the representative TSS. We log(1 + x) normalized then standardized the RNA-seq counts before training models. A list of names of tissues corresponding to the labels can be found here: bulk_rna_expression/label_mapping.csv. *Note the data in this repository for this task has already been log(1+x) normalized and standardized to mean 0 and unit variance.

Task Structure

Type: Multi-variable regression

Task Args:
sequence_length: an integer type, the desired final sequence length

Input: a genomic nucleotide sequence centered around the CAGE representative trancription start site
Output: a 218 length vector of continuous values corresponding to the bulk RNA expression levels in 218 different tissue types

Splits

Train: chromosomes 1-7,9-22,X,Y
Test: chromosome 8

6. Chromatin Features

Predicting chromatin features, such as histone marks and DNA accessibility, is crucial for understanding gene regulation, as these features indicate chromatin state and are essential for transcription activation.

Source

Original data used to generate labels for histone marks and DNase profiles comes from the ENCODE and Roadmap Epigenomics project. We used processed data files from the Deep Sea paper to build this dataset. Sequence data originates from the GRCh37/hg19 genome assembly.

Data Processing

The authors of DeepSea processed the data by chunking the human genome into 200 bp bins where for each bin labels were determined for hundreds of different chromatin features. Only bins with at least one transcription factor binding event were considered for the dataset. If the bin overlapped with a peak region of the specific chromatin profile by more than half of the sequence, a positive label was assigned. DNA sequences were obtained from the human reference genome assembly GRCh37. To make the dataset more accessible, we randomly sub-sampled the chromatin profiles from 125 to 20 tracks for the histones dataset and from 104 to 20 tracks for the DNA accessibility dataset.

Task Structure

Type: Multi-label binary classification

Input: a genomic nucleotide sequence centered on the 200 base pair bin that is associated with the labels
Output: a vector of length 20 with binary entries

Splits

Train set: chromosomes 1-7,10-22
Test set: chromosomes 8,9

7. Regulatory Elements

Cis-regulatory elements, such as promoters and enhancers, control the spatial and temporal expression of genes. These elements are essential for understanding gene regulation mechanisms and how genetic variations can lead to differences in gene expression.

Source

Original data annotations to build labels came from the Search Candidate cis-Regulatory Elements by ENCODE project. Sequence data originates from the GRCh38 genome assembly.

Data Processing

The data is processed as follows, we break the human reference genome into 200 bp non-overlapping chunks. If the 200 bp chunk overlaps by at least 50% or more with a contiguous region from the set of annotated cis-regulatory elements (promoters or enhancers), we label them as positive, else the chunk is labeled as negative. The resulting dataset was composed of ∼15M negative samples and ∼50k positive promoter samples and ∼1M positive enhancer samples. We randomly sub-sampled the negative set to 1M samples, and kept all positive samples, to make this dataset more manageable in size.

Task Structure

Type: Binary classification

Input: a genomic nucleotide sequence centered on the 200 base pair bin that is associated with the label
Output: a single binary value

Splits

Train set: chromosomes 1-7,10-22
Test set: chromosomes 8,9