etrop commited on
Commit
66099dc
1 Parent(s): 46d85ff

update README

Browse files
Files changed (1) hide show
  1. README.md +151 -118
README.md CHANGED
@@ -3,36 +3,40 @@ license: cc-by-nc-sa-4.0
3
  language:
4
  - en
5
  tags:
6
- - biology
7
- - genomics
8
- pretty_name: Genomics Long Range Benchmark
 
 
9
  viewer: false
10
  ---
11
 
12
  ## Summary
13
- 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.
 
14
  While serving as a strong basis of evaluation, the benchmark must also be efficient and user-friendly.
15
  To achieve this we strike a balance between task complexity and computational cost through strategic decisions, such as down-sampling or combining datasets.
16
 
17
- ## Dataset Tasks
18
- The Genomics LRB is a collection of tasks which can be loaded by passing in the corresponding `task_name` into the `load_dataset` function. All of the following datasets
 
19
  allow the user to specify an arbitrarily long sequence length, giving more context
20
  to the task, by passing the `sequence_length` kwarg to `load_dataset`. Additional task
21
  specific kwargs, if applicable, are mentioned in the sections below.<br>
22
  *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
23
  cause indexing outside the boundaries of chromosomes.
24
 
25
- | Task | `task_name` | Sample Output | # Train Seqs | # Test Seqs |
26
- |--------------------------------------|----------------------------------------|-------------------------------------------------------------------------------------------| ------------ | ----------- |
27
- | Variant Effect Causal eQTL | `variant_effect_causal_eqtl` | {ref sequence, alt sequence, label, tissue, chromosome,position, distance to nearest TSS} | 88717 | 8846 |
28
- | Variant Effect Pathogenic Clinvar | `variant_effect_pathogenic_clinvar` | {ref sequence, alt sequence, label, chromosome, position} | 38634 | 1018 |
29
- | Variant Effect Pathogenic OMIM | `variant_effect_pathogenic_omim` | {ref sequence, alt sequence, label,chromosome, position} | - | 2321473 |
30
- | CAGE Prediction | `cage_prediction` | {sequence, labels, chromosome} | 33891 | 1922 |
31
- | Bulk RNA Expression | `bulk_rna_expression` | {sequence, labels, chromosome,position} | 22827 | 990 |
32
- | Chromatin Features Histone_Marks | `chromatin_features_histone_marks` | {sequence, labels,chromosome, position} | 2203689 | 227456 |
33
- | Chromatin Features DNA_Accessibility | `chromatin_features_dna_accessibility` | {sequence, labels,chromosome, position} | 2203689 | 227456 |
34
- | Regulatory Elements Promoter | `regulatory_element_promoter` | {sequence, label,chromosome, start, stop} | 953376 | 96240 |
35
- | Regulatory Elements Enhancer | `regulatory_element_enhancer` | {sequence, label,chromosome, start, stop} | 1914575 | 192201 |
36
 
37
  ## Usage Example
38
  ```python
@@ -57,190 +61,216 @@ dataset = load_dataset(
57
 
58
  ```
59
 
60
-
61
-
62
- ### 1. CAGE Prediction
63
- Cap Analysis Gene Expression(CAGE) is a biological assay used to measure the level of mRNA production rather than steady state values, taking into account both production and
64
- degradation. Being able to accurately predict mRNA levels as measured by CAGE is essential for deciphering tissue-specific expression patterns, transcriptional networks, and
65
- identifying differentially expressed genes with functional significance.
66
 
67
  #### Source
68
- Original CAGE data comes from FANTOM5. We used processed labeled data obtained from the [Basenji paper](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5932613/) which also used to train Enformer and is located [here](https://console.cloud.google.com/storage/browser/basenji_barnyard/data/human?pageState=(%22StorageObjectListTable%22:(%22f%22:%22%255B%255D%22))&prefix=&forceOnObjectsSortingFiltering=false).
 
 
69
  Sequence data originates from the GRCh38 genome assembly.
70
 
71
  #### Data Processing
72
- The original dataset from the Basenji paper includes labels for 638 CAGE total tracks over 896 bins (each bin corresponding to 128 base pairs)
73
- totaling over ~70 GB. In the interest of dataset size and user-friendliness, only a
74
- subset of the labels are selected.
75
- From the 638 CAGE tracks, 50 of these tracks are selected with the following criteria:
76
-
77
- 1. Only select one cell line
78
- 2. Only keep mock treated and remove other treatments
79
- 3. Only select one donor
80
-
81
- 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`.
82
 
83
  #### Task Structure
84
 
85
- Type: Multi-variable regression<br>
86
- 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
87
- you request a sequence length smaller than 114,688 bps than the labels will be subsetted.
88
 
89
  Task Args:<br>
90
- `sequence_length`: an interger type, the desired final sequence length, *must be a multiple of 128 given the binned nature of labels<br>
91
 
92
- Input: a genomic nucleotide sequence<br>
93
- Output: a variable length vector depending on the requested sequence length [requested_sequence_length / 128, 50]
 
94
 
95
  #### Splits
96
- 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
97
- set with the train set and use cross validation to select a new train and validation set from this combined set.
98
-
99
 
100
  ---
101
 
102
- ### 2. Bulk RNA Expression
103
- In comparison to CAGE, bulk RNA sequencing assays measure the steady state level (both transription and degradation) of mRNA in a population of cells.
 
 
 
104
 
105
  #### Source
106
- Original data comes from GTEx. We use processed data files from the [ExPecto paper](https://www.nature.com/articles/s41588-018-0160-6) found
107
- [here](https://github.com/FunctionLab/ExPecto/tree/master/resources). Sequence data originates from the GRCh37/hg19 genome assembly.
 
 
108
 
109
  #### Data Processing
110
- The continuous labels were log(1+x) transformed and standardized. A list of names of tissues corresponding to the labels can be found here: `bulk_rna_expression/label_mapping.csv`.
 
 
 
111
 
112
  #### Task Structure
113
 
114
- Type: Multi-variable regression<br>
115
 
116
  Task Args:<br>
117
- `sequence_length`: an interger type, the desired final sequence length<br>
118
 
119
- Input: a genomic nucleotide sequence centered around the CAGE representative trancription start site<br>
120
- Output: a 218 length vector of continuous values corresponding to the bulk RNA expression levels in 218 different tissue types
121
 
122
  #### Splits
123
- Train: chromosomes 1-7,9-22,X,Y<br>
124
- Test: chromosome 8
125
-
126
- #### Metrics
127
- Mean Spearman correlation across tissues <br>
128
- Mean Spearman correlation across genes <br>
129
- R<sup>2</sup>
130
 
131
  ---
132
 
133
- ### 3. Variant Effect Causal eQTL
134
- In genomics, a key objective is to predict how genetic variants affect gene expression in specific cell types.
 
 
135
 
136
  #### Source
137
- Original data comes from GTEx. However, we used processed data files from the [Enformer paper](https://www.nature.com/articles/s41592-021-01252-x) located [here](https://console.cloud.google.com/storage/browser/dm-enformer/data/gtex_fine/vcf?pageState=(%22StorageObjectListTable%22:(%22f%22:%22%255B%255D%22))&prefix=&forceOnObjectsSortingFiltering=false).
 
 
 
 
138
  Sequence data originates from the GRCh38 genome assembly.
139
 
140
  #### Data Processing
141
- In Enformer the datasets were partitioned in 48 different sets based on the tissue types. In our framing of the task we combine all samples across all tissues into one set
142
- and provide the tissue type along with each sample.
143
-
144
- As data files were used from Enformer, the labels were constructed according to their methodology - variants were labeled as 1 if their posterior inclusion probability was greater than 0.9 as
145
- assigned by the population-based fine-mapping tool SuSiE, while a matched set of negative variants was built with posterior inclusion probabilities of less than .01.
146
 
147
  #### Task Structure
148
 
149
  Type: Binary classification<br>
150
 
151
  Task Args:<br>
152
- `sequence_length`: an interger type, the desired final sequence length<br>
 
153
 
154
- 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<br>
155
- Output: a binary value refering to whether the variant has an effect on gene expression
156
 
157
  #### Splits
158
- Train: chromosomes 1-8, 11-22, X, Y<br>
159
- Test: chromosomes 9,10
160
 
161
  ---
162
 
163
- ### 4. Variant Effect Pathogenic ClinVar
164
- This task differs from the above in that variants in this datasets are linked to disease pathogenicity rather than effect on gene expression. Specifically, this tasks involves
165
- predicting whether a variant located in a coding region is pathogenic or is a common missense variant.
166
 
167
  #### Source
168
- Original data comes from ClinVar and gnomAD. However, we use processed data files from the [GPN-MSA paper](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10592768/) located here :https://huggingface.co/datasets/songlab/human_variants/blob/main/test.parquet .
 
 
169
  Sequence data originates from the GRCh38 genome assembly.
170
 
171
  #### Data Processing
172
- Human labelers generated the pathogenic missense set from ClinVar. The non-pathogenic set comes from common missense variants in gnomAD.
 
 
 
 
 
 
 
 
 
 
173
 
174
  #### Task Structure
175
 
176
- Type: Binary classification<br>
 
 
177
 
178
  Task Args:<br>
179
- `sequence_length`: an interger type, the desired final sequence length<br>
180
 
181
- 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<br>
182
- Output: a binary value referring to whether the variant is pathogenic or not
183
 
184
  #### Splits
185
- Train: chromosomes 1-7, 9-22, X, Y<br>
186
- Test: chromosomes 8
 
187
 
188
  ---
189
 
190
- ### 5. Variant Effect Pathogenic OMIM
191
- This tasks involves predicting whether a variant located in a regulatory region is
192
- implicated in a mendelian disease or is a common missense variant.
 
193
 
194
  #### Source
195
- Original data comes from the Online Mendelian Inheritance in Man (OMIM) and gnomAD
196
- databases.
197
- However, we use processed data files from the
198
- [GPN-MSA paper](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10592768/) located here :https://huggingface.co/datasets/songlab/omim/blob/main/test.parquet.
199
- Sequence data originates from the GRCh38 genome assembly.
200
 
201
  #### Data Processing
202
- Pathogenic variants originate from OMIM. The non-pathogenic set comes
203
- from common missense variants in gnomAD.
 
 
 
 
 
 
 
 
204
 
205
  #### Task Structure
206
 
207
- Type: Binary classification<br>
208
 
209
  Task Args:<br>
210
- `sequence_length`: an interger type, the desired final sequence length<br>
211
- `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)
212
 
213
- 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<br>
214
- Output: a binary value referring to whether the variant is pathogenic or not
215
 
216
  #### Splits
217
- Test: all chromosomes
 
218
 
219
  ---
220
  ### 6. Chromatin Features
221
- Chromatin refers to the mixture of DNA and proteins that form the chromosomes and allow the DNA to exist in a compact structure. There are many factors describing the state of
222
- chromatin which play important roles in regulating the cell. In this task we present two features which represent chromatin state, histone marks and DNA accessibility. Histones are proteins which in combination with
223
- a short strand of DNA form the basic unit of chromatin called the nucleosome. Histone marks correspond to the chemical modifications of histones which include methylation, acetylation,
224
- phosphorylation, and ubiquitination. DNA accessibility as measured through DNase I–hypersensitive sites refers to regions of the chromosome where DNA is not bound in a nucleosome but
225
- rather is open and accessible for other proteins such as transcription factors to bind.
226
 
227
  #### Source
228
  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](https://www.nature.com/articles/nmeth.3547) to build this dataset.
229
  Sequence data originates from the GRCh37/hg19 genome assembly.
230
 
231
  #### Data Processing
232
- As we use the data from Deep Sea the labels were generated according to the their following procedure - the genome was split into 200 base pairs bins and bins were labeled
233
- as 1 if more than half of the bin overlapped with the peak region of the chromatin feature annotations and 0 otherwise.
234
-
235
- Data from Deep Sea contains 919 labels for various chromatin features - 690 TF binding profiles, 125 DNase profiles and 104 histone-mark profiles. To make the dataset more accessible
236
- while still providing a representative set, we randomly sampled 20 histone marks and 20 DNase tissues.
 
 
 
 
237
 
238
  #### Task Structure
239
 
240
  Type: Multi-label binary classification
241
 
242
  Task Args:<br>
243
- `sequence_length`: an interger type, the desired final sequence length<br>
244
  `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)
245
 
246
  Input: a genomic nucleotide sequence centered on the 200 base pair bin that is associated with the labels<br>
@@ -251,26 +281,29 @@ Train set: chromosomes 1-7,10-22<br>
251
  Test set: chromosomes 8,9
252
 
253
  ---
254
- ### 7. Regulatory Element
255
- Regulatory elements are segments of DNA that play a pivotal role in controlling gene expression and cellular processes.
256
- These elements include promoters, enhancers, and other regions that interact with transcription factors and other proteins
257
- to modulate the activity of genes. Identifying regulatory elements is essential for understanding the complex regulatory
258
- networks governing cellular functions.
259
 
260
  #### Source
261
  Original data annotations to build labels came from the Search Candidate cis-Regulatory Elements by ENCODE project. Sequence data originates from the GRCh38
262
  genome assembly.
263
 
264
  #### Data Processing
265
- To generate labels from the annotations we first split the genome into 200 base pairs bins. We label a bin as 1 if more than half of the bin overlaps with any regulatory
266
- element given in the annotations and 0 otherwise.
 
 
 
 
 
267
 
268
  #### Task Structure
269
 
270
  Type: Binary classification
271
 
272
  Task Args:<br>
273
- `sequence_length`: an interger type, the desired final sequence length<br>
274
  `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)
275
 
276
  Input: a genomic nucleotide sequence centered on the 200 base pair bin that is associated with the label<br>
 
3
  language:
4
  - en
5
  tags:
6
+ - Genomics
7
+ - Benchmarks
8
+ - Language Models
9
+ - DNA
10
+ pretty_name: Genomics Long-Range Benchmark
11
  viewer: false
12
  ---
13
 
14
  ## Summary
15
+ The motivation of the genomics long-range benchmark (LRB) is to compile a set of
16
+ biologically relevant genomic tasks requiring long-range dependencies which will act as a robust evaluation tool for genomic language models.
17
  While serving as a strong basis of evaluation, the benchmark must also be efficient and user-friendly.
18
  To achieve this we strike a balance between task complexity and computational cost through strategic decisions, such as down-sampling or combining datasets.
19
 
20
+ ## Benchmark Tasks
21
+ The Genomics LRB is a collection of nine tasks which can be loaded by passing in the
22
+ corresponding `task_name` into the `load_dataset` function. All of the following datasets
23
  allow the user to specify an arbitrarily long sequence length, giving more context
24
  to the task, by passing the `sequence_length` kwarg to `load_dataset`. Additional task
25
  specific kwargs, if applicable, are mentioned in the sections below.<br>
26
  *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
27
  cause indexing outside the boundaries of chromosomes.
28
 
29
+ | Task | `task_name` | Sample Output | ML Task Type | # Outputs | # Train Seqs | # Test Seqs | Data Source |
30
+ |-------|-------------|---------------|-------------------------|-------------|--------------|----------- |----------- |
31
+ | 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](https://www.nature.com/articles/s41592-021-01252-x)) |
32
+ | 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](https://www.biorxiv.org/content/10.1101/2023.10.10.561776v1)) |
33
+ | Variant Effect Pathogenic OMIM | `variant_effect_pathogenic_omim` | {ref sequence, alt sequence, label,chromosome, position} | SNP Classification | 1 | - | 2321473 |OMIM, gnomAD (via [GPN-MSA](https://www.biorxiv.org/content/10.1101/2023.10.10.561776v1)) |
34
+ | CAGE Prediction | `cage_prediction` | {sequence, labels, chromosome} | Binned Regression | 50 per bin | 33891 | 1922 | FANTOM5 (via [Basenji](https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1008050)) |
35
+ | Bulk RNA Expression | `bulk_rna_expression` | {sequence, labels, chromosome,position} | Seq-wise Regression | 218 | 22827 | 990 | GTEx, FANTOM5 (via [ExPecto](https://www.nature.com/articles/s41588-018-0160-6)) |
36
+ | Chromatin Features Histone_Marks | `chromatin_features_histone_marks` | {sequence, labels,chromosome, position} | Seq-wise Classification | 20 | 2203689 | 227456 | ENCODE, Roadmap Epigenomics (via [DeepSea](https://pubmed.ncbi.nlm.nih.gov/30013180/) |
37
+ | Chromatin Features DNA_Accessibility | `chromatin_features_dna_accessibility` | {sequence, labels,chromosome, position} | Seq-wise Classification | 20 | 2203689 | 227456 | ENCODE, Roadmap Epigenomics (via [DeepSea](https://pubmed.ncbi.nlm.nih.gov/30013180/)) |
38
+ | Regulatory Elements Promoter | `regulatory_element_promoter` | {sequence, label,chromosome, start, stop} | Seq-wise Classification | 1| 953376 | 96240 | SCREEN |
39
+ | Regulatory Elements Enhancer | `regulatory_element_enhancer` | {sequence, label,chromosome, start, stop} | Seq-wise Classification | 1| 1914575 | 192201 | SCREEN |
40
 
41
  ## Usage Example
42
  ```python
 
61
 
62
  ```
63
 
64
+ ### 1. Variant Effect Causal eQTL
65
+ Predicting the effects of genetic variants, particularly expression quantitative trait loci (eQTLs), is essential for understanding the molecular basis of several diseases.
66
+ eQTLs are genomic loci that are associated with variations in mRNA expression levels among individuals.
67
+ By linking genetic variants to causal changes in mRNA expression, researchers can
68
+ uncover how certain variants contribute to disease development.
 
69
 
70
  #### Source
71
+ Original data comes from GTEx. Processed data in the form of vcf files for positive
72
+ and negative variants across 49 different tissue types were obtained from the
73
+ [Enformer paper](https://www.nature.com/articles/s41592-021-01252-x) located [here](https://console.cloud.google.com/storage/browser/dm-enformer/data/gtex_fine/vcf?pageState=%28%22StorageObjectListTable%22:%28%22f%22:%22%255B%255D%22%29%29&prefix=&forceOnObjectsSortingFiltering=false).
74
  Sequence data originates from the GRCh38 genome assembly.
75
 
76
  #### Data Processing
77
+ Fine-mapped GTEx eQTLs originate from [Wang et al](https://www.nature.com/articles/s41467-021-23134-8), while the negative matched set of
78
+ variants comes from [Avsec et al](https://www.nature.com/articles/s41592-021-01252-x)
79
+ . The statistical fine-mapping tool SuSiE was used to label variants.
80
+ Variants from the fine-mapped eQTL set were selected and given positive labels if
81
+ their posterior inclusion probability was > 0.9,
82
+ as assigned by SuSiE. Variants from the matched negative set were given negative labels if their
83
+ posterior inclusion probability was < 0.01.
 
 
 
84
 
85
  #### Task Structure
86
 
87
+ Type: Binary classification<br>
 
 
88
 
89
  Task Args:<br>
90
+ `sequence_length`: an integer type, the desired final sequence length<br>
91
 
92
+ 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<br>
93
+ Output: a binary value referring to whether the variant has a causal effect on gene
94
+ expression
95
 
96
  #### Splits
97
+ Train: chromosomes 1-8, 11-22, X, Y<br>
98
+ Test: chromosomes 9,10
 
99
 
100
  ---
101
 
102
+ ### 2. Variant Effect Pathogenic ClinVar
103
+ A coding variant refers to a genetic alteration that occurs within the protein-coding regions of the genome, also known as exons.
104
+ Such alterations can impact protein structure, function, stability, and interactions
105
+ with other molecules, ultimately influencing cellular processes and potentially contributing to the development of genetic diseases.
106
+ 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.
107
 
108
  #### Source
109
+ Original data comes from ClinVar and gnomAD. However, we use processed data files
110
+ from the [GPN-MSA paper](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10592768/)
111
+ located [here](https://huggingface.co/datasets/songlab/human_variants/blob/main/test.parquet).
112
+ Sequence data originates from the GRCh38 genome assembly.
113
 
114
  #### Data Processing
115
+ Positive labels correspond to pathogenic variants originating from ClinVar whose review status was
116
+ described as having at least a single submitted record with a classification but without assertion criteria.
117
+ 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
118
+ variants were defined as those with MAF > 5%.
119
 
120
  #### Task Structure
121
 
122
+ Type: Binary classification<br>
123
 
124
  Task Args:<br>
125
+ `sequence_length`: an integer type, the desired final sequence length<br>
126
 
127
+ 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<br>
128
+ Output: a binary value referring to whether the variant is pathogenic or not
129
 
130
  #### Splits
131
+ Train: chromosomes 1-7, 9-22, X, Y<br>
132
+ Test: chromosomes 8
 
 
 
 
 
133
 
134
  ---
135
 
136
+ ### 3. Variant Effect Pathogenic OMIM
137
+ Predicting the effects of regulatory variants on pathogenicity is crucial for understanding disease mechanisms.
138
+ 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.
139
+ Accurate predictions can identify biomarkers and therapeutic targets, enhancing personalized medicine and genetic risk assessment.
140
 
141
  #### Source
142
+ Original data comes from the Online Mendelian Inheritance in Man (OMIM) and gnomAD
143
+ databases.
144
+ However, we use processed data files from the
145
+ [GPN-MSA paper](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10592768/) located [here](
146
+ https://huggingface.co/datasets/songlab/omim/blob/main/test.parquet).
147
  Sequence data originates from the GRCh38 genome assembly.
148
 
149
  #### Data Processing
150
+ Positive labeled data originates from a curated set of pathogenic variants located
151
+ in the Online Mendelian Inheritance in Man (OMIM) catalog. The negative set is
152
+ composed of variants that are defined as common from gnomAD. gnomAD version 3.1.2 was downloaded and filtered to variants with
153
+ allele number of at least 25,000. Common variants were defined as those with minor allele frequency
154
+ (MAF) > 5%.
155
 
156
  #### Task Structure
157
 
158
  Type: Binary classification<br>
159
 
160
  Task Args:<br>
161
+ `sequence_length`: an integer type, the desired final sequence length<br>
162
+ `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)
163
 
164
+ 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<br>
165
+ Output: a binary value referring to whether the variant is pathogenic or not
166
 
167
  #### Splits
168
+ Test: all chromosomes
 
169
 
170
  ---
171
 
172
+ ### 4. CAGE Prediction
173
+ CAGE provides accurate high-throughput measurements of RNA expression by mapping TSSs at a nucleotide-level resolution.
174
+ This is vital for detailed mapping of TSSs, understanding gene regulation mechanisms, and obtaining quantitative expression data to study gene activity comprehensively.
175
 
176
  #### Source
177
+ Original CAGE data comes from FANTOM5. We used processed labeled data obtained from
178
+ the [Basenji paper](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5932613/) which
179
+ also used to train Enformer and is located [here](https://console.cloud.google.com/storage/browser/basenji_barnyard/data/human?pageState=%28%22StorageObjectListTable%22:%28%22f%22:%22%255B%255D%22%29%29&prefix=&forceOnObjectsSortingFiltering=false).
180
  Sequence data originates from the GRCh38 genome assembly.
181
 
182
  #### Data Processing
183
+ The original dataset from the Basenji paper includes labels for 638 CAGE total tracks over 896 bins (each bin corresponding to 128 base pairs)
184
+ totaling over ~70 GB. In the interest of dataset size and user-friendliness, only a
185
+ subset of the labels are selected.
186
+ From the 638 CAGE tracks, 50 of these tracks are selected with the following criteria:
187
+
188
+ 1. Only select one cell line
189
+ 2. Only keep mock treated and remove other treatments
190
+ 3. Only select one donor
191
+
192
+ 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`.
193
+ *Note the data in this repository for this task has not already been log(1+x) normalized.
194
 
195
  #### Task Structure
196
 
197
+ Type: Multi-variable regression<br>
198
+ 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
199
+ you request a sequence length smaller than 114,688 bps than the labels will be subsetted.
200
 
201
  Task Args:<br>
202
+ `sequence_length`: an integer type, the desired final sequence length, *must be a multiple of 128 given the binned nature of labels<br>
203
 
204
+ Input: a genomic nucleotide sequence<br>
205
+ Output: a variable length vector depending on the requested sequence length [requested_sequence_length / 128, 50]
206
 
207
  #### Splits
208
+ 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
209
+ set with the train set and use cross validation to select a new train and validation set from this combined set.
210
+
211
 
212
  ---
213
 
214
+ ### 5. Bulk RNA Expression
215
+ 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.
216
+ 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
217
+ data is crucial for advancing our understanding of gene regulation, elucidating disease mechanisms, and identifying functional sequence variants.
218
 
219
  #### Source
220
+ Original data comes from GTEx. We use processed data files from the [ExPecto paper](https://www.nature.com/articles/s41588-018-0160-6) found
221
+ [here](https://github.com/FunctionLab/ExPecto/tree/master/resources). Sequence data originates from the GRCh37/hg19 genome assembly.
 
 
 
222
 
223
  #### Data Processing
224
+ The authors of ExPecto determined representative TSS for Pol II transcribed genes
225
+ based on quantification of CAGE reads from the FANTOM5 project. The specific procedure they used is as
226
+ follows, a CAGE peak was associated to a GENCODE gene if it was withing 1000 bps from a
227
+ GENCODE v24 annotated TSS. The most abundant CAGE peak for each gene was then selected
228
+ as the representative TSS. When no CAGE peak could be assigned to a gene, the annotated gene
229
+ start position was used as the representative TSS. We log(1 + x) normalized then standardized the
230
+ RNA-seq counts before training models. A list of names of tissues corresponding to
231
+ the labels can be found here: `bulk_rna_expression/label_mapping.csv`. *Note the
232
+ data in this repository for this task has already been log(1+x) normalized and
233
+ standardized to mean 0 and unit variance.
234
 
235
  #### Task Structure
236
 
237
+ Type: Multi-variable regression<br>
238
 
239
  Task Args:<br>
240
+ `sequence_length`: an integer type, the desired final sequence length<br>
 
241
 
242
+ Input: a genomic nucleotide sequence centered around the CAGE representative trancription start site<br>
243
+ Output: a 218 length vector of continuous values corresponding to the bulk RNA expression levels in 218 different tissue types
244
 
245
  #### Splits
246
+ Train: chromosomes 1-7,9-22,X,Y<br>
247
+ Test: chromosome 8
248
 
249
  ---
250
  ### 6. Chromatin Features
251
+ 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.
 
 
 
 
252
 
253
  #### Source
254
  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](https://www.nature.com/articles/nmeth.3547) to build this dataset.
255
  Sequence data originates from the GRCh37/hg19 genome assembly.
256
 
257
  #### Data Processing
258
+ The authors of DeepSea processed the data by chunking the human genome
259
+ into 200 bp bins where for each bin labels were determined for hundreds of different chromatin
260
+ features. Only bins with at least one transcription factor binding event were
261
+ considered for the dataset. If the bin overlapped with a peak region of the specific
262
+ chromatin profile by more than half of the
263
+ sequence, a positive label was assigned. DNA sequences were obtained from the human reference
264
+ genome assembly GRCh37. To make the dataset more accessible, we randomly sub-sampled the
265
+ chromatin profiles from 125 to 20 tracks for the histones dataset and from 104 to 20 tracks for the
266
+ DNA accessibility dataset.
267
 
268
  #### Task Structure
269
 
270
  Type: Multi-label binary classification
271
 
272
  Task Args:<br>
273
+ `sequence_length`: an integer type, the desired final sequence length<br>
274
  `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)
275
 
276
  Input: a genomic nucleotide sequence centered on the 200 base pair bin that is associated with the labels<br>
 
281
  Test set: chromosomes 8,9
282
 
283
  ---
284
+ ### 7. Regulatory Elements
285
+ Cis-regulatory elements, such as promoters and enhancers, control the spatial and temporal expression of genes.
286
+ These elements are essential for understanding gene regulation mechanisms and how genetic variations can lead to differences in gene expression.
 
 
287
 
288
  #### Source
289
  Original data annotations to build labels came from the Search Candidate cis-Regulatory Elements by ENCODE project. Sequence data originates from the GRCh38
290
  genome assembly.
291
 
292
  #### Data Processing
293
+ The data is processed as follows, we break the human
294
+ reference genome into 200 bp non-overlapping chunks. If the 200 bp chunk overlaps by at least 50%
295
+ or more with a contiguous region from the set of annotated cis-regulatory elements (promoters or
296
+ enhancers), we label them as positive, else the chunk is labeled as negative. The resulting dataset
297
+ was composed of ∼15M negative samples and ∼50k positive promoter samples and ∼1M positive
298
+ enhancer samples We randomly sub-sampled the negative set to 1M samples, and kept all positive
299
+ samples, to make this dataset more manageable in size.
300
 
301
  #### Task Structure
302
 
303
  Type: Binary classification
304
 
305
  Task Args:<br>
306
+ `sequence_length`: an integer type, the desired final sequence length<br>
307
  `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)
308
 
309
  Input: a genomic nucleotide sequence centered on the 200 base pair bin that is associated with the label<br>