tom-ellis-lab/yeast-RNA-seq

Yeast RNA-seq dataset (for Yorzoi training)

This is the processed training data for Yorzoi. It consists of:

• samples.pkl
• track_values/*

samples.pkl contains a bit less than 16k rows, each row being one training sample. The track_values folder contains the actual track_values which are the prediction targets for Yorzoi during training and evaluation.

samples.pkl

samples.pkl has 12 columns:

1. chr describes the chromosome that the sequence comes from. E.g. 926e5_NC_000017.11 is from one of the human inserts (NC_000017.11 is chromosome 17 of the human genome)
2. strand (either + or -) dictates which strand the sequence comes from (not the sequence orientation!)

The next few columns are all genomic positions. No matter which strand the sequence comes from, the positions refer to the axis from 5' to 3'. E.g. strand=+ start=20 is the exact same position as strand=- start=20

3. start_seq defines the genomic start position of the sequence
4. end_seq defines the genomic end position of the sequence (inclusive)

The prediction target of our model are RNA-seq coverage values. These coverage values mostly represent the central 3 kb of the sequence input window.

5. start_coverage: the genomic start position of the RNA-seq coverage
6. end_coverage: the genomic end position of the RNA-seq coverage
7. coverage_sequence_3kb: DNA sequence of the region for which we predict coverage
8. input_sequence: the DNA sequence (of varying length, most of them 5,000 bp long) that serves as the input for our model. This assumes that your dataloader does the padding for sequences shorter than 4,992 bp and that your model crops sequences longer than 4,992 bp.
9. fold: either train, val or test which describes whether the sample is used for training, validation or testing.
10. track_values is the path to a compressed numpy array file storing the actual coverage values. They can be loaded using np.load(path)['a'].
11. hom_graph_id are node names for the homology graph that stores the similarity of sample sequences. We are currently in the process of preparing the homology graph file for distribution.

track_values/*

All coverage value matrices are of shape 162 x 3000. Coverage values are not binned and in their original scale (i.e. not squashed). An example name of a file is row_759h10_NC_000004.12_-_1031_4030.npz

Raw Data

We are currently in the process of preparing the raw data for distribution.