Expression data:
- Table S2 from "Condition-dependent transcriptome reveals high-level regulatory architecture in Bacillus subtilis" (Nicolas, Mäder, Dervyn et al., 2012). Stored as
expression_data.xlsx. Genome sequence: - Bacillus subtilis subsp. subtilis str. 168 genome sequence downloaded from NCBI, reference NC_000964.3. Stored as
bsubt168.fna. Experimental growth conditions: - The different experimental growth conditions tested are summarized in
experiments_order.txt. More information about what exactly the conditions correspond to can be found in the paper by Nicolas et al.
To generate the cleaned dataset, click through the jupyter notebook data processing.ipynb
The notebook performs the following processing steps:
- Retrieve the sequences for the regions 20, 50, 100, 200, 1000bp upstream of each gene
- Compute the mean expression level across different replicates for the same conditions
- For each gene, store all of the expression data in a single array
- For each gene, also compute a normalized (z-scored) array of expression data
- Store all this data in
cleaned_data.csv - Store the order of the conditions stored in the expression array in
experiments_order.txt - Visualize correlatedness of growth conditions
To follow the training protocol, click through the jupyter notebook train CNN.ipynb
Two separate training tasks were performed
- A CNN binary classification model was trained on each experimental growth condition to predict whether a given upstream DNA sequence would lead to up- or downregulation of the gene in the given growth condition
- A CNN regression model was trained to predict the full expression profile across all experimental conditions for a given upstream DNA sequence