DNA is ordered in a 4D-tensor with for each training range, a one-hot encoding (4 element
vector) of all nucleotides in that range (assumed to always be 200). If
a sequence is unknown ('N' in FASTA format), the entire corresponding
one-hot vector is zero. Then, for all nucleotides seperately, ranges are
combined into bits in a uint8 array for compactness (see
numpy.unpackbits).
Eventually the data is pickled into dna.pkl with protocol 2.
ChIP-seq peak conservative calls are collected per range and per
protein-cell combination in a 3D-tensor. The data is then pickled into
chip.conservative.pkl with protocol 2.
For experiments with only one protein (currently the only experiment
supported) one can generate a protein-cell specific dataset by running
python3 pick_chip.py --cell X --protein Y, where X and Y are the
indices of both properties in the original tensor.
One can start training on a specific protein-cell pair by running
python3 train.py --epochs E --chip-path chip-X-Y.conservative.pkl.
Check python3 train.py --help for additional training options.