CNN for enhancing short read STR calls
Pre-processing steps:
- One-hot encode bases at each position
- Custom Dataset, Dataloader for base, depth, metadata
3 methods:
- Baseline CNN -> FCN with concatenated metadata
- Dilated self-designed kernel in CNN -> FCN w/concatenated metadata
- 2 convolutional networks with regular kernel -> aggregate -> FCN w/concatenate metadata
We are maintaining same number of convolutional/FCN layers and number of kernels for the 3 models so they aren't confounding.