ChromBPNet PyTorch Lightning Implementation

A PyTorch Lightning reimplementation of ChromBPNet developed during fall BMI rotation with the Corces Lab. ChromBPNet is a deep learning model for predicting chromatin accessibility from DNA sequence.

Project Structure

models/

• base_modules.py: Contains the fundamental building blocks of the neural network architecture
• _module.py: Arranges the base modules into the complete model architecture
• _model.py: Contains the PyTorch Lightning trainer class implementation
• _data.py: Implements data loading and preprocessing functionality

testing/

• notebooks/: Jupyter notebooks used for model development, testing, and comparisons with the original ChromBPNet implementation
• microglia_train.sh and microglia_train_new.sh: Training scripts for microglia-specific experiments

utils/

• adjust_bed.py: Utilities for handling BED file format adjustments
• attention_utils.py: Helper functions for attention mechanisms
• data_utils.py: General data processing utilities
• losses.py: Loss function implementations (Note: NLL implementation currently in progress)
• one_hot.py: Functions for one-hot encoding of DNA sequences
• shape_utils.py: Utilities for handling tensor shapes and transformations

Development TODOs:

• Complete NLL loss implementation
• Personalized Genome integration (paired WGS, scATAC-Seq)