Interpretability Analysis for Convolutional-Hybrid Architectures

This repository explores different convolutional-hybrid architectures to quantitatively analyze the extent that first layer filters learn motif representations (Task 1) and quantitatively measure how reliable their saliency maps recapitulate ground truth patterns (Task 2).

Dependencies

• Python 3.5 or greater
• Pandas, NumPy, Matplotlib, H5py
• TFomics -- https://github.com/p-koo/tfomics
• TensorFlow 2.0 or greater
• Logomaker (Tareen and Kinney)

Source files

• models.py - many hybrid model architectures
• task1_pipeline.py - callable function to fit synthetic multi-task data and perform tomtom analysis for first layer filters
• task2_pipeline.py - callable function to fit sythetic regulatory code data and perform saliency analysis and quantify efficacy of saliency maps
• commands_task1.py - list of function calls to task1_pipeline.py for various hybrid architectures
• commands_task2.py - list of function calls to task2_pipeline.py for various hybrid architectures
• Analyze_task1_results.ipynb - notebook to reproduce figure 1
• Analyze_task2_results.ipynb - notebook to reproduce figure 2

Data

• synthetic_dataset.h5 - data used in Koo & Eddy, PLoS Comp Bio (2019) -- multi-task classification for 12 TF binding sites
• synthetic_code_dataset.h5 - data used in Koo & Ploenzke, Nature Machine Intelligence (2021) -- synthetic regulatory code prediction task