This repository contains code supporting the data processing, model training, and analysis of the MpraNet model. This is a deep learning model trained to predict whether SNP sites have the kind of functional effect detected by MPRA assay. The model is trained on positive variants from existing MPRA experiments on the GM12878 cell line, with matched negatives sampled from the background genome. The model uses numerous features including epigenetic annotations and prior functional scores.

Features were obtained from ROADMAP, Eigen, and RegBase databases, and were processed with setup_project_data.py, setup_neighbor_data.py, and setup_mixed_datasets.py, which also generates the train and validation splits.

MPRANet was tuned with 5-fold CV using search_hparam.py, after which the final model was trained using fit_mpra_models.py

Subsequent analysis of the model was performed using evaluation scripts, such as evaluate_results.py and sample_size_sim.py.

We additionally provide an example script for using a trained MPRANet to predict on a user-given set of genomic variant locations in predict_script.py.