DeepMotifSyn: a deep learning approach to synthesize heterodimeric DNA motifs

Introduction

DeepMotifSyn is a deep-learning-based model to synthesize heterodimeric motifs from monomeric motif pairs.

This model consists of heterodimeric motif generator and evaluator. The generator is a U-Net-based neural network that down-convolutes a monomeric motif pair and then up-convolute to generate a heterodimeric motif. A downstream machine learning model is used as the evaluator to compute for the predicted probability that a generated heterodimeric motif is the true one, based on the motif sequence features and DNA-binding family. Together, the generator and evaluator provide an integrated tool that enables users to conveniently synthesize heterodimeric motifs using any motif pair of interests.

Requirement

• python == 3.7
• pytorch == 1.7
• xgboost == 1.3.3
• MATLAB Engine API

Here is a tutorial showing how to install MATLAB Engine API

Usage

Here is a synthesis example notebook which can be run on colab:

CONTAINS:

• notebook/cross_validation_DeepMotifSyn.ipynb: Python notebook to evaluate DeepMotifSyn under leave-one-motif-pair-out cross-validation
• notebook/deepMotifSyn_motifSynthesis.ipynb: Python notebook to run DeepMotifSyn on synthesizing FLI1-FOXI1 motif
• data/*: Dataset for training and evaluating DeepMotifSyn containing 614 CAP-SELEX heterodimeric motifs from 313 monomeric motif pairs
• reproduced_paper_figure/plot_paper_figure.ipynb: Python notebook to plot the figures in our manuscript
• model/*: A well-trained DeepMotifSyn model containing a U-Net-based generator and a XGBoost-based evaluator

(Unzip the zip files in data/ before running notebook)