jiahuaqu/aire_target

Aire's target choices

This repository contains code related to the paper: Aire relies on Z-DNA to flag gene targets for thymic T-cell tolerization.

Code for the CNN model and various types of sequencing experiments can be found in the corresponding folders. Instructions for scripts of the CNN model can be found below. Other scripts were written in R or Bash using standard packages.

CNN instructions

To investigate Aire’s target choices, we built and trained a CNN model with dilated convolutional layers and residual skip connections to distinguish the extended-promoter sequences of Aire-induced and Aire-neutral genes. We followed the pre-training and then fine-tuning paradigm, given the relatively small number of Aire-induced and Aire-neutral genes compared with the typical training sizes employed for large-scale models. Below detailed the required coding environment and script usage.

Pre-training

Environment setup

Code for building and pre-training the CNN was developed using Python 3.6 and a variety of well-developed packages. To prepare for the pre-training environment using Anaconda, run the following:

conda env create -f pretrain.yml

Scripts

All the utility functions are included in the script model_main.py.

pretrain_data.py is used to modify the datasets of Basenji2¹ for the pre-training purpose. Here is an example of the usage:

python pretrain_data.py ${data_path}

## Mandatory input:
  ${data_path}: The directory containing the original datasets of Basenji2.

run_pretrain.py is used to pre-train the CNN model. Example usage:

python run_pretrain.py ${tfr_path} ${loss} ${weight}

## Mandatory input:
  ${tfr_path}: The directory containing the modified datasets of Basenji2 for pre-training.
  ${loss}: The id of the loss used for pre-training: 
    ${loss} == 1: tf.keras.losses.Poisson()
    ${loss} == 2: MSEMultinomialPretrain(total_weight = ${weight})
    ${loss} == 3: PoissonMultinomialPretrain(total_weight = ${weight})
  ${weight}: An argument for the loss function.

Fine-tuning

Scripts

All the utility functions are included in the scripts model_main.py and utility_functions.py.

finetune_data.py included the functions used to generate the TFRecords for fine-tuning.

run_finetune.py is used to fine-tune the CNN model. Example usage:

python run_finetune.py ${batch_size} ${tfr_train_path} ${tfr_test_path} \
${tfr_valid_path} ${pretrained_path} ${pretrained_prefix} ${train_sequence_num} \
${valid_sequence_num} ${test_sequence_num} ${finetuned_path} ${finetuned_prefix}

## Mandatory input:
  ${batch_size}: The batch size for fine-tuning.
  ${tfr_train_path}: The directory containing the training set for fine-tuning.
  ${tfr_test_path}: The directory containing the test set for fine-tuning.
  ${tfr_valid_path}: The directory containing the validation set for fine-tuning.
  ${pretrained_path}: The directory containing the saved pre-trained model.
  ${pretrained_prefix}: The name of the saved pre-trained model.
  ${train_sequence_num}: The number of training samples.
  ${valid_sequence_num}: The number of validation samples.
  ${test_sequence_num}: The number of test samples.
  ${finetuned_path}: The directory to save the fine-tuned model.
  ${finetuned_prefix}: The name of the fine-tuned model.

Footnotes

1. Kelley, D.R. Cross-species regulatory sequence activity prediction. PLoS Comput. Biol 16, e1008050 (2020). ↩