DeepM6ASeq

DeepM6ASeq is a deep-learning-based framework to predict m6A-containing sequences and visualize saliency map for sequences.

Dependency

• Python3

• numpy==1.14.2

• torch==0.4.0a0+0e24630 ( pytorch )

• scikit_learn==0.19.1

Content

• data: data of human, mouse and zebrafish for model training ( uncompressed by "tar -zxvf data.tar.gz")
• demo_test:
  • test.fa: data for testing scripts
  • expected results
    • predout.tsv: prediciton results
    • sal.out: saliency map output
    • sal_heatmap: visualization for saliency map
• result: uncompressed by "tar -zxvf result.tar.gz"
  • trained_models: trained models for mammalian, human,mouse and zebrafish
  • tomtom: TOMTOM results for aligning learned motifs to known motifs
  • RSAT: clusters of learned motifs

Usage

1. train model

The script main_train.py is used to train model. The required arguments

• model_type: cnn or cnn-rnn
• pos_fa/neg_fa: the fasta file for positive samples/negative samples ( the length of sequences should be no more than 101bp)
• out_dir: the path of output directory

This script ouput the trained model and prediction result in the out_dir.

python main_train.py -m model_type -pos_fa pos_fa -neg_fa neg_fa -od out_dir

2. predict m6A-containing sequences

The script main_test.py is used to predict if a given sequence contain m6A sites. The required arguments

• model_type: cnn or cnn-rnn
• input_fa: a fasta file for test samples ( the length of sequences should be no more than 101bp)
• model_dir: the path of model directory
• out_fn: output file

This script ouput the prediction scores for given sequences.

python main_test.py -m model_type -infa input_fa -md model_dir -outfn out_fn

3. saliency map for m6A-containing sequences

The script saliency_map.py is used to get saliency map for given sequences. The required arguments

• model_type: cnn or cnn-rnn
• input_fa: a fasta file for test samples ( the length of sequences should be no more than 101bp)
• model_dir: the path of model directory
• out_fn: output file

This script ouput the saliency map file for given sequences. The output could be visualized using saliency_heatmap.R. The visualization result is a pdf file which contains saliency map for each sequence.

python saliency_map.py -m model_type -infa input_fa -md model_dir -outfn out_fn

Rscript saliency_heatmap.R saliency_map_out pdf_name

Other

1. Confidence threshold

The table lists confidence thresholds of prediction scores for mammlian, human, mouse and zebrafish. Moderate, High and Very high correspond to 90%, 95% and 99% specificity respectively.

| | Moderate | High | Very high | | --------- | -------- | ----- | --------- | | Mammalian | 0.725 | 0.818 | 0.929 | | Human | 0.772 | 0.841 | 0.92 | | Mouse | 0.724 | 0.813 | 0.89 | | Zebrafish | 0.715 | 0.82 | 0.9 |