Eukaryotic-Promoter-Recognition-using-CNN

DeePromoter combines a convolutional neural network (CNN) and a long short-term memory (LSTM). Additionally, instead of using non-promoter regions of the genome as a negative set, we derive a more challenging negative set from the promoter sequences. The proposed negative set reconstruction method improves the discrimination ability and significantly reduces the number of false positive predictions.

Requirements

• Please install torch==1.9 from https://pytorch.org

• You can install others Python dependencies with

  pip3 install -r requirements.txt

Dataset

Current supported dataset is:

• EPDnew : A collection of experimentally validated promoters for selected model organisms. Evidence comes from TSS-mapping from high-throughput expreriments such as CAGE and Oligocapping

Preprocessing

Dataset for Human and Mouse had been processed and stored in ./data

Procedure for create negative dataset as described in paper:

• Step 1: Break the protein sequence to N part(20 as in the paper)

• Step 2: Random choose M part of the original protein to keep it, and random initialize the rest

• Step 3: For every training step mix the positive batch with negative batch and perform training

Training

python3 train.py -d data/human/nonTATA/hs_pos_nonTATA.txt --experiment_name human_nonTATA

Early stop had been implement and train will automatically stop when Mathews correlation coefficient is saturated

The results will be saved in to ./output/experiment_name

You can do continue training by pass the path to weight by flag -w or --weight

Inference

Prepare your dataset in txt format with each DNA sequence(length 300) on a line

Run inference by

python3 test.py -d data/human/nonTATA/hs_pos_nonTATA.txt -w path_to_weight

Output will be save into file infer_results.txt in the main folder

Implementation Issues

Negative sampling

1. In addition to using negative sampling as in the paper described(see Preprocessing) I added a random dataset to help the model generalize.

Parallel convolution

1. The author use grid search to find optimal parameters for the network. I used the final set of parameter from the paper. Kernel size = [27, 14, 7], and maxpooling with kernel = 6

References

1. If Scanpy is useful for your research, consider citing DeePromoter paper