Promotech-CNN is a Bacterial Promoter Prediction Tool built using Dilated Convolution Neural Networks.

The body of the code was built using a Class module that read fasta files to retrieve DNA sequences. Once read, the sequences are converted to one-hot-coding and passed to the CNN for building and training. Our CNN architecture was guided by the BPNet architecture found at https://www.nature.com/articles/s41588-021-00782-6, however, this solution does not use as many layers and have several tweeks in architecture.

Environment:

-- Operating System: Linux
-- Python: 3.8.2
-- tensorflow: 2.5.0
-- numpy: 1.2.0
-- pandas: 1.2.4

About the Files:

-- Promotech-CNN-20210720.h5 - H5 module built from the trained network. This is loaded during predictions
-- PromotechCNN.py	- Base Class module that is used to build and train a netwok if required. It handles promoter sequence conversion before they are fed into the Neural Network. The main prediction method is processed here.
-- PromotechCNNPredict.py - Main file to load CNN module and obtain predictions for sequences in FASTA format

Making Predictions:

• Predictions are made after loading the H5 data file and then running the prediction on an input promoter sequence file.

    -- You may run predictions on fasta files or text files with only one 40nt sequence per line. Sequences must contain only A, C, G and T.
    	python PromotechCNNPredict.py Promotech-CNN-20210720.h5 sequences.fasta
  

• The output is a comma-delimited text file with four columns: sequence, probability of not being a promoter, probability of being a promoter and the prediction (0 - non promoter, 1 - promoter). The first row contains the column headers.

Seq,Neg_C,Pos_C,Pred
CAACCAGCAAGAGAACATCTGGCCTGCGGGCACCACTGCT,0.9882957,0.0117043005,0
GGCCTGCGGGCACCACTGCTACAGTGGTGGTCACTGGCAA,0.9797492,0.020250821,0
ACAGTGGTGGTCACTGGCAAACAAGATCGCGACGAAAGCC,0.9842748,0.015725216,0
CCGCGACCTTTGCCGCAAGCCTCGGTCTGAAAGTTAACGC,0.9870631,0.012936891,0
CTCGGTCTGAAAGTTAACGCCGGACACGGTCTGACCTATC,0.98490417,0.015095865,0

Training the Network from Scratch:

• The model was built for training on a Linux environment, you may change line 115 and 116 to run on Windows - the '/' should be switched to '\'.

• If you intend to rebuild the network from scratch, you may run the commands below - this is ideal for new training data.

    -- Import the Python Class.
    	from PromotechCNN import PromotechCNN
    	code = PromotechCNN()
    
    -- Setup the training data(Path to directory with training data is used).
    	code.createTrainFastaFiles("/TrainingData/40nt-sequences/bacteria-1-10-ratio")
    	code.buildData()
  
    -- Build and Train the network with 100 epochs, early stopping is in place so the network might stop before 100 epochs.
    	code.buildBPNet(100)
  
    -- Predict the validation data. Used to build AUPRC and AUROC graphs
    	code.predictValidation("/ValidationData/40nt-sequences/bacteria-1-1-ratio")
  
    -- Run prediction on some promoter sequence.
    	code.predictSequenceFile("BACILLUS_negative.fasta")