This repo is a submission for the 8200Bio - BiomX data challenge. The challenge was binary classification of DNA sequences (bacteria/pahge). My solution was a CNN implemented using TensorFlow.
my_utils.py contains all helper functions for dataset generation, preporcessing and visualisations.
encodeFASTA.py is a fork of fasta_one_hot_encoder with my addition of zero padding to support DNA sequences of varying lengths.
Given a balanced training set of 10k labeled sequences, classify DNA sequences of varying to either bacteria of phage (a virus that attackes bacteria). Test set consists of 2k previously unseen sequences.
Here's a sample entry from the train set:
>Phage-4995
ATGACGGCTGATCAGGTGTTTAACCAAGTGCTGCCTGAAGCTTACAAGCTT...
The main challenge was the varying length of sequences:
In order to overcome the difference in sequence length, I slice each sequence to a fixed length and duplicate the label accordingly. During inference, the class is determined using the mean probability score for each slice.
My model is a CNN inspired by the following article Enhancer Identification using Transfer and Adversarial Deep Learning of DNA Sequences:
Why Use 1D Convolutions?
The information we seek to learn from DNA sequences comes mostly in the form of specific protein sequences (k-mers, motifs, etc..). This means that a 1D covolution is more suited for this model than the "classic" 2D convolution which learns spatial features.
The trained network reached a validation accuracy of 96%.
It is important to note that validation set metrics seems to be much better than the train set metrics. This can ber explained by the relatively high dropout rate and by the small size of the validation set.
Results on the test
I scored 76.8% accuracy, which is not bad but the winner got an 88.8%. Given that the winner was a Bionformatics PHD student, I can be pretty proud of the results :)
Feel free to use this work as long as you refrence this repo.
Contact: doronser@gmail.com