GeneSeq2Taxonomy

Taxonomy Classifer Using Gene Sequence.

Using FFT and Neural Net to identify gene sequence's taxonomy.

Requires

• Python3
• numpy
• oct2py
• tensorflow
• Octave

How to use

Collect lots of sequence data

|-GeneSeq2Taxonomy
  |-archaea
    |- *.fna
  |-bacteria
    |- *.fna
  |-protozoa
    |- *.fna
  |-calc_fft.m
  |-extract_feature.py
  |-train.py

Extract features

Extract features executing extract_features.py for each directory.

>python3 extract_feature.py archaea
Candidatus_Altiarchaeales_archaeon_WOR_SM1_SCG.fna
Candidatus_Altiarchaeales_archaeon_WOR_SM1_SCG.fna_1
Candidatus_Altiarchaeales_archaeon_WOR_SM1_SCG.fna_2
Candidatus_Altiarchaeales_archaeon_WOR_SM1_SCG.fna_3
...
>python3 extract_feature.py bacteria
Arthrobacter_globiformis.fna
Arthrobacter_globiformis.fna_1
Arthrobacter_globiformis.fna_2
Arthrobacter_globiformis.fna_3
...
>python3 extract_feature.py bacteria
protozoa.1.1.genomic.fna
protozoa.1.1.genomic.fna_1
protozoa.1.1.genomic.fna_2
...

Then .npy files are generated at each directory.

Train data

Open Python shell, execute train method at train.py

python3
Python 3.6.0 |Anaconda 4.3.0 (64-bit)| (default, Dec 23 2016, 12:22:00) 
[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import train from train
>>> clf= train(steps=50000)

2/3 of data are used for train and 1/2 data are used for calculate precision. Predictions are printed after training. Tensorflow checkpoint is saved at GeneSeq2Taxonomy/dnn_model.

How it works

1. Translate gene sequence to z-curve sequence. It's for reduce training cost by reducing features. I think there's no meaningful difference comparing other methods. http://www.mecs-press.org/ijitcs/ijitcs-v4-n8/IJITCS-V4-N8-3.pdf

2. Apply Fourier transform to extract it's spectral features. Manipulate Fourier transformed data to train neural net. I split frequence axis by 1000 area and picked up max value for each area. I've got over 90% precision using this method classifying archaea, bacteria, protozoa and fungi.(Best method I've tried. See below to other methods I've tried) For example if Fourier transformed data is [1, 2, 3, 6, 8, 7, 5, 3, 1, 4, 6, 4]. I split frequence axis to 4 and pick max value, then result will be [3, 8, 5, 6].

3. Use neural net, train data.

Another methods manipulating FFT result.

• Pick frequency order by magnitude. Suppose magnitudes [1, 2, 3, 4, 8, 7, 5, 3, 6, 4] and frequence of each magnitudes [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0]. Pick 3 frequences order by magnitude. [0.5, 0.6, 0.9].

Result: Failed cause there's lots of peak around 2/3 pi area. It means there's many 3-periodic sequence; Codon. Data looks like this. [0.333333, 0.333321, 0.3333342, 0.333432 ....]

• Pick frequency order by magnitude at band pass filtered data. I wiped up data around 2/3 pi and 0 pi. And pick frequences order by it's magnitude.

Result: Precision about 80% classifing Archaea and Bacteria. But classifying Archaea, Bacteria and Protozoa, precision falls around 65%.

• Pick magnitude at random frequency and pick best classifier.

Result: Failed. Frequency is near real number. Too many learning case. Precisions are about 70% classifing Archaea and Bacteria.