A Keras implementation of Wx in preprint, :
[Wx: a neural network-based feature selection algorithm for next-generation sequencing data Sungsoo Park, Bonggun Shin, Yoonjung Choi, Kilsoo Kang, and Keunsoo Kang] (https://www.biorxiv.org/content/biorxiv/early/2017/11/18/221911.full.pdf)

Differences with the paper:

• We use the Keras as Neural Network running framework, original paper used naive tensorflow framework
• Selected Features can be different as which backend learning framework used
• Some TCGA samples updated, so we have more samples than paper.

Contacts

• Your contributions to the repo are always welcome. Open an issue or contact me with E-mail [email protected]

Step 1.

Experienment Environments

$ python 3.4
$ tensorflow gpu 1.4.0
$ keras 2.1.2

Install instructions

$ pip install tensorflow-gpu
$ pip install keras

Step 2. Clone this repository to local.

$ git clone https://github.com/deargen/DearWXpub.git
$ cd DearWXpub

Step 3. Download the TCGA RNA-seq data

1. Download rna-seq via TCGA-Assembler tool

• we downloaded tool (Module_A.R) via http://www.compgenome.org/TCGA-Assembler/index.php

2. Run tcga_download.R

• you may have to install releative packages

 - Shall Terminal
 $ sudo apt-get install libssl-dev
   * deb: libssl-dev (Debian, Ubuntu, etc)
   * rpm: openssl-devel (Fedora, CentOS, RHEL)
   * csw: libssl_dev (Solaris)
   * brew: [email protected] (Mac OSX)
 
 $ sudo apt-get install libcurl4-openssl-dev
   * deb: libcurl4-openssl-dev (Debian, Ubuntu, etc)
   * rpm: libcurl-devel (Fedora, CentOS, RHEL)
   * csw: libcurl_dev (Solaris)

- R-studio
 > install.packages('curl')
 > install.packages('httr')
 > install.packages('stringr')
 > install.packages('rjson')

• you can see the 'TCGA_DATAS' folder in current DearWXpub path.

3. Data Status

Step 4. Do the Feature selection and Get Classification Accuracy

$ python src/wx_tcga.py

It will generate preprocessed TCGA data set. And Select features, Get the scores.

Keras Wx 14 ['EEF1A1','FN1','GAPDH','SFTPC','AHNAK','KLK3','UMOD','CTSB','COL1A1','GPX3','GNAS','ATP1A1','SFTPB','ACTB']

Peng 14 ['KIF4A','NUSAP1','HJURP','NEK2','FANCI','DTL','UHRF1','FEN1','IQGAP3','KIF20A','TRIM59','CENPL','C16ORF59','UBE2C']

with WX ranking ['KIF4A'(12970),'NUSAP1'(15886),'HJURP'(11479),'NEK2'(11939),'FANCI'(17123),'DTL'(12600),'UHRF1'(11825), 'FEN1'(17497),'IQGAP3'(14173),'KIF20A'(13057),'TRIM59'(11113),'CENPL'(10344),'C16ORF59'(9463),'UBE2C'(9039)]

edgeR 14 ['LCN1','UMOD','AQP2','PATE4','SLC12A1','OTOP2','ACTN3','KRT36','ATP2A1','PRH2','AGER','PYGM','PRR4','ESRRB']

with WX ranking ['LCN1'(3847),'UMOD'(7),'AQP2'(106),'PATE4'(5878),'SLC12A1'(165),'OTOP2'(7785),'ACTN3'(4939),'KRT36'(9441), 'ATP2A1'(14193),'PRH2'(2155),'AGER'(934),'PYGM'(13688),'PRR4'(2151),'ESRRB'(9630)]

TCGA data( Downloaded at Dec. 26th. 2017 ), Half of data for feature selection / Half of data for validation

Accuracy of 14 Biomarker

WX 14 AUC curve (LUSC, LUAD, BRCA)

Accuracy of 7 Biomarker

Keras wx 7

'EEF1A1', 'FN1', 'GAPDH', 'SFTPC', 'AHNAK', 'KLK3', 'UMOD'

Martinez 7

BLCA : 'SMAD2', 'RUNX2', 'ABTB1', 'ST5', 'CEBPB', 'SETDB1', 'CEBPG'

BRCA : 'JAK2', 'NFKBIA', 'TBP', 'RXRA', 'VAV1', 'HES5', 'NFKBIB'

HNSC : 'DUSP16', 'KRT8', 'RAF1', 'MED1', 'PPARG', 'YWHAB', 'FABP1'

KIRC : 'AR', 'HGS', 'RUNX1', 'BCL3', 'BRCA1', 'STAT2', 'ITGA8'

LUAD : 'DOK1', 'FUT4', 'INSR', 'ITGB2', 'SHC1', 'PTPRC', 'KHDRBS1'

LUSC : 'BRCA1', 'ETS2', 'HIF1A', 'JUN', 'LMO4', 'PIAS3', 'RBBP7'

$ python src/wx_examples.py

$ python src/wx_gse_multi_class.py

GSE72056

Cancer type : melanoma

Sample size
malignant : 1257
normal : 3256
total : 4513

Dataset splits(trn,dev,tst) = (3611, 902)
We perfomred the 5 fold cross validation

Accuracy Result (test set)
WX14 : 90.71%
Peng14 : 70.22%

GSE40419

Cancer type : luad(lung)

Sample size
malignant : 87
normal : 77
total : 164

Dataset splits(trn,dev,tst) = (132,32)
We perfomred the 5 fold cross validation

Accuracy Result(test set)
WX14 : 80.00%
Peng14 : 56.87%

GSE103322

Cancer type : Head and Neck(sinle cell)

Sample size
malignant : 2215
normal : 3687
total : 5902

Dataset splits(trn+dev,tst) = (4722,1180)
We perfomred the 5 fold cross validation

Accuracy Result(test set, 1180)
WX14 : 81.10%
Peng14 : 68.28%