novasmedley / deepradiogenomics

Methods for training and interpreting deep radiogenomic neural networks

Radiogenomic neural networks

deepRadiogenomics contains the source code to analyses in the paper:

Nova F Smedley, Suzie El-Saden, William Hsu, Discovering and interpreting transcriptomic drivers of imaging traits using neural networks, Bioinformatics, btaa126, https://doi.org/10.1093/bioinformatics/btaa126

Note: an updated version has been made and will be added in the future.

• repo contents
• data
• install
• usage + demos
• citation

Repo contents

• data full gbm dataset used in analysis - to be released
• demo_data toy data
• R: scripts for most post-modeling analyses, association testing, etc.

• general modeling functions:

  • neuralnet.py
  • other_models_utils.py
  • bootstrap.py
  • custom_callbacks.py
  • sparse_x_generator.py

• glioblastoma (gbm) specific functions:

  • training:
    • setup.py
    • train_gene_ae.py deep transcriptomic autoencoder
    • train_nn.py supervised radiogenomic neural network
    • train_others.py comparative models (logit, gbt, rf, svm)
  • extract radiogenomic associations
    • gene_masking.py
    • get_masking.py
    • gene_saliency.py
    • get_saliency.py
  • misc.
    • parse_cv.py
    • demo_gene_masking.py
    • demo_gene_saliency.py
  • all others, see R

Data

All data was originally taken from public repositories, where identifiable information was scrubbed.

• Transcriptomic data was downloaded from the legacy version of The Cancer Genome Archive (TCGA).

• Imaging studies were download from from The Cancer Imaging Archive (TCIA). Vasari traits were annotated by Dr. Suzie El-Saden and based on pre-operative magnetic resonance imaging studies.

  • Vasari MR Feature Guide_v1.1.pdf Vasari guidelines for imaging annotations
  • Our annotation form was based on the Round 2 Google Form used by Vasari Project

Datasets are available in the data folder.

• training data
  • gene_expression.txt - gene expression profiles
  • vasari_annotations.csv - imaging traits
  • nationwidechildrens.org_clinical_patient_gbm.txt - TCGA-GBM clinical traits
• gene sets
  • TCGA_unified_CORE_ClaNC840.txt - gene sets from Verhaak, Roel GW, et al. "Integrated genomic analysis identifies clinically relevant subtypes of glioblastoma characterized by abnormalities in PDGFRA, IDH1, EGFR, and NF1." Cancer cell (2010)
  • gene_sets_Puchalski - gene sets from Puchalski, Ralph B., et al. "An anatomic transcriptional atlas of human glioblastoma." Science (2018)
  • msigdb_v6.2_GMTs - gene set collections from Molecular Signatures Database.

For more details, see our paper.

Install

• Neural networks were trained on Amazon Web Services using Deep Learning AMI with Ubuntu 16.04.4 LTS and the tensorflow_p36 environment. All other classifiers were implemented on an Ubuntu 18.04.1 LST machine.

  • Check out AWS's environment documentation.

  • Python 3.6 dependencies (training of comparative models, gene masking, gene saliency):

    keras 2.2.2
    keras-vis 0.4.1
    numpy 1.14.3
    pandas 0.23.0
    scikit-learn 0.20.0
    scipy 1.1.0
    seaborn 0.8.1
    tensorflow 1.10.0
    xgboost 0.80
    

  • R 3.4.4 dependencies (gene set enrichment analysis, but mostly figure generation):

    awtools 0.1.0
    broom 0.5.1
    cowplot 0.9.3
    data.table 1.11.8
    doParallel 1.0.14
    dplyr 0.7.6
    egg 0.4.0
    fgsea 1.4.1
    foreach 1.4.4
    ggrepel 0.8.0
    ggplot2 3.1.0
    grid 3.4.4
    gridExtra 2.3
    ggpubr 0.2
    pheatmap 1.0.12
    plyr 1.8.4
    qvalue 2.10.1
    rcartocolor 1.0.0
    RColorBrewer 1.1
    reshape2 1.4.3
    scales 1.0.0
    tidyr 0.8.1
    tidyverse 1.2.1
    viridis 0.5.1
    wesanderson 0.3.6
    

• install from Github using git:

  git clone https://github.com/novasmedley/deepRadiogenomics.git
  

Usage

Demos were run using demo data, a small subset of the published dataset, on Ubuntu 18.04.1 LTS with 15.5 GB memory. It has also been tested on macOS 10.14.5.

Neural network pipeline:

1. Train gene expression autoencoder (ae) - cross-validation(cv), 15 secs:

   $ python3 train_gene_ae.py --exp ae_cv --dir demo --data demo_data \
   --label autoencoder --predType regression --loss mae --opt Nadam --act tanh \
   --h1 200 --h2 100 --h3 50 --epoch 2 --folds 2 --patience 2
   

2. (optional) Parse cv results: $ python3 parse_cv.py --dir demo/ae_cv --model nn

3. Retrain ae - 15 secs:

   run train_gene_ae.py from Step 1 except: change --exp ae_retrain and add --retrain 1

4. Train radiogenomic model - cv, 21 secs:

   $ python3 train_nn.py --exp nn_cv --dir demo --data demo_data \
   --pretrain demo_results/ae_retrain/autoencoder/neuralnets/200_100_50_0_0_tanh_decay_0_drop_0_opt_Nadam_loss_mae_bat_10_eph_2 \
   --label f5 --opt Nadam --act tanh \
   --h1 200 --h2 100 --h3 50 --epoch 2 --folds 2 --patience 2 --freeze 0 --num_ae_layers 3
   

5. (optional) Parse cv results: $ python3 parse_cv.py --dir demo/nn_cv --model nn

6. Retrain radiogenomic model - 16 secs:

   run train_nn.py from Step 4 except: change --exp nn_retrain and add --retrain 1

7. Gene masking - 11 secs

   $ python3 demo_gene_masking.py --label f5 --geneset verhaak --cpus 7

8. Gene saliency - 19 secs

   $ python3 demo_gene_saliency.py

Train other models:

1. Fit logit with l1 regularization - 2 secs, fitting 1000 hyperparameters

   $ python3 train_others.py --exp other_cv --dir demo --data demo_data \
   --dataType vasari --predType binaryClass --label f5 --model logit1 --folds 2 --cpus 7
   

2. Parse cv results:

   $ python3 parse_cv.py --dir demo/other_cv --model other

Citation

If you want to cite this work, please cite the paper:

TBA!

and the repo:

@misc{smedleyRadiogenomics,
  title={deepRadiogenomics},
  author={Smedley, Nova F},
  year={2019},
  publisher={GitHub},
  howpublished={\url{https://github.com/novasmedley/deepRadiogenomics}},
}