Multimodal autoencoders for subtypes and survival prediction of breast cancer

Implementation of our paper titled "Prognostically Relevant Subtypes and Survival Prediction for Breast Cancer Based on Multimodal Genomics Data" submitted to IEEE Access journal, August 2019. In this implementation, a multimodal autoencoders(MAE) is used to predict different clinical status of breast cancer patients based on multiplatformic genomics data. The MAE is trained with genomics data such as DNA methylation, gene expression, miRNA expressionfrom, and clinical outcomes from The Cancer Genome Atlas(TCGA).

Predicted clinical status

1. Breast cancer subtypes which is determined by the estrogen receptor (ER), progesterone receptor (PGR), and HER2/neu status
2. Survival rate (0-1, with 1 being the best chance of survival).

Requirements

• Python 3
• TensorFlow
• Keras.

About the the dataset used

DATASET_IDX	Data Types	Data size(GB)
1	DNA Methylation	148
2	Gene Expression	9
3	miRNA Expression	0.24
4	Gene Expression + miRNA Expression	10
5	DNA Methylation + Gene Expression + miRNA Expression	162

Training the neural networks

• python3 main_run.py <options>, with the below supported options:

Option	Values	Details	Required
-p PLATFORM --platform PLATFORM	int [1-2]	[1] Tensorflow, [2] Theano	yes
-t TYPE --type TYPE	int [1-2]	[1] Breast cancer type classification [2] Survival rate regression	yes
-d DATASET --dataset DATASET	int [1-15]	[1] DNA Methylation GPL8490 [2] DNA Methylation GPL16304 [3] Gene Expression Count [4] Gene Expression FPKM [5] Gene Expression FPKM-UQ [6] miRNA Expression [7] Gene Expression Count + miRNA Expression [8] Gene Expression FPKM + miRNA Expression [9] Gene Expression FPKM-UQ + miRNA Expression [10] DNA Met GPL8490 + Gene Count + miRNA [11] DNA Met GPL16304 + Gene Count + miRNA [12] DNA Met GPL8490 + Gene FPKM + miRNA [13] DNA Met GPL16304 + Gene FPKM + miRNA [14] DNA Met GPL8490 + Gene FPKM-UQ + miRNA [15] DNA Met GPL16304 + Gene FPKM-UQ + miRNA	yes
--pretrain_epoch PRE_EPOCH	int	Pre-training epoch. Default = 100	no
--train_epoch TRAIN_EPOCH	int	Training epoch. Default = 100	no
--batch BATCH	int	Batch size for pre-training and training. Default = 10	no
--pre_lr PRE_LR	int	Pre-training learning rate. Default = 0.01	no
--train_lr TRAIN_LR	int	Training learning rate. Default = 0.1	no
--dropout DROPOUT	int	Dropout rate. Default = 0.2	no
--pca PCA	int [1-2]	[1] Use PCA [2] Don't use PCA Default = [2] Don't use	no
--optimizer OPTIMIZER	int [1-3]	[1] Stochastic gradient descent [2] RMSProp [3] Adam Default = [1] Stochastic gradient descent	no

Example

If we want to perform breast cancer subtype classification based on the dime sion reduced DNA methylation dataset using PCA on TensorFlow platform, one can issue the following command from the terminal: python3 main_run.py --platform 1 --type 1 --dataset 1 --batch 10 --pretrain_epoch 5 --train_epoch 5 --pca 1 --optimizer 3

In the preceding command, we define: -- 10 as the batch size -- 5 as the number of pretraining epoch -- 5 is the fine tuning epoch -- 3 is the idx for the Adam optimizer.