Overall survival prediction of non-small cell lung cancer by integrating microarray and clinical data with deep learning
This repository provides a reference implementation of Bimodal DNN as described in the paper submitted to Scientific Reports: Overall survival prediction of non-small cell lung cancer by integrating microarray and clinical data with deep learning.
The built gene interaction networks based on microarray data. Based on our previous research, we selected 7 prognostic biomarkers with high prognsotic relevance values (PRV) and combined with 8 eight well-known biomarkers to form a set of 15 prognostic biomarkers for NSCLC patient overall survival prediction. Their interactions are illustrated with the following figure generated with STRING:
The Bimodal DNN consists of two subnetworks for microarray and clincial data, as well as a merge subnetwork. The two subnetworks were first pre-trained with microarray and clinical data, respectively. They were then merged with the merge ubnetwork to obtain the final prediction. The detailed model architecture is illustrated in the following figure and more details can be found in the submitted paper.
The combined GSE cohort is available here. The data were first pre-processed and split into train/test splits.
You can train the microarray/clinical subnetworks with the following lines. Trained model and its details are saved in /models.
python3 pretrain_x.py
python3 pretrain_c.pyAfter pre-training, the merged Bimodal DNN can be trained with
python3 merge.py
Loss per epoch for the microarray subnetwork.
Loss per epoch for the clinical subnetwork.
Loss per epoch for the merge network.
If you find our work useful for your research, please consider citing our work. Please feel free to contact us. A pre-print version of our paper is available on bioRxiv. The full version accepted to Scientific Reports is also available.