Deep learning has performed well and led the third wave of artificial intelligence. Most of the current top-performing applications use deep learning, and the big hit AlphaGo uses deep learning. Modern machine learning methods have been extensively utilized in gene expression data processing. In particular, autoencoders (AE) have been employed in processing noisy and heterogenous RNA-Seq data. However, the input features AEs uses to generate their output are hidden or a "black box," making their outputs challenging to interpret, hindering downstream experimental validations and clinical translations. To bridge the gap between an AE and its biological interpretations, we developed a tool called eXplainable AutoEncoder for Expression data (XAE4Exp) which applies the SHapley Additive exPlanations (SHAP), the flagship technique in the field of eXplainable AI (XAI), to AEs. Through the use of SHAP, XAE4Exp quantitatively evaluates the contribution of each gene to the hidden features learned by an AE, substantially improving the applications of AEs. This tool will enable researchers and practitioners to analyze high-dimensional expression data intuitively, paving the way towards broader uses of deep learning.
-
Required Software
- Python 3.8.3 and above
- R 1.4.1 and above
-
Python Libraries
- numpy
- pandas
- shap
- sklearn
- torch
- torchvision
- matplotlib
-
R Libraries
- WebGestaltR
- data.table
- biomaRt
Check out requirements.txt for more details.
Please follow the following steps to access the data source:
- Go to https://portal.gdc.cancer.gov/
- Click on the Exploration tab
- Check the “TCGA-BRCA” and “ductal and lobular neoplasms” options on the left.
- Click “View files in Repository”
- Under “Workflow Type” in the left, select “HTSeq-Counts”.
- Click the “Add All Files to Cart” button on the right.
- Then, go to “Cart” on the upper right-hand side of the top bar.
- Download the sample sheet (this indicates the type of sample) and then click Download: Cart.
Similar process for other cancers as well. Once downloaded, the compressed files need to be extracted and put in one folder (the ‘.count’ files). Then convert the count data to the TPM expression matrix by data_collection1.R and data_collection2.R. Please note that this still contains both tumor and normal samples. In this project, we just used tumor data.
AutoEncoder
Autoencoder is an unsupervised neural network model that learns the implicit features of the input data, here we use Deep AutoEncoder for representation learning.
SHAP
SHAP (SHapley Additive exPlanations) is a game theoretic approach to explain the output of any machine learning model. It connects optimal credit allocation with local explanations using the classic Shapley values from game theory and their related extensions. By using SHAP, we can track contribution (weights) of each gene from cancer gene for the resluts of representation learning.
For more details, please reach out to:
https://github.com/slundberg/shap#citations
- To demonstrate the use of XAE4Exp, we applied it to The Cancer Genome Atlas (TCGA) breast cancer data (sample size N= 1041, number of genes M= 56497). I would like to thank TCGA for their data support!
-
(optional) To make the following feature importance figure more clear, we converted the data column of gene id to column of gene name by gene_annotation.R.
-
According to data quality control.py. Before runing AutoEncoder.py (This script already contains data quality control coding, so basically you can just ran AutoEncoder.py), we need to do gene data quality control to make sure the the variance of each gene is less than 1, which denoises the gene data.
-
According to AutoEncoder.py (This script already contains data quality control). After gene quality control, we proceed to representation learning. By tuning the parameters in AutoEncoder.py, we would have representations in different precision level. In this Deep AE, a three-hidden-layer (exclude input and output layer) structure is used, and loss function is designed by Mean Square Error (MSE).
-
Because we don't have a certain rule to tuning parameter, we always compare the results and find the best combination. Following is the parameter that would affect the results:
- batch_size (number of data cloumns that computer runs each time)
- test_size (train test split rate)
- smallest_layer (the number that you want to compress gene data, by changing smallest_layer, you need to change the compress rate inside the AutoEncoder Neural Network correspondly.)
- number_epochs (the number of round that you want data to be ran)
Besides, you might change the number of hidden layer, which is so-called the depth of AutoEncoder.
-
According to SHAP_combo.py. SHAP is a strong deep learning explainer, by SHAP, you are able to get figures of distribution of each gene contribution and exact number of each gene contribution. Based on the flow of SHAP, a pre-designed prediction model and an explainer are need. Here we used a randon forest with 100 decision trees with maximun depth of 20 as the prediction model, and TreeExplainer as the explainer.
-
The SHAP_combo.py results contains all feature importance in txt form (two columns with first column of gene name and second column of feature importance), top 20 global interpretation figure and top 20 feature importance figure. Separately, SHAP_value.py generates feature importance in txt file, SHAP_figure_scatter.py generates top 20 global interpretation figure, SHAP_figure_bar.py generates top 20 feature importance figure.
-
Basically, you can just run SHAP_combo.py, or you can run SHAP_value.py, SHAP_figure_scatter.py and SHAP_figure_bar.py one-by-one seperately.
-
SHAP figures:
- Example for Global interpretation figure (SHAP_figure_scatter.py):
- Example for Feature importance figure (SHAP_figure_bar.py):
- Example for Global interpretation figure (SHAP_figure_scatter.py):
-
According to ORA.R and GSEA.R. The enrichment analysis is preceeded by WebGestaltR package in R. ORA stands for Over-Representation Analysis and GSEA stands for Gene Set Enrichment Analysis. A .csv should be generated in this step.
-
(optional) WebGestalt also provides online Enrichment Analysis service with clear figures. They provide Over-Representation Analysis (ORA), Gene Set Enrichment Analysis (GSEA) and Network Topology-based Analysis (NTA). Here is the website: http://www.webgestalt.org/#
- Example data: please find the example data at data_example.csv. Please note that the example data is fake and small size, the training result could be bad.
- If you are confused about the process, feel free to check out the demonstration in the folder example.
This project is licensed under the MIT License
MIT License
Copyright (c) 2021 Yang Yu, Qingrun Zhang, Wenyuan Liao, Pathum Kossinna & Qing Li.
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
- Yang Yu, Qingrun Zhang, Wenyuan Liao, Pathum Kossinna & Qing Li.
- Qingrun Zhang
- Email: qingrun.zhang@ucalgary.ca
- Yang Yu
- Email: yang.yu2@ucalgary.ca