A Method to facilitate cancer detection and type classification from gene expression using a deep auto-encoder and neural network
- 1. A Method to Facilitate Cancer Detection and Type Classification from Gene Expression Data using a Deep Autoencoder and Neural Network By Xi Chen March 27, 2019
- 2. Gene Expression Data Properties. • Gene expresses differently depending upon various factors such as the type of cells, environment and disease conditions. • Gene expression data are highly available due to the increased affordability of the sequencing technology. • Gene expression data are multimodality, high dimensional with small observation number (#row << #column). • Gene expression data can be used for disease detection and classification, and drug suggestion. 2
- 3. Gene Expression Data With Dimension Reduction • Using dimension reduction methods, such as PCA, for feature selection, since gene expression data have high dimension. • Apply traditional statistical and machine learning methods for application such as disease detection or classification. • Problem: how to explain the selected features. E.g. Each PC is a linear combination of the gene expression features. 3
- 4. Proposed Drug Suggestion Scheme. 2D Gene Expression Representation Feature 1 Feature2 Drug Sensitivity Drug A Drug B Drug C Drug D Cluster Approaches: • K-means • Gaussian Mixture Models 4
- 5. Problem: Current Gene Expression Data Don’t Include Drug Results. • Most gene expression data aren’t associated with well documented medical records. • Available records often miss drug information and patient disease outputs. 5
- 6. Solving The Harder Classification Problem First, Then We Could Infer Cluster Approach Works • In general, a classification problem is similar to a cluster problem, e.g. k-Nearest Neighbors algorithm. • If using gene expression data we could achieve high accurate classification results, we might be able to suggest clustering gene expression data for drug suggestion. 6
- 9. Autoencoder For Feature Learning Minimize 𝑓(𝐼𝑛𝑝𝑢𝑡 − 𝑂𝑢𝑡𝑝𝑢𝑡) 100 50 25 50 100 Training Autoencoder 1st hidden layer: 2nd hidden layer: 3rd hidden layer: 4th hidden layer: 5th hidden layer: Model Configuration 9
- 10. Learned Feature + Neural Network 10
- 11. Single Type Classification Lung cancer, abundant and balanced data 11
- 12. Why Not PCA? • PCA is a descriptive model. • Each component is a linear combination of all the features. • Hard to explain. 12
- 13. Cancer Type Acronym Full Name LGG Lower Grad Glioma UVM Uveal Melanoma LUSC Lung squamous cell carcinoma GBM Glioblastoma Multiforme Multiple Type Classification • Misclassifications are due to small sample size. • Misclassifications are sparse, clustering potential. 13
- 14. Conclusion • Autoencoder to automatically generate feature representations, thus addressing the very high dimensionality of gene expression data. • The extracted feature vector captures the non-linearity of the data. • This approach is scalable for new data after training, and it can generalize in multi-classification of different types of cancer. • We have demonstrated the high accuracy and low FNR/FPR of this method for the majority of the abundant cancer types, and its potential for handling sub-classification within certain cancers and identifying metastasis cancers. 14
- 15. Other Projects—Deep Learning Behind The Scenes • Almost all machine learning applications use similar approaches—Feature Engineering + Deep Learning. • E.g. Self-driving cars = CNN + DNN • Feature engineering CNN • Deep Learning training DNN • Deployment 15