Multi-run Concrete Autoencoder to Identify Prognostic lncRNAs for 12 Cancers

Abstract

Long non-coding RNA plays a vital role in changing the expression profiles of various target genes that leads to cancer development. Thus, identifying prognostic lncRNAs related to different cancers might help in developing cancer therapy. To discover the critical lncRNAs that can identify the origin of different cancers, we proposed to use the state-of-the-art deep learning algorithm Concreate Autoencoder (CAE) in an unsupervised setting, which efficiently identifies a subset of the most informative features. However, CAE does not identify reproducible features in different runs due to its stochastic nature. We proposed a multi-run CAE (mrCAE) to identify a stable set of features to address this issue. The assumption is that a feature appearing in multiple runs carries more meaningful information about the data under consideration. The genome-wide lncRNA expression profiles of 12 different types of cancers, a total of 4,768 samples available in The Cancer Genome Atlas (TCGA), were analyzed to discover the key lncRNAs. The lncRNAs identified by multiple runs of CAE were added to the final list of key lncRNAs, which are capable of identifying 12 different cancers. Our results showed that mrCAE performs better in feature selection than single-run CAE, standard autoencoder (AE), and other state-of-the-art feature selection techniques. This study discovered a set of top-ranking 128 lncRNAs that could identify the origin of 12 different cancers with an accuracy of 95%. Survival analysis showed that 76 of 128 lncRNAs have the prognostic capability in differentiating high- and low-risk groups of patients in different cancers. The proposed mrCAE outperformed the standard autoencoder, which selects the latent features and is thought to be the upper limit in dimension reduction. Since the proposed mrCAE selects actual features and outperformed AE, it has the potential to provide information that can be used for precision medicine, such as identifying prognostic lncRNAs (this work) and mRNAs, miRNAs, and DNA methylated genes (future work) for different cancers.

Quantifying Intratumor Heterogeneity by Key Genes Selected using Concrete Autoencoder

“…Concrete Autoencoder (CAE) [15], a unsupervised deep learning approach, is used to identify cancer-specific key genes. CAE identifies features that are most informative for a given dataset [15], [17].…”

“…We kept three of the parameters the same as used in the original CAE developed by Abid et al [15]. These parameters are (i) leaky ReLU as activation function with a threshold value of 0.1, (ii) 10% dropout rate, and (iii) temperature T reduced using a simple annealing schedule from 10 to 0.1, For decoder part of CAE, we used a neural network with two hidden layers, each layer containing 300 units as used in our previous work with expression profile data [17].…”

Explainable Machine Learning to Identify Patient-specific Biomarkers for Lung Cancer

“…There has been much work on classifying lung cancers using gene expression values. The researchers also applied various well-known and novel machine learning and deep learning techniques for the feature selection and classification of lung cancers and other cancer types[14]–[20]. But they have mostly used machine learning and deep learning models as “black boxes.” Recently people have been using various approaches to explain the black box model.…”