Rapid advances in high-throughput measurement technologies have led to a data rich era. Such big data provide opportunities to discover more biological phenomena but also challenge our ability to efficiently and effectively analyze them. While supervised machine learning algorithms are powerful in studying a known phenomenon, they cannot make novel discoveries. Therefore, unsupervised approaches that can directly extract knowledge from big data without requiring metadata annotation are needed. We adapted a neural network-based approach, denoising autoencoders (DAs), to fulfill this need. We first applied DAs to breast cancer gene expression data. DAs built features descriptive of important molecular characteristics of breast cancer, and the features can be robustly generalized to an independent dataset. We next developed ADAGE, A̲nalysis using D̲enoising A̲utoencoders for G̲ene E̲xpression, a DAs-based unsupervised feature extraction approach that integrates a complete gene expression compendium of an organism. We demonstrated ADAGE in the bacterial pathogen Pseudomonas aeruginosa. ADAGE constructed features that reflect biological states, such as strain variations and oxygen abundance in the environment, and achieved better resolution than the traditional feature extraction approaches PCA and ICA. Aiming for a more robust model, we next developed eADAGE, an ensemble of ADAGE models. eADAGE consolidates 100 individual ADAGE models into one model with improved robustness, less noise, and better concordance with known biology. We found that features in eADAGE models can recapitulate expert-curated biological pathways, and they obtained higher pathway coverage and better pathway resolution than ADAGE, PCA, and ICA. Using eADAGE, we compared 78 different medium types used in the P.a. compendium and identified media-specific patterns. Finally, to put eADAGE features into researchers’ data analysis workflow, we developed the ADAGE signature analysis pipeline. This pipeline aims to identify perturbed biological processes in an input dataset. Unlike traditional pathway analysis that relies on expert-curated processes, we leverage features extracted from big data that mimic biological processes. This pipeline is especially useful for less wellannotated organisms with few pathway resources available. As data continue to grow, we expect that unsupervised knowledge extraction methods like ADAGE will become the mainstream in big data analysis.