In many machine learning applications datasets are collected from different sources and can be described with multiple representations or views. In this thesis, the methodology for the inference of microbial phenotypes from multi-view genomic and textual data is proposed. The proposed methodology consists of two main parts: (i) non-negative matrix factorization based framework for defining novel phenotypic concepts from multi-view textual data, and (ii) supervised machine learning framework for phenotype annotation where individual predictors from different textual and genomic representations are combined using the late integration approach. The proposed computational pipeline covers more than 3,000 microbial species across more than 400 phenotypic traits, and provides ~545,000 novel phenotypic annotations. Besides late integration of multiple views, this thesis introduces new multi-view subspace clustering algorithms that learn joint subspace representation matrix across all views, while at the same time promote sparsity and low-rankness of the representation. The rank and sparsity are measured using the nuclear and l1 norms, respectively. To estimate rank and sparsity more accurately, this thesis also introduces new low-rank sparse subspace clustering algorithms with l0 quasi-norm inspired regularizations of rank and sparsity. Proposed multi-view and single-view subspace clustering algorithms outperform state-of-the-art methods on datasets from various domains, including microbial textual and genomic datasets.