Comparative genomics is essentially a form of data mining in large collections of n-ary relations between genomic elements. Increases in the number of sequenced genomes create a stress on comparative genomics that grows, at worse geometrically, for every increase in sequence data. Even modestly-sized labs now routinely obtain several genomes at a time, and like large consortiums expect to be able to perform all-against-all analyses as part of these new multi-genome strategies. In order to address the needs at all levels it is necessary to rethink the algorithmic frameworks and data storage technologies used for comparative genomics.To meet these challenges of scale, in this thesis we develop novel methods based on NoSQL and MapReduce technologies. Using a characterization of the kinds of data used in comparative genomics, and a study of usage patterns for their analysis, we define a practical formalism for genomic Big Data, implement it using the Cassandra NoSQL platform, and evaluate its performance. Furthermore, using two quite different global analyses in comparative genomics, we define two strategies for adapting these applications to the MapReduce paradigm and derive new algorithms. For the first, identifying gene fusion and fission events in phylogenies, we reformulate the problem as a bounded parallel traversal that avoids high-latency graph-based algorithms. For the second, consensus clustering to identify protein families, we define an iterative sampling procedure that quickly converges to the desired global result. For both of these new algorithms, we implement each in the Hadoop MapReduce platform, and evaluate their performance. The performance is competitive and scales much better than existing solutions, but requires particular (and future) effort in devising specific algorithms.
