NSF EAGER: Topic Models for Population Genetics

Columbia University, New York NY

Investigators

Abstract

The project breaks new ground by revealing the compelling analogy between analysis of natural language and genetics. In text analysis, documents are modeled as discussing different topics, each with its characteristic vocabulary. Similarly, modern day individuals can be thought of as having ancestry in multiple populations, each with its characteristic genetic patterns. Applied to state-of-the-art genomic data from contemporary individuals and archaeological remains, the unified framework proposed by this project is expected to resolve great historical mysteries, such as the decline of the Mayans, the spread of agriculture, or the evolution of the Indian caste system. The project is expected to adapt Topic Modeling techniques, a framework from Natural Language Processing which employs Latent Dirichlet Allocation to population genetics. The project will pursue three goals: 1. Formulate existing analysis methods in population genetics as Topic Models, leveraging the existing framework in other domains to improve efficiency and accuracy of genomic analysis 2. Introduce the domain-specific concepts of time and space, across which populations evolve in a theoretically understood way. 3. Integrate the components of the model to create a graphical model of ancestral populations, which describes the genetic history of contemporary and historical populations whose genomes had been sequenced. The project will compare accuracy and efficiency of models vs. the existing standards in the field. All software tools that will be developed as part of the project will be made available to the research community.