Exploiting Latent Features of Text and Graphs

May 7, 2020

Dissertation

  • Justin Sybrandt

The fastest growing online data sources include unstructured text and large graphs. These data are particularly prevalent in biomedical science, as large databases of scientific papers and archives of experimentally derived relationships entomb a vast amount of knowledge. While collections of text and graphs may be useful for human readers to understand, their ever-expanding size and scope creates a scalability challenge. To aid human researchers in scientific explorations, we propose new techniques that enable machine understanding of scientific text and graphs. To do so, we use embeddings—semantically rich vectors of latent features—to convert human constructs into mathematical representations.

In this dissertation we propose a range of new techniques and systems for automatic hypothesis generation, the process of using existing scientific information to anticipate fruitful new research directions. This exploration includes the Moliere system, which uses text embeddings to construct a large semantic graph containing associations between biomedical concepts. Through graph and text analytics the Moliere system uncovers nontrivial implicit connections within biomedical science. We augment this system to produce a novel ranking criteria, which enables large-scale validation and massive real-world experiments. We apply the Moliere system to the problem of HIV-associated neurodegenerative disorder in order to automatically identify a new gene-treatment target in DDX3. We confirm this mathematically derive hypothesis in a laboratory setting.

Additionally, we propose a new graph-embedding technique suited for bipartite graphs, such as the gene-disease or the drug-side effect graphs, and demonstrate that it can better capture type-specific latent features. We further demonstrate this new embeddings effectiveness by proposing an embedding-based solution strategy to the NP-Hard algorithmic problem of hypergraph partitioning. By using the global structural features of graph embeddings we substantially improve partitioning result quality, which in real-world applications corresponds to a significant decrease in communication overhead in large scientific workloads.

Using recent advances in deep learning and graph embedding, we construct a next-generation hypothesis generation system, Agatha, which challenges a range of assumptions present in Moliere. Using our proposed validation technique, we demonstrate a substantial quality improvement in ranking plausible future research directions. Furthermore, the Agatha system queries potential hypotheses two orders of magnitude faster than its counterpart, enabling whole new discovery paradigms.

To supplement our quantitative ranking criteria with more interpretable information, we propose a text generation model that can author biomedical summaries. We propose the conditional biomedical abstract generator, a deep-learning language model that allows us to modify the content of automatically generated text based on a set of desired elements of interest. As a result, we gain the ability to generate human-understandable content that could help describe automatically derived hypotheses.