Pathway Hypotheses Knowledge-base: A Knowledge Source for the Biomedical Data Translator

Dartmouth College, Hanover NH

Investigators

Abstract

We propose to develop a Pathway Hypotheses Knowledgebase (PHK), a new knowledge source that will analyze and hypothesize novel relationships and interactions driven by researcher data together with the wealth of knowledge as captured by the Biomedical Data Translator project. The capability to bring together different biomedical data knowledge sources, including experimental data, in order to discover yet unknown relationships has been difficult to realize. The fundamental gap lies in a lack of rigorous and robust framework for linking or creating sophisticated lattices of relationships needed for different lines of evidence from heterogeneous knowledge sources. Such a framework must address systematic and mathematically driven algorithms for hypotheses exploration, construction, and assessment in order to bridge gaps, derive, and ultimately discover new knowledge beyond existing sources by assessing existing molecular relationships and generating additional actionable pathway information that can be queried directly or via API. A critical feature of a successful framework must include formal, welldefined mechanisms for sensitivity analyses (measures of fragility and reliability of constructed hypothesis), impact analyses (measures of importance and knowledge novelty), and parsimony analyses (wholistic measures of congruence of hypotheses). Lastly, the functional mechanism behind newly derived knowledge must be invertible and provide an unambiguous, precise, auditable provenance from the original knowledge sources serving as the basis for explainability. To realize this vision, PHK will employ a mature AI knowledge representation called Bayesian Knowledge Bases (BKBs). BKBs model knowledge, relationships, and uncertainty within a rigorous graph-based probabilistic framework capable of managing inconsistent, incomplete, and cyclic knowledge. It fully subsumes a variety of well-known models including (dynamic) Bayesian networks and temporal representations such as hidden Markov models. BKBs can be learned from data but its most critical contribution is the ability to fuse multiple BKBs and their underlying distributions without any loss of information. This inherently provides end-to-end forward to backward auditability of computational derivations which admits ready sensitivity, contribution, and impact analysis. This further leads to a formal mechanism to explore, discover, and create new hypotheses that links multiple heterogenous knowledge sources. PHK will provide a rich AI ready encoding of data, information, and knowledge from any number of new and existing sources including curated databases (e.g., NIH Cancer Genome Atlas (TCGA)) and raw experimental data. PHK?s encoding enables additional knowledge augmentation through probabilistic and statistical inferencing capabilities. This augmentation is further enhanced through knowledge unification and fusion algorithms closely coupling disparate knowledge in a well-defined, rigorous manner. Altogether, PHK can systematically discover and develop novel hypotheses. BKBs have been applied and deployed across a number of projects for over two decades with a mature software base for ready integration into the Translator framework?s target prototypes. Our multi-institutional team (Dartmouth College, Tufts University Clinical and Translational Sciences Institute (CTSI)) is comprised of senior researchers and software engineers in the computer and data sciences, cheminformatics, bioinformatics, molecular biology, and biochemistry. Dr. Eugene Santos Jr. is Professor of Engineering at Dartmouth. He will serve as the PI and will also lead in the technical development of the core BKB component for PHK. Joseph Gormley is the Director for Advanced Systems Development for Tufts/CTSI. He will serve as Project Manager for all software deliverables under this proposal. The PHK capabilities proposed herein will also be based on strategies developed by our team for scalable intelligent information retrieval where the desire for greater transparency when reasoning over experimental data is a primary aim. PHK will provide a powerful new computational representation of pathway structures and molecular components in support of both human and machine-driven interpretation and pathway-based biomarker discovery and drug development. PHK will enable more efficient joint human-machine exploration with explanation.