Biomedical Data Translator Technical Feasibility Assessment of Reasoning Tool: Broad Institute

Broad Institute, Inc., Cambridge MA

Investigators

Linked publications & trials

Abstract

In the framework of the Translator, user questions can be considered requests to identify connections between biological entities. The central idea of our proposed prototype is to lift this task from specific instances (?imatinib?) to high-level entities (?drug?). This allows us to efficiently plan knowledge retrieval and reason about specific instances by using knowledge about abstract entities. The agent uses probabilistic inference and online replanning to handle uncertainties that arise when moving from abstract concepts to instances (e.g., a drug?target database may not contain a specific drug; binding strengths vary for specific drug?target pairs). Consistent with the ?identify-query-analyze? framework of the Translator, the agent breaks the task of answering a given user query into four steps: (1) parsing the query into a standardized format for internal representation, (2) high-level planning to identify relevant KSs and develop a query strategy, (3) knowledge retrieval by executing the query, and (4) analysis and inference on the obtained knowledge to produce an answer to the query (Figure 1). We implemented the planning module (2) as a Markov decision process (MDP). Analysis and inference (4) is performed using a combination of probabilistic graphical models (PGMs) and conventional graph algorithms. To allow replanning, the agent can iterate over steps (2) ? (4) to respond to failed actions or improve on a preliminary answer.