Using Generative LLMs to Extract Pathways from the Scientific Literature

Tracking biological and chemical research of dual-use concern is extraordinarily time consuming, given the vast size and ever-increasing nature of the scientific literature.  Fortunately, new artificial intelligence technologies are being developed what can help analysts rapidly by automatically harvesting relevant information from the literature, effectively “unlocking” this vast pool of data. In this project, we propose to extend research on information extraction that our team members have previously done with several relevant systems, including BioDDEx and SynKB. Specifically, we will employ generative LLMs (Large Language Models) to extract pathway data from the scientific literature.  We believe that generative LLMs will enable us to significantly improve upon the accuracy as well as scope of previous work in this area.  Additionally, we plan to experiment with new techniques our team has recently developed that can boost the accuracy of generative LLMs on complex extraction tasks. In addition to addressing the challenges of the extraction problem, our will explore how pathway data can best searched and browsed in a visual way that facilitates the work of analysts.  This will include investigating how the extracted data can best be integrated, adapted, and utilized by downstream systems that support DoD analysts.   Finally, we will also develop both a feasibility demonstration and an evaluation to establish that the research plan is viable and that the results we produce will have practical value.