Abstract A099: Creating an interconnected pancreatic cancer research knowledge system using a Large Language Model and wiki publishing platform
TL;DR
The authors have created a fully functioning, curated, dynamic knowledge tool that can be used by researchers to deeply explore focused research areas as well as visualize the interconnections between different treatments and research foci across pancreatic cancer research.
Abstract
The authors sought to utilize Large Language Models (LLMs) to interconnect previously separate areas of pancreatic cancer research into a dynamic research tool for the research community. The historical weaknesses of LLMs include hallucinations, dependence on unverified internet sources, and lack of deterministic outcomes. However, dramatic advances in the power of Artificial Intelligence (AI) models and agentic AI tools occurred in early 2026. Using Andrej Karpathy’s LLM-Wiki framework, the authors have created a fully functioning, curated, dynamic knowledge tool that can be used by researchers to deeply explore focused research areas as well as visualize the interconnections between different treatments and research foci across pancreatic cancer research. Published journal articles and major conference abstracts are the source of truth for the wiki (an online hypertext linked publication), not the public internet or AI model training knowledge. 15 Research Lanes (e.g. KRAS, Immunotherapy, etc.) currently form the spine of the wiki. These research lanes are dynamic, and are adjusted as new research areas develop. The wiki content is also dynamic - each source type (paper, abstract, trial record) has its own automated ingestion timing cycle. Each selection, ingestion, and authoring process is clearly defined by a set of “skills” that the authors have created in cooperation with the LLM that the LLM must follow. The authors set the rules and provide oversight, but the LLM triages candidate sources and creates and links the wiki content. Accuracy is enforced by regular adversarial LLM agent review and human oversight. Each candidate paper or abstract in the candidate pool goes through an LLM driven triage process that includes relevance to Pancreatic Ductal Adenocarcinoma. Sources which potentially fill a gap, reinforce a thinly sourced finding, contradict an earlier finding, or connect research lanes or treatments are favored for selection. The source pages that are created from those papers and abstracts are the source of truth for the wiki, and essentially every downstream synthesis in the wiki stems from and links to those primary research pages. Additional pages such as concepts, interconnections, asked and answered queries, and clinical trial records are synthesized by the LLM directly from the source pages and clinicaltrials.gov. The wiki currently contains syntheses of 286 published papers and 132 abstracts, with a total of 1,397 pages connected by 21,522 links. The wiki structure and LLM integration resulted in a dynamic interconnected knowledge system that is curated, easily accessible, up to date, and intuitively navigable. The wiki is fully operational and currently available to the research community. Bernard A. Kroll, Joshua P. Raff, Edward A. Fisher. Creating an interconnected pancreatic cancer research knowledge system using a Large Language Model and wiki publishing platform [abstract]. In: Proceedings of the AACR Conference on Pancreatic Cancer: New Frontiers in Biology and Therapeutic Development; 2026 Sep 25-28; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(18_Suppl_2):Abstract nr A099.