Ayesha Saleem
Research Mentor(s): Yongqun He
Mentor Department:
Authors: Ayesha Saleem, Anthony Huffman, Yongqun He
Session: Session 7 (4:00pm – 4: 50pm)
Presentation Type: Poster 7
Abstract
After the COVID-19 pandemic of 2020, the Omicron variant of SARS-CoV-2 has since become the dominant strain that continues to spread throughout the world. Different variants of SARS-CoV-2 are characterized by specific mutations that effect viral transmission, disease severity, and the efficacy of COVID-19 vaccines. Human Phenotype Ontology (HPO) provides a standardized language for precise phenotypic analysis; current HPO tools are inefficient and struggle to analyze incomplete phenotype assignments, as these tools are prone to “hallucinations†and are circumstantially unreliable. This study focused on using Coronavirus Infectious Disease Ontology (CIDO) to train a large-language model (CIDO-RAG) to identify and report what specific mutations in Omicron variants may impact vaccine efficacy, severity of disease, and transmission efficacy of the virus and to evaluate LLM accuracy in HPO term assignment and time and resource efficiency. Using CoVariants.org, missing mutations were surveyed and annotated and then uploaded to CIDO. Three large-language model (LLM) modules were generated using Maizey and a Retrieval Augmented Generation (RAG) module to read COVID-19 studies. To assess CIDO-RAG approach, a set of questions were provided to each LLM, and the responses were checked against an answer sheet and graded on metrics such as accuracy, precision, and hallucinations. The annotation determined that there are 394 AA mutations and 20 COVID-19 variants, of which are all new variants of COVID-19 Omicron. Additionally, there were common AA variants that emerge in current active strains of COVID-19 that were identified.
