Cyberinfrastructure and Artificial Intelligence Platforms

National Human Genome Research Institute

Investigators

Abstract

The Section Cyberinfrastructure and Artificial Intelligence Platforms (the laboratory of Dr. Sean Mooney) completed its first full fiscal year in September 2025. During that time, we completed laboratory space development within the NIH clinical center and continued to expand the staff and research efforts in the laboratory. For much of the year, we consisted of two postdoctoral fellows and one staff scientist. Our focus has remained steady, that is, research at the intersection of artificial intelligence and human genetics. To that end we have contributed to several projects. Update on Lab Activities We have established regular group meetings (Thursdays at noon) and regular 1:1s with Dr. Mooney (Thursday afternoons) and we participate in NHGRI programs such as the monthly Center for Genomics and Data Science Research team meetings. We have presented there twice. We participated in several conferences including the Pacific Symposium of Biocomputing 2025 (PSB) in January and the Intelligent Systems in Molecular Biology (ISMB) Conference in July. In July we worked with CIT and ODSS to run a workshop on NIH cyberinfrastructure at ISMB. Recruitment We are restarting recruitment of postdoctoral fellows to build the laboratory. We are posting positions in computational genomics and artificial intelligence model development. We intend to recruit 5-8 postdoctoral fellows with the key goal of collaborating with other laboratories within the Intramural Research Program at the NIH. Our space in the Clinical Center (7th floor) is completed. Meetings and Presentations We are running a workshop at PSB 2026 focused on NIH Cyberinfrastructure with CIT in January. Dr. Mooney is presenting to the AMIA Clinical Research Informatics WG on December 8th. Dr. Kaiyrbekov is presenting at the AMIA Annual Symposium on November 18th. We have several ongoing projects that fit the theme of the lab. Our long-term interests are to develop automated computational tools for interpreting human genomes. We also are interest in the translation of AI based methods to clinical practice, and in understanding their effects on patients and providers when used in practice. Below some, but not all, laboratory projects are highlighted. Update on Use of a Large Language Model conversational agent to collect clinical research information from participants (Drs. Kaiyrbekov and Dobbins) Phone surveys are an essential tool for collecting health data, but they are often expensive, time-consuming for personnel, and difficult to scale. To address these limitations, we present an innovative framework that uses advanced artificial intelligence (AI) models called Large Language Models (LLMs), a type of advanced AI systems capable of processing and generating text and speech like humans. In our framework, one LLM powers a conversational agent that conducts phone surveys and transcribes the conversations, while another LLM analyzes the transcripts and extracts responses to individual survey questions. In our pilot study involving 40 surveys to test our framework, the LLM analyzing the transcripts of collected surveys achieved an impressive 98% accuracy in extracting responses to survey questions, even in the presence of some transcription errors in conversations. Survey participants noted occasional mistakes by the conversational phone agent but praised its ability to understand questions and maintain engaging conversations. Our pilot study demonstrates that the framework has potential to conduct real world phone surveys in an efficient, scalable, and cost-effective fashion. Update on Functional characterization pathogenic, benign, and variants of unknown significance in the PTPN11 gene (SHP2 protein) (Dr. Um) Understanding the molecular basis of pathogenic variants is essential for improving variants interpretation and clarifying the mechanisms that drive genetic diseases. In this project, we investigated molecular mechanisms underlying monogenic disease, with a focus on PTPN11 missense variants associated with Noonan syndrome (NS), LEOPARD syndrome (LS), and juvenile myelomonocytic leukemia. Using MutPred2, a machine-learning model that predicts structural and functional consequences of amino acid substitutions, the study quantified how these variants alter SHP2 protein structure and function. The results showed that NS- and LS-associated variants differ in their functional mechanisms related to sodium and DNA binding. Additionally, NS-associated fetal and familial variants exhibited a higher likelihood of pathogenicity, providing insights relevant to variant interpretation. This approach highlights the utility of variant-effect prediction tools in revealing potential pathogenic mechanisms, prioritizing candidate variants, and generating hypotheses for future experimental validation. Update on Development of a Large Language Model to Advance Health IT Price Transparency (Drs. Kaiyrbekov and Dobbins) Health spending totaled $74.1 billion in 1970. By 2000, health expenditures had reached about $1.4 trillion, and in 2023 the amount spent on health more than tripled to $4.9 trillion. On a per capita basis, health spending has increased in the last five decades from $353 per year in 1970 to $14,570 per year in 2023. Public health insurance programs managed in part by CMS, Medicare, Medicaid, and the Children’s Health Insurance Program (CHIP), accounted for 43.0% of all health consumption spending. One of the main culprits of increase in healthcare cost is the lack of transparency of prices for services. Transparency in prices would foster competition between providers and allow customers to shop for the financially viable treatment options thus driving prices down. To make pricing more transparent the US Government has made some initial steps. Starting January 1, 2021, each hospital operating in the United States is required to provide clear, accessible pricing information online about the items and services they provide in two ways: 1. As a comprehensive machine-readable file with all items and services. 2. In a display of shoppable services in a consumer-friendly format. However, the machine-readable files could be published in different formats such as JSON, XLSX, or CSV, and the contents within these files differ between different providers. So, even if two hospitals have the same files in a comma separated format the content in these files could differ significantly requiring custom parsers. In this project we are working on leveraging large language models to parse the diverse data and display prices in a consumer-friendly manner. We plan to employ retrieval augmented generation (RAG) and LLM finetuning.