Keerthana Nallamotu will present her General Exam "Reliable Reasoning in Genomics" on Wednesday, May 6, 2026 at 11:15 AM in CS 302 and via zoom. Zoom link: https://princeton.zoom.us/j/3674686562 Committee Members: Olga Troyanskaya (advisor), Ellen Zhong, Yuri Pritykin Abstract: Understanding the fundamental drivers of disease requires bridging the gap between molecular mechanisms and clinical phenotypes across different biological scales. However, the fragmentation of genomic data across isolated repositories and the tendency of Large Language Models (LLMs) to hallucinate domain-specific facts make it difficult to design a system for reliable discovery. To address these challenges, we first introduce MiSTIC-KG, a multi-scale, cell type-specific knowledge graph that serves as a structured foundation for biological inquiry. The schema integrates diverse entities, including cell types, tissues, genes, diseases, phenotypes, and drugs, and embeds tissue-specificity directly into gene-gene functional interaction edges to model more complex biological processes. By capturing these relationships, MiSTIC-KG enables advanced downstream applications, including synthetic lethality prediction, drug repurposing, and drug toxicity assessment. We then introduce Alvessa, an agentic framework designed to orchestrate complex reasoning over static knowledge bases and dynamic foundation models. We evaluate this approach by constructing GenomeArena, a comprehensive benchmark suite that demonstrates how Alvessa’s evidence-grounded strategy significantly reduces hallucination and outperforms general-purpose LLMs in genomic reasoning tasks. By advancing both high-resolution knowledge representation and verifiable agentic reasoning, this work takes a step towards utilizing fragmented genomic data for reliable biological discovery. Reading List: https://docs.google.com/document/d/1qsh3eItEWunmqCqGrzp-4nDLMOu__CSfGu9L_iAu... Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.