Alvessa: An Agentic Evidence-Grounded Research Assistant for Genomics

Ksenia Sokolova, Sanketh Vedula, Keerthana Nallamotu, Guillermo Sapiro, Olga G Troyanskaya

Additional Submission Instructions: For the camera-ready version, please include the author names and affiliations, funding disclosures, and acknowledgements.

Track: Track 1: Original Research/Position/Education/Attention Track

Keywords: agentic models, computational biology, genomics, ai for science

Abstract:

The rapid proliferation of genomic data, large-scale experiments, and biological foundation models presents major opportunities for biological discovery, but also creates significant integration challenges. Researchers often face a landscape of heterogeneous databases with inconsistent formats, a challenge compounded by the difficulty of integrating this static knowledge with dynamic predictions from foundation models. Furthermore, knowledge from publications quickly becomes outdated and disconnected from new evidence. To address these gaps, we present Alvessa, a research assistant that orchestrates modular agents to perform user-intent understanding, context-specific tool calling, reasoning, and evidence-backed summarization. Alvessa integrates a diverse array of genetic databases, specialized foundation models, and bioinformatics tools, and dynamically selects tools needed for a given query.
Unlike conventional portals that return data without reasoning, or general-purpose language models whose conclusions may be outdated due to static training data, Alvessa actively retrieves relevant evidence for a given query, reasons over it to synthesize a conclusion, and presents both the answer and its supporting evidence in an interactive interface.
For reproducible assessment, we introduce GenomeArena, a suite of curated benchmarks for evaluating the framework's core components, enabling granular measurement of entity extraction, tool selection, reasoning, and evidence verification. Our results show high performance while highlighting limitations when necessary databases are not yet integrated.

Submission Number: 301