KODA: An agentic framework for KEGG orthology-driven discovery of antimicrobial drug targets in gut microbiome

Javad Aminian Dehkordi, Mohammad S. Parsa, Mohsen Naghipourfar, Mohammad Mofrad

Keywords: AI Agents, Large Language Models, Knowledge Graphs, Gut microbiome, Antimicrobial Drug Targets, KEGG Orthology

TL;DR: We present KODA, a multi-agent LLM framework that uses knowledge graphs to identify potential antimicrobial drug targets in the gut microbiome via KEGG-based metabolic analysis.

Abstract:

The gut microbiome significantly impacts human health and disease by modulating key biological functions, including immune responses and nutrient processing. Despite its importance, the intricate web of microbial interactions and their metabolic interdependencies remains largely elusive. In this work, we introduce KODA—a novel agent-based framework that employs large language models (LLMs) and knowledge graphs (KGs) to streamline the identification of antimicrobial drug targets within the gut microbiome. KODA operates through a collaborative multi-agent architecture that transforms natural language queries into structured graph queries, facilitating user-friendly exploration of complex microbiome datasets.

By focusing on KEGG orthologs associated with essential microbial genes, KODA pinpoints candidate targets for antimicrobial intervention through the analysis of metabolic pathways. At its core lies a Neo4j-powered microbiome knowledge graph, which integrates data on microbial interactions, metabolic networks, and KEGG-derived annotations. To ensure robustness, the system incorporates an evaluation pipeline where LLM-based agents review both query quality and analytical outputs.

Our findings highlight KODA’s capability to yield biologically relevant insights, especially in uncovering conserved, essential genes that may serve as promising drug targets. This framework not only enhances antimicrobial research but also aims to broaden access to microbiome analytics, lowering the technical threshold for researchers and accelerating early-stage drug discovery.

Submission Number: 27