Interpretable RNA-Seq Clustering with an LLM-Based Agentic Evidence-Grounded Framework

TL;DR

This work addresses the interpretive bottleneck in RNA-seq cluster analysis by focusing on finding concrete, literature-grounded explanations for gene modules rather than generic enrichment results. It introduces CITE V.1, an agentic LLM-based framework with Retriever, Interpreter, and Critics to ground RNA-seq cluster interpretations in PubMed and UniProt evidence. In comparative evaluations against a Gemini baseline, CITE V.1 provides structured, auditable interpretations with reliability flags and confidence scores, reducing speculative outputs and false citations. The approach enhances transparency and reproducibility in biomedical AI and offers a scalable foundation for evidence-grounded analyses across pathogens.

Abstract

We propose CITE V.1, an agentic, evidence-grounded framework that leverages Large Language Models (LLMs) to provide transparent and reproducible interpretations of RNA-seq clusters. Unlike existing enrichment-based approaches that reduce results to broad statistical associations and LLM-only models that risk unsupported claims or fabricated citations, CITE V.1 transforms cluster interpretation by producing biologically coherent explanations explicitly anchored in the biomedical literature. The framework orchestrates three specialized agents: a Retriever that gathers domain knowledge from PubMed and UniProt, an Interpreter that formulates functional hypotheses, and Critics that evaluate claims, enforce evidence grounding, and qualify uncertainty through confidence and reliability indicators. Applied to Salmonella enterica RNA-seq data, CITE V.1 generated biologically meaningful insights supported by the literature, while an LLM-only Gemini baseline frequently produced speculative results with false citations. By moving RNA-seq analysis from surface-level enrichment to auditable, interpretable, and evidence-based hypothesis generation, CITE V.1 advances the transparency and reliability of AI in biomedicine.

Figures (6)

• Figure 1: Overview of the CITEV.1 framework for interpretable RNA-seq analysis. Clusters are enriched with PubMed and UniProt evidence, interpreted by the Interpreter Agent, and validated through a panel of Critic Agents, producing transparent literature-grounded outputs.
• Figure 2: Supporting analyses from CITEV.1. (a) Distribution of references (specific, generic, and UniProt) demonstrates the framework’s layered evidence grounding, with the absence of PubMed-specific hits in this dataset reflecting dataset-specific coverage. (b) Keyword extraction reveals recurrent biological themes and functional linkages across clusters.
• Figure 3: Comparison of interpretations for Cluster 1. (a) CITEV.1 provides structured, reference-grounded outputs with explicit limitations. (b) Gemini baseline yields speculative, less verifiable interpretations with placeholder references explicitly returned as "[Citation Needed]" in the original Gemini output.
• Figure 4: Comparative visualization of cluster interpretations: (a) CITEV.1 produced structured, reference-grounded interpretations with critic-based reliability for Cluster 1, while (b) Gemini baseline on Cluster 1 was speculative, misclassified, and lacked evidence.
• Figure 5: MultiQC summary of sequence counts across samples. Blue bars indicate unique reads, while black bars denote duplicate reads.