SCTA: An Agentic Framework for Stable and Interpretable Target Gene Discovery from Single-Cell RNA Sequencing

Identifying therapeutic targets from single-cell RNA sequencing (scRNA-seq) data is notoriously fragile. Small variations in analytical choices—such as normalization, clustering, or differential expression testing—can lead to vastly different candidate gene lists. This paper addresses the lack of reproducibility in target discovery by introducing SCTA (Single-Cell Target Agent), a framework designed to prioritize stable and biologically coherent target selection over simple, single-run performance.

SCTA moves away from general-purpose analysis agents by using a decision-centric, multi-agent architecture. The framework decomposes the target discovery pipeline into five specialized, sequential roles:

1. Preprocessing Agent: Handles data quality and initial differential expression.
2. Annotation Agent: Maps clusters to biological cell types.
3. Insight Agent: Identifies disease-relevant populations using enrichment evidence.
4. Filter Agent: Aggregates multi-modal evidence (e.g., gene summaries, interaction networks) to prioritize a final target set.
5. Reviewer Agent: Manages error handling and retries.

By enforcing unidirectional information flow and restricting each agent to a predefined toolset, SCTA ensures that downstream decisions are explicitly conditioned on upstream outputs. The authors evaluated the framework by measuring the stability of target gene selection across multiple independent runs, comparing it against general-purpose baselines and ablated versions of the SCTA pipeline.

SCTA consistently outperforms general-purpose baselines in both target stability and biological coherence. In a case study on hereditary chronic pancreatitis, the full SCTA configuration achieved significantly higher pairwise Jaccard similarity across independent runs compared to ablated versions where biological knowledge or enrichment evidence was removed. The results demonstrate that integrating structured biological evidence—specifically network-context and pathway-level support—is critical for ensuring that the final target shortlist remains consistent across repeated analyses. This stability is vital for translational research, where the high cost of experimental validation makes the reproducibility of candidate prioritization a primary concern.