Title: BRAINCELL-AID: A Resource for Community Annotation of Brain Cell Types via Agentic AI

Abstract: The advent of single-cell RNA sequencing has revolutionized the identification of diverse cell types and their associated transcriptomic profiles. Despite this progress, annotating these profiles—particularly when they involve poorly understood genes—presents a significant hurdle. Conventional techniques like Gene Set Enrichment Analysis (GSEA) rely heavily on pre-existing, high-quality annotations and frequently yield suboptimal results in such complex scenarios. While Large Language Models (LLMs) present a viable alternative, they often fail to adequately capture intricate biological knowledge within structured ontological frameworks.

To overcome these limitations, we introduce BRAINCELL-AID (available at https://biodataai.uth.edu/BRAINCELL-AID), an innovative multi-agent AI system designed to enhance the accuracy and robustness of gene set annotation by combining free-text descriptions with ontology labels. By leveraging Retrieval-Augmented Generation (RAG), we established a resilient agentic workflow that refines predictions through the integration of relevant PubMed literature. This approach significantly minimizes hallucinations and improves the interpretability of the results.

In evaluations, this workflow successfully identified correct annotations for 77% of mouse gene sets within their top predictions. We applied this methodology to annotate 5,322 brain cell clusters derived from the comprehensive mouse brain cell atlas produced by the BRAIN Initiative Cell Census Network. This effort has facilitated new insights into brain cell functionality, including the discovery of region-specific gene co-expression patterns and the inference of functional roles for various gene ensembles. Furthermore, BRAINCELL-AID successfully identified Basal Ganglia-related cell types accompanied by neurologically significant descriptions. Ultimately, this project establishes a valuable resource to facilitate community-driven cell type annotation.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC