Single-cell RNA sequencing has become one of the most powerful tools for studying complex biological systems. By measuring gene expression in individual cells, researchers can identify different cell types, track cellular development, and better understand disease processes. However, one major challenge remains, accurately determining the identity of each cell after sequencing data has been generated.

This process, known as cell type annotation, is often time-consuming and requires significant biological expertise. Traditional computational approaches typically rely on reference datasets containing previously characterized cell types. While effective in many situations, these methods can struggle when analyzing novel datasets or previously unseen cell populations.

More recently, researchers have begun exploring the use of large language models (LLMs) to automate cell type annotation. These artificial intelligence systems can interpret lists of marker genes and suggest likely cell identities. However, relying on a single LLM can introduce model-specific biases and may provide limited information about uncertainty in the predictions.

To address these challenges, researchers at Texas A&M University have developed a new framework called mLLMCelltype.

Rather than relying on a single AI model, mLLMCelltype uses multiple large language models that work together through an iterative deliberation process. The concept is inspired by collective intelligence, the idea that groups of independent problem-solvers can often arrive at more accurate conclusions than individuals working alone.

In practice, several LLMs independently evaluate cell type information derived from single-cell RNA sequencing data. The models then compare and refine their predictions through structured discussions before reaching a final consensus. This collaborative approach helps reduce individual model biases and improves confidence in the final annotation.

The researchers evaluated mLLMCelltype using 49 diverse datasets. Across these datasets, the framework achieved an average annotation accuracy of 77.2%, substantially outperforming the best single-LLM approach, which achieved an average accuracy of 61.5%.

An important advantage of the framework is its robustness. The consensus-based approach remained effective even when the input data contained noise and successfully generalized to datasets released after the underlying language models had been trained. This suggests that the method can adapt to new biological discoveries without requiring extensive retraining.

The framework also provides transparent reasoning chains that allow researchers to understand how cell type assignments were made. In addition, it generates confidence scores based on consensus among the participating models, giving users a clearer understanding of prediction reliability.

These features could significantly reduce the amount of manual effort required for cell type annotation while improving confidence in downstream analyses. As single-cell RNA sequencing datasets continue to grow in size and complexity, tools that automate data interpretation will become increasingly important.

The researchers have made mLLMCelltype available as both an open-source software package and a web server, allowing the broader scientific community to apply the technology to their own datasets. By combining the strengths of multiple AI systems, mLLMCelltype demonstrates how collaborative artificial intelligence approaches may help accelerate biological discovery and improve interpretation of increasingly complex single-cell datasets.

Availability – https://github.com/cafferychen777/mLLMCelltype