CellTypeAgent: Trustworthy cell type annotation with Large Language Models

Furthermore, the subsequent verification based on average gene expression correctly identifies pericyte cell, in contrast to GPTCellType's misclassification as fibroblasts. We further assessed various factors that could influence CellTypeAgent's performance. First, we examined how the number of initial candidate cell types in the 4 inference step affected accuracy (Figure 2c). Although performance remained relatively stable, using the top three candidates yielded a slight higher performance. We also evaluated the impact of the number of marker genes, finding that including more genes generally enhanced annotation quality (Figure 2d). Moreover, we tested CellTypeAgent's ability to handle mixtures of different cell types (Figure 2e). When explicitly prompted that multiple cell types might be present, the agent successfully identified one or more components within the mixed sample. While performance declined compared to annotating pure cell types, CellTypeAgent still demonstrated the capacity to accurately detect multiple cell types.