Determining Cell Type

This function allows you to determine which cell type is most likely to be the true cell type of a cluster, utilizing multiple LLMs. Under default settings, 3 state-of-the-art LLMs are used to score the candidate cell types based on the marker genes.

Function Parameters

compareCelltypes(
    tissue,        # The tissue type being analyzed (e.g., "large intestine")
    celltypes,     # Vector of cell types to compare (e.g., c("Plasma Cells", "IgA-secreting Plasma Cells"))
    marker,        # String of marker genes separated by commas
    species,       # Species of origin ("human" or "mouse")
    output_file,   # Name for the output file (e.g., "plasma_cell_subtype")
    model_list     # Optional: List of LLM models to use (has default values)
)

Parameter Details

1. tissue:

   • Type: character string
   • Specifies the tissue source of your data
   • Example: "large intestine", "small intestine", "brain"

2. celltypes:

   • Type: character vector
   • List of cell types you want to compare
   • Maximum recommended: 4-5 cell types for optimal results
   • Example: c("Plasma Cells", "IgA-secreting Plasma Cells", "IgG-secreting Plasma Cells")

3. marker:

   • Type: character string
   • Comma-separated list of marker genes
   • Can include both up and down-regulated markers
   • Example: "IGLL5, IGLV6-57, JCHAIN, FAM92B, IGLC3"

4. species:

   • Type: character string
   • Options: "human" or "mouse"
   • Specifies the species origin of your data

5. output_file:

   • Type: character string
   • Name for the output file (without extension)
   • Example: "plasma_cell_subtype"

6. model_list:

   • Type: vector of strings
   • Optional parameter
   • Default models (if none provided):

     model_list = c(
         "anthropic/claude-3.5-sonnet",  # Anthropic's latest model
         "openai/o1-mini",              # OpenAI's model
         "google/gemini-pro-1.5"        # Google's model
     )
   • These default models are selected as they represent state-of-the-art LLMs

Output Format

1. Console Output:

   • Similarity scores from each LLM for each cell type
   • Consensus results (if reached)
   • Warning messages (if any)

2. Output File (saved as "[output_file].txt"):

   • Detailed comparison results from each LLM
   • Marker gene analysis
   • Final consensus (if reached)

Interpretation Guide

High Confidence Result

• A high confidence result is obtained when ALL LLMs give a score above 80% for the same cell type
• This indicates a clear, unambiguous cell type identification

No Consensus Reached

If no clear consensus is reached, consider these possible scenarios:

1. Low Quality Cluster

   • Symptom: Inconsistent or low scores across LLMs
   • Solution: Increase the number of marker genes in your analysis

2. Mixed Cluster

   • Symptom: Different LLMs strongly favor different cell types
   • Solution: Perform subclustering to separate potential distinct populations

3. Last Resort

   • If issues persist after trying the above solutions
   • Consult domain experts for manual review