Why Multiple Models?

Single large language models (LLMs) can produce impressive results for cell type annotation, but they also have limitations:

1. Knowledge gaps: No single model has perfect knowledge of all cell types across all tissues
2. Hallucinations: LLMs can sometimes generate plausible-sounding but incorrect annotations
3. Biases: Each model has biases based on its training data and architecture
4. Inconsistency: The same model may give different answers to the same question

The multi-LLM consensus approach addresses these limitations by leveraging the complementary strengths of different models. This is similar to how a panel of experts might collaborate to reach a more reliable conclusion than any single expert could provide alone.

Model Diversity

mLLMCelltype deliberately incorporates models with different architectures and training data:

• Anthropic Claude models: Known for careful reasoning and biological knowledge
• OpenAI GPT models: Strong general knowledge and pattern recognition
• Google Gemini models: Good at integrating information from multiple sources
• X.AI Grok models: Newer architecture with different training approach
• Other models: Provide additional diversity in reasoning approaches

This diversity is crucial for the consensus mechanism to work effectively. Models with different “perspectives” can catch each other’s errors and provide complementary insights.

1. Initial Independent Annotation

Each model independently annotates the cell clusters based on the marker genes:

# Conceptual representation of the initial annotation process
initial_results <- list()
for (model in models) {
  initial_results[[model]] <- annotate_cell_types(
    input = marker_data,
    tissue_name = tissue_name,
    model = model,
    api_key = api_keys[[get_provider(model)]]
  )
}

This step ensures that each model forms its own opinion without being influenced by others, similar to how jurors might form initial opinions before deliberation.

2. Identification of Controversial Clusters

The system identifies clusters where there is significant disagreement among models:

# Conceptual representation of controversial cluster identification
controversial_clusters <- identify_controversial_clusters(
  initial_results,
  threshold = discussion_threshold
)

A cluster is considered “controversial” if the proportion of models agreeing with the most common annotation is below a certain threshold (default: 0.5). This means that if less than half of the models agree on the cell type, the cluster requires further discussion.

3. Structured Discussion for Controversial Clusters

For controversial clusters, the system initiates a structured discussion process:

# Conceptual representation of the discussion process
discussion_results <- facilitate_cluster_discussion(
  controversial_clusters,
  initial_results,
  marker_data,
  tissue_name,
  discussion_model,
  api_key
)

This discussion follows a specific format:

1. Initial positions: Each model’s initial annotation and reasoning are presented
2. Evidence evaluation: The discussion model evaluates the evidence for each proposed cell type
3. Counter-arguments: Potential weaknesses in each argument are identified
4. Synthesis: The discussion model synthesizes the arguments to reach a conclusion

This structured approach mimics how human experts might deliberate on a difficult case, considering multiple perspectives and critically evaluating the evidence.

4. Final Consensus Formation

After discussion, the system forms a final consensus for all clusters:

# Conceptual representation of consensus formation
final_annotations <- combine_results(
  initial_results,
  discussion_results,
  controversial_clusters
)

For non-controversial clusters, the most common annotation among the initial results is used. For controversial clusters, the result from the structured discussion is used.

Consensus Proportion

The consensus proportion measures the level of agreement among models:

# Conceptual calculation of consensus proportion
consensus_proportion <- sapply(clusters, function(cluster) {
  annotations <- sapply(models, function(model) initial_results[[model]][cluster])
  most_common <- names(which.max(table(annotations)))
  sum(annotations == most_common) / length(annotations)
})

This metric ranges from 0 to 1: - 1.0: Perfect agreement (all models agree) - 0.5: Moderate agreement (half of the models agree) - < 0.5: Low agreement (less than half of the models agree)

The consensus proportion helps identify which annotations are more reliable and which might require further investigation.

Shannon Entropy

Shannon entropy quantifies the uncertainty in the annotations:

# Conceptual calculation of Shannon entropy
shannon_entropy <- sapply(clusters, function(cluster) {
  annotations <- sapply(models, function(model) initial_results[[model]][cluster])
  p <- table(annotations) / length(annotations)
  -sum(p * log2(p))
})

Shannon entropy is a measure from information theory: - 0: No uncertainty (all models give the same answer) - Higher values: More uncertainty (models give diverse answers)

Unlike consensus proportion, Shannon entropy captures the full distribution of annotations, not just the most common one. This makes it particularly useful for identifying clusters with high uncertainty.

Cross-Model Verification

By requiring multiple independent models to agree, the system naturally filters out many hallucinations. A hallucinated annotation from one model is unlikely to be independently hallucinated by other models.

Evidence-Based Reasoning

The structured discussion process explicitly requires models to ground their annotations in the marker gene evidence. This reduces the likelihood of hallucinations that aren’t supported by the data.

Critical Evaluation

During the discussion process, models critically evaluate each other’s reasoning. This helps identify and correct potential hallucinations or reasoning errors.

Collective Error Correction

Even if some marker genes are noisy or misleading, the consensus approach can still reach the correct conclusion if enough models can identify the true signal in the data.

Focus on Strong Signals

By using the top_gene_count parameter, the system focuses on the strongest marker genes, which are less likely to be noise.

Uncertainty Flagging

When the input data is too noisy to make a reliable annotation, the system will show low consensus proportion and high Shannon entropy, flagging the cluster for human review.

Prompt Engineering

The prompts used in mLLMCelltype are carefully designed to:

1. Structure the reasoning process: Guide models through a step-by-step analysis
2. Enforce evidence-based reasoning: Require models to cite specific marker genes
3. Encourage critical thinking: Ask models to consider alternative explanations
4. Standardize output format: Ensure consistent, parseable responses

Here’s a simplified example of the annotation prompt structure:

You are an expert in single-cell RNA sequencing analysis.

TASK:
Identify the cell type for a cluster based on its marker genes.

MARKER GENES:
[List of marker genes with fold changes and p-values]

TISSUE:
[Tissue name]

STEPS:
1. Analyze the marker genes and their expression levels
2. Identify key cell type-specific markers
3. Consider multiple possible cell types
4. Determine the most likely cell type based on the evidence
5. Provide your reasoning

OUTPUT FORMAT:
Cell Type: [Your answer]
Reasoning: [Your step-by-step reasoning]
Confidence: [High/Medium/Low]

Discussion Orchestration

The discussion process is orchestrated by a single “discussion model” (typically Claude) that:

1. Presents each model’s initial annotation and reasoning
2. Evaluates the evidence for each proposed cell type
3. Identifies potential weaknesses in each argument
4. Synthesizes the arguments to reach a conclusion

This approach allows for a coherent discussion while still incorporating the diverse perspectives of multiple models.

vs. Single LLM Annotation

Compared to using a single LLM: - Advantages: Higher accuracy, uncertainty quantification, reduced hallucinations - Disadvantages: Higher computational cost, more complex implementation, requires multiple API keys

vs. Traditional Annotation Methods

Compared to traditional methods (e.g., reference-based, marker-based): - Advantages: No reference dataset required, more flexible, captures rare cell types, provides reasoning - Disadvantages: Depends on LLM knowledge, potentially higher cost, requires internet connection

vs. Human Expert Annotation

Compared to human expert annotation: - Advantages: Faster, more scalable, consistent methodology, transparent reasoning - Disadvantages: May miss novel cell types not in literature, lacks domain-specific expertise for very specialized tissues