Cell type annotation by committee, with entropy

What it does

The framework takes marker gene lists from single-cell RNA-seq clusters and submits them to multiple commercial LLMs for cell type identification. It then aggregates the responses through an iterative discussion loop, yielding a consensus label plus uncertainty scores. The tool plugs into existing Scanpy or Seurat workflows and, according to the README, requires no pre-trained reference atlas.

The interesting bit

Instead of treating a single LLM as ground truth, the framework leans into model disagreement: it runs a multi-round deliberation and surfaces metrics like Consensus Proportion and Shannon Entropy so you can see exactly which clusters confuse the machines. The README cites benchmark accuracy up to 95% on tested datasets.

Key highlights

• Reference-free annotation using only cluster marker genes
• Modular provider support; the README lists OpenAI, Anthropic, Google, DeepSeek, Alibaba, and others, with optional per-provider dependencies
• Built-in uncertainty quantification via consensus and entropy metrics
• Hierarchical annotation with consistency checks across resolution levels
• Complete reasoning logs for every annotation decision, plus a web interface at mllmcelltype.com

Caveats

• Requires valid API keys for each LLM provider you want to include; costs scale with the number of models polled.
• Gene symbols are mandatory—Ensembl IDs will not work without manual conversion.
• Optional dependency management means missing provider SDKs trigger ImportError at runtime rather than install time.

Verdict

Wet-lab biologists drowning in unlabeled scRNA-seq clusters should try this, especially if they lack a matching reference atlas. Anyone trying to avoid cloud API costs or keep pipelines fully offline should skip it.