single-cell-foundation-model-langcell

Zero-shot single-cell annotation aligns cell embeddings with textual descriptions of candidate cell identities. You provide ontology context and text descriptions, and the framework matches unlabeled cells to identities without requiring labeled training data.

Preprocessing AnnData for single-cell tokenization converts raw expression matrices into the tokenized format expected by language-assisted models. This step ensures cell profiles and ontology JSONs are correctly structured for multimodal text-cell matching.

Few-shot annotation works when you have a small number of labeled cells by leveraging both text descriptions and limited labels. It aligns cell embeddings with ontology context to classify the remaining cells, reducing the labeled data requirement.

Cell-encoder finetuning adapts a pretrained model to your specific single-cell dataset by updating cell embeddings. Use it when zero-shot or few-shot annotation lacks accuracy and you need the encoder to learn dataset-specific biological signals.

Multimodal cell-text matching requires ontology JSONs for cell identity definitions and external checkpoints for the pretrained encoders. These assets provide the structured vocabulary and model weights necessary to align text and cell embeddings.

Preparing text descriptions for candidate cell identities involves writing biological summaries of each expected cell type using ontology terms. These descriptions guide the language model to match cell embeddings against the correct textual context.