vandijklab/cell2sentence

What it solves

Cell2Sentence (C2S) addresses the difficulty of applying Large Language Models (LLMs) to single-cell transcriptomics data. Traditionally, this data consists of expression vectors that LLMs cannot process natively. C2S solves this by transforming these vectors into "cell sentences"—sequences of gene names ordered by their expression levels—allowing LLMs to treat biological data as natural language.

How it works

The framework uses a rank-ordering transformation to convert gene expression vectors into space-separated gene names (descending by expression). This allows LLMs (such as those based on Pythia or Gemma-2 architectures) to model scRNA-seq data. The system supports next-gene prediction, cell type prediction, and the generation of synthetic cells conditioned on specific cell types. It also enables the unification of transcriptomic data with textual data, such as paper abstracts.

Who it’s for

This tool is designed for computational biologists and researchers specializing in single-cell analysis and transcriptomics who want to leverage the latest LLM architectures for biological discovery.

Highlights

• Scalable Architectures: Supports models up to 27B parameters using Pythia and Gemma-2.
• Multi-task Capabilities: Enables perturbation prediction, dataset summarization, cluster captioning, and biological question answering.
• Comprehensive Model Zoo: Provides pretrained models for cell type prediction, conditioned cell generation, and diverse multi-cell tasks.
• Flexible Workflow: Includes built-in support for custom prompt templates and multi-cell prompt formatting for finetuning.