Single-cell analysis model driven by plain-language instructions, covering cell type annotation, pseudo-cell generation, and drug response prediction.
No providers recorded yet. Browse all providers
Single-cell RNA sequencing produces data that most analysis pipelines expose only through code: annotating cell types, simulating a cell population, or asking whether a cell will respond to a drug each means assembling a bespoke script around a task-specific model. InstructCell removes that layer by making natural language the interface. A user types an ordinary sentence alongside an expression profile, and the model decides which task is being asked for and returns either a textual answer or a synthesised cell.
InstructCell was developed at Zhejiang University by the group behind ChatCell, and released as a preprint in January 2025. Where ChatCell converted cells into sentences of gene names for a text-only model, InstructCell keeps gene expression in its native numeric form and gives the language model a dedicated encoder and decoder for it. That makes the transcriptome a first-class modality rather than something serialised into text, and it lets the same checkpoint both read cells and write them.
The model sits alongside single-cell foundation models such as scGPT, Geneformer, and scBERT, but occupies a different position: those models supply general-purpose cell embeddings that a practitioner adapts per task, whereas InstructCell packages three concrete analyses behind one conversational entry point.
InstructCell uses a T5-base encoder-decoder as its language backbone. Expression profiles enter through a Q-Former module with four transformer layers, eight learnable query tokens, and six key-value pairs, feeding a three-layer MLP with skip connections; ablations show eight query tokens to be optimal, with additional queries introducing redundancy. Cell generation runs through a conditional VAE whose encoder produces posteriors over a latent and a library-size term, and whose decoder emits zero-inflated negative binomial parameters over genes; a fully connected layer projects the 768-dimensional hidden state to the 256-dimensional conditioning vector that drives it.
Training data comprises 299,155 scRNA-seq profiles drawn from ten public datasets covering human and mouse tissues including pancreas, blood, lung, and bone marrow. Mouse genes are mapped to human orthologs, and the top 3,600 highly variable genes per dataset are merged into a shared vocabulary of 18,961 genes. Instruction-response templates were synthesised with GPT-4o across 79 personality traits and per-task motivations, deduplicated at a ROUGE-L threshold of 0.75, yielding 2,787 templates for annotation, 2,395 for drug sensitivity, and 2,080 for pseudo-cell generation. On annotation the model matches or exceeds scBERT, scGPT, and Geneformer on weighted F1, macro F1, and accuracy; on generation it achieves lower maximum mean discrepancy than scDiffusion and scGAN; on drug sensitivity it leads on all three metrics for the GSE117872 and GSE110894 cohorts and is comparable on GSE149383.
The intended user is a bench biologist or clinician-scientist who has scRNA-seq data and a question, but not a computational pipeline. Annotating an unlabelled dataset, generating synthetic cells of a specified type to augment a rare population, and screening predicted drug response across a tumour cell population all become single-sentence requests against a released checkpoint. Because the gene vocabulary spans human and mouse orthologs, the same model serves cross-species comparisons without a separate checkpoint per organism.
InstructCell is an argument that the bottleneck in single-cell analysis is increasingly the interface rather than the representation: general-purpose cell embeddings are plentiful, but turning one into an answer still demands expertise most wet-lab groups do not have. By pairing a modest 220M-parameter backbone with purpose-built encoders and decoders for expression data, it reaches parity with much larger single-cell foundation models on the tasks it covers, which suggests that instruction tuning on a well-constructed multi-modal dataset can substitute for raw scale. Its coverage is deliberately narrow — three tasks, ten training datasets — and the work remains a preprint awaiting peer review. Code, both checkpoints, and the preprocessing pipeline are released under the MIT license.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.