Compare · 14 models
Single-cell foundation models you can run online
Foundation and predictive models for single-cell transcriptomics, enabling cell type annotation, latent embedding generation, gene regulatory network reconstruction, and in silico perturbation forecasting.
What single-cell foundation models models do
Single-cell foundation models address the analytical challenges of high-dimensional single-cell RNA sequencing (scRNA-seq) and multi-omics profiles. Traditional single-cell workflows often struggle to integrate heterogeneous datasets, resolve complex cell identities across disparate tissues, or forecast how cellular states shift under experimental interventions. By pre-training transformer architectures, diffusion models, or variational autoencoders on massive corpora spanning tens to hundreds of millions of transcriptomic profiles, these tools learn generalised biological representations of cell states and gene-gene interactions.
Across these methods, cellular transcriptomes are typically abstracted into structured tokens, rank-value representations, or count matrices. While several approaches focus on broad representation learning to enable cross-study batch integration and cell annotation, others explicitly model cellular responses to physical perturbations such as drug compounds, cytokine treatments, or genetic knockdowns. In addition to pure transcriptomic inputs, specialised architectures incorporate multi-modal contexts, linking single-cell transcriptomics with paired chromatin accessibility profiles or fluorescence microscopy imaging.
Compared
- CellFMMulti-institution consortiumA 100-million-cell foundation model for pan-tissue single-cell annotation and cross-dataset integration.Commercial OKNot yet
- eva-rnaScientaLabA transformer foundation model generating sample- and gene-level embeddings from human and mouse transcriptomic profiles.See its pageNot yet
- GeneformerTheodoris et al., Gladstone InstitutesA transformer model for gene network inference and target prioritisation from single-cell transcriptomes.Commercial OKNot yet
- PhenoSeqSentinal4DPredicts single-cell scGPT RNA-seq embeddings from multi-channel fluorescence microscopy imaging features.See its pageNot yet
- RhaistertahoebioA statistical model predicting cellular perturbation responses across drugs, knockdowns, and cytokines.See its pageNot yet
- SCENIC+ / pySCENICVIB-KU LeuvenInfers enhancer-driven gene regulatory networks from paired single-cell ATAC-seq and RNA-seq data.Check termsNot yet
- scGPTUniversity of TorontoA foundation model trained on 33 million human cells for single-cell annotation and perturbation prediction.Commercial OKNot yet
- SE-600MarcinstituteGenerates embeddings from human single-cell RNA profiles to aid cell-state and perturbation analysis.See its pageNot yet
- STACK LargearcinstituteA single-cell transcriptomics model for in-context learning across cell types and biological perturbations.See its pageNot yet
- Stack-Large AlignedarcinstituteA 217M-parameter single-cell foundation model aligned for cross-batch consistency and perturbation analysis.See its pageNot yet
- Tahoe-100M-SCVItahoebioA single-cell variational autoencoder mapping drug perturbations across cancer cell lines into a latent space.See its pageNot yet
- Tahoe-x1tahoebioPerturbation-trained transformer foundation models for single-cell cancer research and drug discovery.See its pageNot yet
- TEDDYMerckSingle-cell foundation models trained on 116 million cells for cross-donor and cross-disease representations.See its pageNot yet
- X-CellXaira-TherapeuticsA diffusion model predicting single-cell transcriptomic responses to CRISPRi perturbations across cellular contexts.See its pageNot yet
How to choose
When selecting a model for generalised cell type annotation, batch integration, or representation learning, the scale of pre-training data and specific architectural designs define the best fit. CellFM is trained on a 100-million-cell pan-tissue corpus, making it well-suited for harmonising labels across disparate atlases. scGPT, pre-trained on 33 million human cells, formulates transcriptomes as sentences of genes and supplies confidence estimates alongside cell type annotations. Both CellFM and scGPT demand at least 40 GB of GPU memory. Where cross-study alignment is critical, Stack-Large Aligned builds upon STACK Large's alternating cell-wise and gene-wise attention architecture, fine-tuning specifically for cross-batch consistency across repositories such as CellxGene and Parse PBMC. If the goal is prioritising disease targets rather than general annotation, Geneformer offers rank-value encoding over 30 million transcriptomes to perform computational in silico gene deletions.
For projects centred on predicting cellular perturbation responses, models diverge significantly in compute requirements and target modalities. Rhaister predicts responses to drugs, cytokines, and knockdowns by operating on aggregate summary statistics using alternating least squares and ridge regression, allowing it to train in seconds and infer in milliseconds; however, its accuracy drops substantially on CRISPR knockdowns and in zero-shot regimes. In contrast, X-Cell is a set-level diffusion language model built specifically for genome-scale CRISPR interference (CRISPRi) knockdowns, integrating multi-modal biological priors via cross-attention. For pharmacological screens, Tahoe-x1 provides transformer foundation models ranging from 70M to 3B parameters pre-trained on 266 million profiles, while Tahoe-100M-SCVI offers an efficient 40-million-parameter scVI variational autoencoder that requires only around 1 GB of GPU memory to host, though it is trained strictly on cancer cell lines under chemical treatments.
For multi-modal data integration or inputs beyond standard human single-cell RNA-seq, input format requirements dictate tool choice. SCENIC+ reconstructs enhancer-driven gene regulatory networks specifically by combining paired single-cell RNA-seq and single-cell ATAC-seq profiles. PhenoSeq bridges imaging and transcriptomics by using a conditional Gaussian diffusion model to translate 5,120-dimensional ViT-L multi-channel fluorescence microscopy features into 512-dimensional scGPT transcriptomic embeddings. For studies incorporating bulk, microarray, or mouse transcriptomes, EVA-RNA processes both human and mouse profiles into sample-level and gene-level embeddings using NCBI Gene IDs, whereas SE-600M, STACK Large, and Stack-Large Aligned remain restricted to human single-cell records.
Questions
Can I run these models without a GPU or installing anything?
0 of the 14 models here run on our servers from their page: open one, fill in the inputs and press Run. Those marked "Not yet" are listed with their licence and documentation but have no loader here yet.
Which models support organisms other than humans?
EVA-RNA natively accommodates both human and mouse transcriptomic profiles across bulk, microarray, and pseudobulked single-cell formats. TEDDY is also pre-trained on single-cell RNA-seq profiles spanning human and mouse tissues. In contrast, models such as SE-600M, STACK Large, and scGPT are explicitly restricted to human cells.
What are the hardware requirements for hosting these models?
Memory footprints vary widely. CellFM, Geneformer, and scGPT each require at least 40 GB of GPU memory to operate. Tahoe-100M-SCVI requires only about 1 GB of GPU memory to host and run inference, though loading its accompanying minified reference AnnData file demands 41 GB of storage and system RAM. Rhaister bypasses heavy GPU requirements by training in seconds on aggregate summary statistics.
Can I predict the effects of genetic knockdowns directly?
Yes, but approaches differ by perturbation mechanism. X-Cell is specifically built as a set-level diffusion model to predict single-cell transcriptomic responses to CRISPR interference (CRISPRi) knockdowns. Geneformer models gene knockouts computationally via in silico deletion within its rank-encoded network context. Rhaister also models knockdowns, though its performance on CRISPR knockdown benchmarks is notably lower than on small-molecule drug screens.
Which tools handle multi-modal data like imaging or chromatin accessibility?
SCENIC+ requires paired single-cell RNA-seq and single-cell ATAC-seq data to reconstruct enhancer-driven gene regulatory networks. PhenoSeq translates multi-channel fluorescence microscopy imaging features (specifically 5,120-dimensional ViT-L/14 embeddings across five channels) into single-cell scGPT gene expression embeddings.
Written from each model's own documentation, as summarised on its page here. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.