CAPTAIN

CAPTAIN (ContrAstive Proteome and Transcriptome pre-trAINing) is a multimodal single-cell foundation model that learns a joint representation of the transcriptome and the cell surface proteome. Existing single-cell foundation models such as scGPT, scFoundation, and Geneformer are trained on RNA alone, so the cellular states they encode are inferred entirely from transcript abundance. Because proteins are the terminal effectors of cellular function — and because surface proteins in particular define immunophenotype, lineage identity, and therapeutic targetability — a transcriptome-only view systematically under-resolves the distinctions that immunologists and cell biologists care about most.

CAPTAIN was developed by Fulong Yu's group at Guangzhou Medical University and Guangzhou National Laboratory, with collaborators at Hunan University, Zhengzhou University, Central South University, and Southern Medical University. It was posted as a preprint in July 2025 and published in Nature Communications in May 2026. The model is pretrained on scT&P-4M, a corpus the authors assembled by harmonizing 252 publicly available co-assay samples into over 4.2 million human and mouse cells with matched RNA and a unified vocabulary of 382 cell surface proteins — 92% of it CITE-seq, the rest from compatible co-assay platforms.

The practical consequence is that CAPTAIN puts protein-level readouts on RNA-only data: having learned the transcriptome-to-proteome mapping during pretraining, it predicts abundance for surface markers that were never measured in a query dataset.

#Key Features

• Dual-encoder cross-attention architecture: RNA and protein modalities are encoded separately, then fused by an RNA-guided cross-attention block, so transcriptomic context directly conditions the prediction of each surface protein token.
• Zero-shot protein expansion: The model accurately predicts markers absent from any fine-tuning data — including CD21, CD79, CD4, and CD94 — recovering canonical B cell, CD4+ T cell, and NK cell patterns without paired protein measurements.
• Fine-grained cell type annotation: On a bone marrow CITE-seq dataset with 13 literature-defined T cell subtypes, CAPTAIN reached a macro-F1 of 0.73 against 0.61 for Seurat and 0.04 for scGPT, which collapsed nearly all T cells into a single naive CD4 label.
• Protein-informed cell-cell communication: Pairing transcript-derived ligand levels with imputed receptor abundance recovered 22 significant CD4+ naive T to NK cell interactions, 18 of them literature-supported, against six for the strongest transcriptome-only tool.
• Curated multimodal corpus: scT&P-4M resolves antibody nomenclature drift across studies into 382 standardized, functionally annotated surface proteins, and is released through a public data portal.

#Technical Details

CAPTAIN adopts a CLIP-style contrastive backbone. Its RNA encoder is a masked-attention transformer whose architecture and 19,264-gene vocabulary follow scGPT, initialized from scGPT weights pretrained on over 50 million single-cell profiles. The RNA pathway takes gene tokens, binned expression values, species tokens for cross-species transfer, and four channels of prior knowledge — co-expression, gene regulatory network, promoter sequence, and gene family embeddings — imported from GeneCompass in a shared 512-dimensional space. The protein encoder embeds 382 surface protein tokens plus a <cls> token, applies self-attention, then a cross-attention layer in which the RNA embedding supplies keys and values; the final <cls> state is the joint cell representation. Pretraining jointly optimizes masked gene expression regression, cross-modal protein abundance regression, and quantile regression over prediction intervals.

Across four held-out CITE-seq datasets spanning human PBMC, MALT, monocyte, and mouse PBMC panels, CAPTAIN achieved the highest protein-prediction accuracy by both Pearson correlation and RMSE against Seurat, sciPENN, and totalVI; in the zero-shot setting it still delivered the lowest RMSE, and on mouse PBMCs it beat every baseline without dataset-specific tuning. Cell type annotation on human PBMC CITE-seq reached 96.1% accuracy. For batch integration across three multi-batch cohorts, it scored an AvgBIO of 0.832 versus 0.725 for scGPT, 0.649 for Harmony, and 0.569 for Seurat.

#Applications

CAPTAIN is aimed at immunologists and single-cell analysts who want protein-level resolution without the reagent cost of a wide antibody panel. Typical uses include expanding surface marker profiles on existing scRNA-seq atlases, annotating closely related lymphocyte subsets that RNA alone cannot separate, harmonizing cohorts across CITE-seq, ECCITE-seq, and TEA-seq into one embedding space, and reconstructing receptor-aware signaling networks. Applied to a multi-sample COVID-19 cohort, it traced rising platelet-to-monocyte signaling with disease severity and severity-dependent activation of the pro-inflammatory S100A9-CD36 axis.

#Impact

CAPTAIN is among the first single-cell foundation models to treat co-assayed protein measurement as a pretraining signal rather than a downstream evaluation target, and its scT&P-4M corpus is a reusable resource independent of the model. The main constraints are scope and access: the protein vocabulary is fixed at the 382 surface markers present in public CITE-seq data, leaving intracellular proteins, chromatin, and metabolites outside the model, and the corpus skews toward immune tissue and human samples (85% human, 14% mouse). Code is MIT-licensed, but the CAPTAIN_Base and CAPTAIN_PBMC checkpoints are distributed through Google Drive links rather than a model hub, making programmatic access and version pinning harder than for its RNA-only predecessors.