henhaobio
Sign in

bioRxivCell BiologyPreprint

Generative machine learning unlocks the first proteome-wide image of human cells

Huangqingbo SuniD, Konstantin KahnertiD, Jan N. HanseniD, William D. LeineweberiD, Mingyang LiiD, Wanyue Feng, Frederic Ballllosera NavarroiD, Ulrika AxelssoniD, Wei OuyangiD, Emma K. LundbergiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

The spatial organization of proteins within cells governs virtually all cellular functions, yet current imaging can simultaneously visualize only tens of proteins, orders of magnitude below the thousands populating a single human cell. Here we present ProtiCelli, a deep generative model that simulates microscopy images for 12,800 human proteins from just three cellular landmark stains. Trained on 1.23 million Human Protein Atlas images, ProtiCelli outperforms existing methods in reconstruction accuracy and textural fidelity, and generalizes to unseen cell types and drug perturbations. Simulated images preserve hierarchical subcellular organization, recapitulate known protein protein interaction landscapes, and resolve compartment-specific functions of moonlighting proteins at single cell resolution. Remarkably, the model infers drug-induced changes in protein expression and localization from cell morphology alone, predicts cell cycle stage without dedicated markers, and enables unsupervised segmentation of subcellular compartments and spatial decomposition of gene sets into functional regions. We leverage ProtiCelli to generate Proteome2Cell, a dataset of 30.7 million simulated images spanning 2,400 virtual cells across 12 human cell lines, enabling hierarchical single-cell models that distinguish conserved from dynamic protein architectures. Integrated into the Human Protein Atlas, Proteome2Cell democratizes exploration of these virtual cells. By computationally bridging the experimental scalability gap, ProtiCelli establishes a foundation for spatial virtual cell modeling.

In-depth summary

Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 3 (10 Sept 2026) · CC BY 4.0

A conditional latent diffusion transformer predicts immunofluorescence images of 12,800 human proteins from nucleus, ER and microtubule stains, outperforming PUPS and CELL-Diff, including on an unseen drug-treated cell line.

Multiplexed imaging reaches on average 37 proteins per cell, far short of the thousands present, and earlier virtual-staining models (Statistical Cell, PUPS, CELL-Diff) are described as biased toward nuclear and cytosolic proteins and judged mostly on technical benchmarks. ProtiCelli swaps sequence-embedding conditioning for learned per-protein embeddings to gain fidelity and adds biology-based tests (co-localization, drug response, cell-cycle staging).

Key results

  1. ShownCompared with PUPS and CELL-Diff on held-out HPA cells, ProtiCelli matched or beat them on pixel-level and texture metrics and captured a cell-line-specific localization they missed; vesicular compartments stayed poorly predicted by every model.

    On HPA-testing (80,000 images), Pearson r and mutual information were higher or equal for every organelle class except peroxisomes, endosomes and lysosomes; Haralick and MS-SSIM texture scores were best; only ProtiCelli placed EGFR at the Golgi in SiHa cells. Ablation: protein-sequence conditioning raised FID from 6.46 to 38.27. · ProtiCelli design; Subcellular patterns and cell topology

  2. ShownGiven no cell-line or drug label, the model transferred to an independently acquired triple-negative breast cancer line (MDA-MB-468) and recovered many drug-induced changes in protein intensity and radial distribution from landmark morphology alone.

    CM4AI-MDA (112,410 cells, 453 proteins; Paclitaxel, Vorinostat or untreated): kept its lead over PUPS and CELL-Diff. Experimental vs simulated Cliff's δ for radial redistribution, Spearman ρ = 0.64 (Paclitaxel) and 0.47 (Vorinostat); average precision for top-10% responders 0.60 and 0.55 vs chance 0.10; top decile ρ = 0.76, 92% directional agreement. Missed BPTF's move to the plasma membrane. · Unseen cell line and cell states; Drug-response prediction; Profiling drug-induced protein spatial responses

  3. ShownSimulated Proteome2Cell images resemble real HPA images in SubCell embedding space, carry protein-specific localization signal, reproduce the real proteome's cluster hierarchy relative to a noise ceiling, and place STRING interaction partners closer than localization-matched pairs.

    59-66% of a protein's 15 nearest neighbours share its source (50% if fully mixed); simulated vs real label transfer top-1 45.2% vs 50.5%, macro F1 0.17 vs 0.25, morphology-only F1 0.04; hierarchy AMI 0.50 falling to 0.32 at 86 clusters vs real-vs-real 0.78 to 0.55 (57-64% of ceiling); 309,411 STRING pairs (score >850) more correlated than matched null in all 12 lines. · Hierarchical protein organization; Fidelity across cellular hierarchies

  4. ShownAn ensemble-agreement reliability score (20 samples per cell and protein) indicates how far predicted positions can be trusted: high for nuclear structures and microtubules, low for vesicles, tracking how well the three landmarks determine each compartment.

    Reliability vs an independent information-theoretic determinability score computed on real images: Spearman ρ = 0.827. Across HPA-testing, reliability correlated with pixel Pearson r (r = 0.84) more than with MI (0.56), MS-SSIM (0.47), Haralick (0.24) or PSNR (0.10). · Reliability-aware staining simulations; Reliability-aware prediction by ensemble medoid…; Reference-channel determinability

  5. SuggestedThe base model only partly captured cell-cycle marker dynamics, but brief fine-tuning on a small new FUCCI dataset recovered the expected stage-dependent CDT1 and GMNN trends in an in-distribution and an out-of-distribution cell line.

    Predicted vs measured intensity for 12 cell-cycle-regulated proteins: Pearson 0.587 (p < 0.001). Base model gave non-monotonic CDT1 in A-549 and flat markers in MDA-MB-468; after fine-tuning on 100 images per line (2,245 cells) for 8,000 iterations, CDT1 fell and GMNN rose significantly from G1 to G2 on held-out images (A-549 222, MDA-MB-468 152). · Cell-state prediction; Figure legends

Methods, data and reproducibility

A conditional latent diffusion model (Stable Diffusion 3.5 VAE, DiT-L/2 transformer, EDM training) generates a 512x512 protein channel from nucleus, ER and microtubule stains plus learned protein and cell-line embeddings, with the cell-line label dropped at probability 0.2. Outputs were assessed with image metrics, SubCell embeddings, an ensemble reliability score and biological tasks.

  • Human Protein Atlas Subcellular section v24 single-cell IF crops, 12,800 protein entries, 39 cell lines (HPA-training 1.15 million; HPA-testing 80,000) · human · n = approximately 1.23 million single-cell images
  • CM4AI-MDA: confocal IF of MDA-MB-468 cells treated with Paclitaxel, Vorinostat or untreated, 453 proteins (CM4AI Beta release, June 2025) · human · n = 112,410 cropped cells · doi:10.18130/V3/F3TD5R
  • FUCCI-CellCycle (new): 3D confocal stacks with CDT1, GMNN, DAPI, ER and microtubule stains in A-549 and MDA-MB-468 · human · n = 574 images (322 A-549, 252 MDA-MB-468)
  • New 3D confocal stacks of MDA-MB-468 landmark stains for subcellular segmentation · human
  • Proteome2Cell (simulated): 12,800-plex images, 200 cells per line across 12 cell lines · human (cell lines) · n = 30,720,000 images; 2,400 cells
  • High-confidence STRING interactions (score >850) used to test co-localization · human (pairs among the model's human protein entries) · n = 309,411 interactions
  • Protein assemblies in U2OS derived from integrated microscopy and biophysical interaction data · human · n = 275 assemblies
  • Reactome Metabolism of lipids gene set, simulated in Hep-G2 · human · n = 758 proteins (566 simulated) · R-HSA-556833
Methods & tools
ProtiCelli, Stable Diffusion 3.5 VAE, DiT-L/2 (Diffusion Transformer), EDM (Elucidating Diffusion Models), SubCell (MAE-CellS-ProtS-Pool), PUPS, CELL-Diff, STRING, g:Profiler, Reactome ContentService API, HPACellSegmentatorPortable, ScanPy (Leiden clustering)
Software & environment
Python 3.10; NumPy 1.24, SciPy 1.11, PyTorch 2.6.0, torchvision 0.21.0, diffusers 0.36.0, scikit-learn 1.8.0, scikit-image 0.26.0, ScanPy 1.10.2, umap-learn 0.5. No container or requirements file stated.
Assay conditions & compute
Training: one million iterations on 8 NVIDIA H100 GPUs, per-GPU batch size 32 (effective 256); inference uses 50 Euler denoising steps. Training time and memory not stated.
Reproducibility gaps
FUCCI-CellCycle, the 3D segmentation images and Proteome2Cell are only promised for bioimage archive deposit before publication; the HPA v26 virtual-cell interface is slated for Nov 2026; availability of trained weights is not stated.
Limitations (6)
  • Proteins are encoded with learned per-protein embeddings, so the model covers only its 12,800-entry vocabulary and cannot predict unseen proteins or transfer across species. · stated by the authors
  • Training used only standard cell lines in normal conditions; relocalizations absent from HPA (e.g. BPTF moving to the plasma membrane under Paclitaxel) cannot be predicted. · stated by the authors
  • Positional accuracy depends on how strongly nucleus, ER and microtubule stains constrain a compartment; vesicular structures (peroxisomes, endosomes, lysosomes) are poorly positioned, and the segmentation method still needs microtubule and ER references. · stated by the authors
  • Limited cell states per protein in training restrict single-cell variability; cell-cycle marker dynamics required task-specific fine-tuning. Co-localization in simulated images is a prerequisite for, not proof of, interaction. · stated by the authors
  • Out-of-distribution and drug-response tests rest on one cell line (MDA-MB-468), two drugs and the authors' own CM4AI dataset. · noted in this summary
  • Hypotheses drawn from simulated images (a nuclear mRNA-processing role for P4HA2, sub-pathway localization of lipid metabolism in Hep-G2) are single illustrative cases not validated experimentally here. · noted in this summary

Read this if you build or use virtual-staining or image-based virtual cell models, or analyse HPA subcellular data, and want proteome-scale simulated images with code, a per-prediction reliability score and biology-based evaluations.

Citations
1

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISGupta '24Dihan '24Kambhampati '26Khwaja '22Zhang '26Wang '24
68% semantic matchCell Biology
Ankit Gupta, Zoe Wefers, Konstantin Kahnert+10 more·bioRxiv·Dec 6, 2024· 30 citations