CellPlex: Benchmarking Single-Cell Spatial Proteomics Prediction from H&E

Zijun Gao; Yijiang Chen; Xiangde Luo; Pheng-Ann Heng

cs.CVNov 2, 2026

CellPlex: Benchmarking Single-Cell Spatial Proteomics Prediction from H&E

Authors: Zijun Gao, Yijiang Chen, Xiangde Luo, Pheng-Ann Heng

Organizations: The Chinese University of Hong Kong · Stanford University School of Medicine

Abstract

Spatial proteomics measures tens of proteins in every cell of a tissue section, but its cost confines it to small research cohorts, whereas hematoxylin and eosin (H&E) slides exist for nearly every patient. Inferring single-cell spatial proteomics from H&E would bring its analyses to routine pathology, and a growing number of virtual staining models pursue this goal. Each is evaluated on its own cohort, panel and pixel-level metric, and released models transfer poorly to new cohorts, leaving open which approach to build on and which proteins H&E can reveal. We introduce CellPlex, a benchmark that poses the task at the level of the cell: seven cohorts with same-section H&E and multiplexed immunofluorescence, from colorectal and breast cancer to lymph-node metastases, bone marrow and glioma, with 717 samples, 44.8 million cells and 149 proteins, processed by one pipeline into per-cell targets. Pathology foundation models (PFMs) are adapted with one cell-level protocol, specialist models are fine-tuned under a matched budget, and all are scored on the same cells. Adapted PFMs are the most accurate models on every cohort, ahead of the specialists even after fine-tuning, but keep less than half of their accuracy on cohorts they were not trained on. Which proteins can be predicted depends on the tissue, not on how many cells were measured for them. By fixing the cells, targets and splits on which predictions are judged, CellPlex sets a common standard for inferring costly molecular measurements from routine H&E and shows how far H&E can stand in for spatial proteomics. We will release the data, splits, evaluation code and baselines as a testbed for developing and comparing future pathology foundation models and virtual staining methods.

Figures & tables

Figure 1: Overview of CellPlex. Left: the seven cohorts with their cells after quality control, patients, samples and proteins, and the construction of per-cell targets, in which nuclei segmented on the H&E are grown into cell regions and the registered multiplexed image is averaged over each region. Right: the evaluated models, pathology foundation models adapted with a cell-level protocol and two specialist models, Eva and GigaTIME, whose pixel-level outputs are averaged over the same cell regions.

	Orion-CRC		CRC-TMA		Multi-tumor		BC/CRC		HNSCC-LN		Bone marrow		Glioma		Avg
Model	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM
Specialist models
Eva, zero-shot	0.381	0.159	0.190	0.133	0.105	0.082	0.188	0.161	0.142	0.107	0.168	0.104	0.118	0.076	0.185	0.117
Eva, fine-tuned	0.500	0.322	0.491	0.340	0.419	0.298	0.456	0.397	0.441	0.295	0.548	0.411	0.365	0.239	0.460	0.329
GigaTIME, zero-shot	0.173	0.059	0.104	0.082	0.006	0.018	0.053	0.050	0.078	0.036	$-$ 0.055	$-$ 0.014	0.024	0.031	0.055	0.038
GigaTIME, fine-tuned	0.426	0.248	0.356	0.237	0.297	0.217	0.317	0.257	0.333	0.222	0.358	0.311	0.323	0.207	0.344	0.243

Table 1: Single-cell protein prediction on CellPlex. For each cohort, Pearson correlation (PCC) and cell-level structural similarity (SSIM) between predicted and measured protein levels. Each score is averaged over the proteins of a test sample and then over the test samples of a cohort; Avg: mean over the seven cohorts; standard errors over test samples are in Tab. S2 . Bold : best per column. Zero-shot GigaTIME is scored only on the panel proteins matching one of its 21 channels (7–14 per cohort).

Figure 2: Protein maps on held-out samples of three cohorts, two per cohort (four more cohorts in Fig. S3 ). Each case shows the H&E of one view and its cells colored by one protein, as measured and as predicted by CELLO (H-optimus-0) and by fine-tuned Eva and GigaTIME; the inset magnifies the boxed region. Every panel is scaled to the 1st–99th percentile of its cells in the view; scale bar, 100 $\mu$ m.

Figure 3: What makes a protein predictable, over the 341 cohort–protein pairs of the benchmark. Top: protein PCC against the spatial autocorrelation of the measured protein (Moran’s $I$ over the 10 nearest cells); lines: median and interquartile range per quintile of Moran’s $I$ ; $\rho$ : Spearman correlation. Bottom left: median protein PCC per quintile of sparsity (Gini coefficient), dynamic range (99th minus 50th percentile of the measured level) and Moran’s $I$ . Bottom right: protein PCC by subcellular localization; boxes: interquartile range and median, whiskers: 5th–95th percentile; $n$ : cohort–protein pairs.

Cohort	Tissue	Platform	Format	$\mu$ m/px	Samples	Cells	Proteins	Split unit
Orion-CRC [ 25 ]	Colorectal cancer	Orion	WSI	0.325	41	38.0 M	16	patient (40)
CRC-TMA [ 37 ]	Colorectal cancer	CODEX	TMA	0.377	130	234.5 K	56	core
BC/CRC [ 31 ]	Breast, colorectal cancer	CODEX	regions	0.507	45	516.9 K	36	region
Multi-tumor [ 37 ]	Multiple tumor types	CODEX	TMA	0.377	66	147.9 K	57	core
HNSCC-LN [ 14 ]	HNSCC lymph nodes	CODEX	TMA	0.377	388	1.47 M	57	patient (79)
Bone marrow [ 3 ]	Bone marrow	CODEX	WSI	0.249	11	772.4 K	52	patient (11)

Table S1: Cohorts of CellPlex after quality control. H&E and multiplexed immunofluorescence were acquired on the same tissue section in every cohort. Samples are whole-slide images (WSI), tissue-microarray cores or imaged regions; cells are nuclei segmented on the H&E. The total protein count merges synonymous marker names across panels.

PCC
Model	Orion-CRC	CRC-TMA	Multi-tumor	BC/CRC	HNSCC-LN	Bone marrow	Glioma
	( $n=8$ )	( $n=130$ )	( $n=66$ )	( $n=45$ )	( $n=388$ )	( $n=11$ )	( $n=9$ )
Eva, zero-shot	0.381 $\pm$ 0.019	0.190 $\pm$ 0.007	0.105 $\pm$ 0.010	0.188 $\pm$ 0.009	0.142 $\pm$ 0.003	0.168 $\pm$ 0.016	0.118 $\pm$ 0.013
Eva, fine-tuned	0.500 $\pm$ 0.026	0.491 $\pm$ 0.008	0.419 $\pm$ 0.011	0.456 $\pm$ 0.013	0.441 $\pm$ 0.003	0.548 $\pm$ 0.024	0.365 $\pm$ 0.021
GigaTIME, zero-shot	0.173 $\pm$ 0.022	0.104 $\pm$ 0.007	0.006 $\pm$ 0.010	0.053 $\pm$ 0.005	0.078 $\pm$ 0.003	$-$ 0.055 $\pm$ 0.011	0.024 $\pm$ 0.021
GigaTIME, fine-tuned	0.426 $\pm$ 0.019	0.356 $\pm$ 0.008	0.297 $\pm$ 0.012	0.317 $\pm$ 0.010	0.333 $\pm$ 0.003	0.358 $\pm$ 0.025	0.323 $\pm$ 0.020

Table S2: Mean

\pm

standard error over test samples of the PCC and SSIM in Tab. 1 . Each sample’s score is its mean over the proteins of the panel;

n

: test samples.

	vs. UNI2-h		vs. Virchow2		vs. Eva (FT)
Cohort	wins	$p$	wins	$p$	wins	$p$
Orion-CRC	7/8	0.016	8/8	0.008	8/8	0.008
CRC-TMA	117/130	$<\!10^{-19}$	123/130	$<\!10^{-21}$	116/130	$<\!10^{-16}$
BC/CRC	40/45	$<\!10^{-7}$	37/45	$<\!10^{-6}$	44/45	$<\!10^{-12}$
Multi-tumor	54/66	$<\!10^{-6}$	46/66	$<\!10^{-3}$	45/66	0.003
HNSCC-LN	330/388	$<\!10^{-39}$	367/388	$<\!10^{-58}$	340/388	$<\!10^{-51}$

Table S3: Paired comparison of CELLO (H-optimus-0) with the other PFMs and with fine-tuned Eva: number of test samples on which H-optimus-0 scores higher, and two-sided Wilcoxon signed-rank

p

-value over test samples.

Trained on	Multi-tumor	CRC-TMA	BC/CRC	HNSCC-LN	Bone marrow	Glioma	Orion-CRC
Multi-tumor	0.44	0.40 (0.53)	0.24 (0.55)	0.21 (0.47)	0.15 (0.59)	0.02 (0.48)	0.30 (0.60)
CRC-TMA	0.39 (0.44)	0.53	0.32 (0.54)	0.24 (0.47)	0.16 (0.59)	0.04 (0.46)	0.42 (0.60)
BC/CRC	0.17 (0.44)	0.31 (0.53)	0.53	0.26 (0.49)	0.28 (0.58)	0.08 (0.48)	0.47 (0.61)
HNSCC-LN	0.18 (0.44)	0.28 (0.53)	0.34 (0.53)	0.48	0.12 (0.59)	0.07 (0.47)	0.33 (0.60)
Bone marrow	0.14 (0.41)	0.19 (0.54)	0.21 (0.53)	0.17 (0.49)	0.62	0.15 (0.46)	0.26 (0.61)
Glioma	0.10 (0.42)	0.09 (0.53)	0.12 (0.52)	0.10 (0.48)	0.26 (0.60)	0.50	0.24 (0.60)

Table S4: Transfer of CELLO with H-optimus-0 across cohorts. Rows: training cohort, whose fold-0 model is applied without further training; columns: target cohort. Off-diagonal: PCC on the proteins shared by the two panels, with the target cohort’s own models on the same proteins and cells in parentheses. Diagonal (bold): the target cohort’s own models on its full panel ( Tab. 1 ).

Figure S1 : Per-protein accuracy across cohorts. Top: protein PCC of CELLO (H-optimus-0) for every protein measured in at least three cohorts, grouped into lineage, functional-state and structural proteins; bars: median over cohorts; hatched: not measured. Bottom left: Spearman correlation of protein PCC between pairs of cohorts over their shared proteins. Bottom right: share of the variance of protein PCC explained by protein, cohort and their interaction (46 proteins measured in at least three cohorts), and protein PCC of the other methods against CELLO; $\rho$ : Spearman correlation.

Figure S2 : Accuracy against the number of cells measured. Top: for each of the 149 proteins, the number of cells in which it was measured over all cohorts, and the PCC of CELLO (H-optimus-0), mean and range over cohorts. Bottom: mean protein PCC against cells measured for four methods; $\rho$ : Spearman correlation.

Figure S3 : Protein maps on held-out samples of the four cohorts not shown in Fig. 2 , displayed in the same way. Each case shows the H&E of one view and its cells colored by one protein, as measured and as predicted; the inset magnifies the boxed region; every panel is scaled to the 1st–99th percentile of its cells in the view; scale bar, 100 $\mu$ m.

	Eva				GigaTIME				Linear probe						CELLO
	zero-shot		fine-tuned		zero-shot		fine-tuned		UNI2-h		Virchow2		H-optimus-0		UNI2-h		Virchow2		H-optimus-0
Protein	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM
E-cadherin	0.687	0.293	0.809	0.490	–	–	0.769	0.436	0.866	0.666	0.859	0.702	0.866	0.676	0.858	0.654	0.846	0.653	0.864	0.689
Pan-CK	0.594	0.253	0.795	0.552	0.373	0.220	0.728	0.507	0.830	0.642	0.821	0.625	0.826	0.623	0.834	0.663	0.819	0.656	0.842	0.675
CD45	0.672	0.377	0.697	0.544	–	–	0.729	0.416	0.829	0.663	0.818	0.682	0.825	0.674	0.830	0.608	0.810	0.587	0.832	0.612
CD45RO	0.572	0.316	0.612	0.413	–	–	0.554	0.296	0.703	0.510	0.696	0.523	0.698	0.511	0.723	0.589	0.692	0.528	0.716	0.584

Table S5: Per-protein PCC and SSIM on Orion-CRC (16 proteins, 8 test samples): mean over test samples of the protein’s score across the cells of a sample. Proteins are sorted by the PCC of CELLO with H-optimus-0. Bold : best per row and metric; –: protein without a matching channel of zero-shot GigaTIME.

	Eva				GigaTIME				Linear probe						CELLO
	zero-shot		fine-tuned		zero-shot		fine-tuned		UNI2-h		Virchow2		H-optimus-0		UNI2-h		Virchow2		H-optimus-0
Protein	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM
CD45	0.427	0.339	0.605	0.458	–	–	0.474	0.218	0.652	0.604	0.636	0.567	0.653	0.599	0.699	0.676	0.699	0.678	0.709	0.688
$\beta$ -catenin	0.343	0.216	0.705	0.587	–	–	0.573	0.548	0.598	0.555	0.604	0.569	0.601	0.561	0.660	0.633	0.666	0.641	0.681	0.658
SMA	0.360	0.218	0.528	0.261	–	–	0.350	0.173	0.569	0.509	0.545	0.477	0.569	0.510	0.624	0.586	0.619	0.583	0.639	0.602
Na/K-ATPase	0.086	0.082	0.562	0.311	–	–	0.368	0.133	0.522	0.435	0.490	0.399	0.516	0.431	0.621	0.565	0.614	0.560	0.633	0.578

Table S6: Per-protein PCC and SSIM on CRC-TMA (56 proteins, 130 test samples), as in Tab. S5 .

	Eva				GigaTIME				Linear probe						CELLO
	zero-shot		fine-tuned		zero-shot		fine-tuned		UNI2-h		Virchow2		H-optimus-0		UNI2-h		Virchow2		H-optimus-0
Protein	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM
MMP9	0.026	0.030	0.577	0.247	–	–	0.346	0.148	0.496	0.346	0.518	0.351	0.516	0.360	0.576	0.485	0.588	0.502	0.602	0.518
CD123	0.238	0.167	0.685	0.578	–	–	0.595	0.570	0.469	0.452	0.504	0.482	0.491	0.469	0.555	0.541	0.578	0.559	0.586	0.575
CD20	$-$ 0.021	0.028	0.567	0.345	0.026	0.043	0.293	0.187	0.401	0.162	0.422	0.168	0.401	0.159	0.541	0.427	0.559	0.445	0.562	0.448
Granzyme B	0.072	0.058	0.528	0.267	–	–	0.313	0.150	0.466	0.308	0.478	0.313	0.482	0.319	0.538	0.439	0.546	0.451	0.562	0.469

Table S7: Per-protein PCC and SSIM on Multi-tumor (57 proteins, 66 test samples), as in Tab. S5 .

	Eva				GigaTIME				Linear probe						CELLO
	zero-shot		fine-tuned		zero-shot		fine-tuned		UNI2-h		Virchow2		H-optimus-0		UNI2-h		Virchow2		H-optimus-0
Protein	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM
Pan-CK	0.600	0.508	0.739	0.627	0.333	0.161	0.612	0.454	0.743	0.561	0.707	0.500	0.731	0.523	0.790	0.722	0.782	0.706	0.792	0.721
CD45	0.495	0.436	0.668	0.616	–	–	0.565	0.522	0.671	0.497	0.646	0.450	0.668	0.478	0.712	0.609	0.708	0.620	0.726	0.630
SMA	0.433	0.294	0.587	0.366	–	–	0.406	0.236	0.652	0.586	0.625	0.546	0.656	0.587	0.709	0.683	0.696	0.675	0.717	0.692
EPCAM	0.463	0.374	0.635	0.497	–	–	0.542	0.408	0.653	0.572	0.631	0.549	0.645	0.557	0.695	0.605	0.693	0.605	0.704	0.622

Table S8: Per-protein PCC and SSIM on BC/CRC (36 proteins, 45 test samples), as in Tab. S5 .

	Eva				GigaTIME				Linear probe						CELLO
	zero-shot		fine-tuned		zero-shot		fine-tuned		UNI2-h		Virchow2		H-optimus-0		UNI2-h		Virchow2		H-optimus-0
Protein	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM
CD30	0.231	0.174	0.683	0.457	–	–	0.560	0.310	0.506	0.477	0.534	0.508	0.522	0.490	0.661	0.634	0.663	0.638	0.674	0.648
HIF1A	$-$ 0.004	0.032	0.709	0.559	–	–	0.604	0.513	0.554	0.493	0.549	0.491	0.565	0.494	0.649	0.609	0.644	0.611	0.662	0.627
CD45	0.439	0.365	0.601	0.444	–	–	0.521	0.292	0.586	0.563	0.568	0.552	0.589	0.563	0.648	0.621	0.641	0.617	0.657	0.632
CD4	0.300	0.208	0.621	0.492	0.131	0.048	0.564	0.465	0.525	0.491	0.520	0.488	0.533	0.494	0.619	0.584	0.614	0.584	0.631	0.600

Table S9: Per-protein PCC and SSIM on HNSCC-LN (57 proteins, 388 test samples), as in Tab. S5 .

	Eva				GigaTIME				Linear probe						CELLO
	zero-shot		fine-tuned		zero-shot		fine-tuned		UNI2-h		Virchow2		H-optimus-0		UNI2-h		Virchow2		H-optimus-0
Protein	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM
CD11b	0.306	0.085	0.722	0.624	–	–	0.361	0.357	0.695	0.682	0.687	0.675	0.702	0.689	0.762	0.746	0.758	0.742	0.774	0.753
CXCR4	0.351	0.233	0.722	0.582	–	–	0.582	0.544	0.686	0.657	0.685	0.644	0.694	0.674	0.743	0.727	0.742	0.729	0.752	0.738
CD141	0.074	0.032	0.658	0.553	–	–	0.214	0.209	0.674	0.638	0.645	0.583	0.677	0.644	0.734	0.706	0.729	0.701	0.745	0.718
HIF1A	0.182	0.078	0.668	0.540	–	–	0.203	0.143	0.651	0.527	0.623	0.462	0.662	0.541	0.729	0.676	0.726	0.664	0.745	0.686

Table S10: Per-protein PCC and SSIM on bone marrow (52 proteins, 11 test samples), as in Tab. S5 .

	Eva				GigaTIME				Linear probe						CELLO
	zero-shot		fine-tuned		zero-shot		fine-tuned		UNI2-h		Virchow2		H-optimus-0		UNI2-h		Virchow2		H-optimus-0
Protein	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM	PCC	SSIM
HSPG2	0.255	0.093	0.715	0.195	–	–	0.507	0.135	0.753	0.681	0.724	0.594	0.754	0.667	0.832	0.817	0.822	0.809	0.835	0.813
CD31	0.243	0.031	0.596	0.233	–	–	0.482	0.185	0.643	0.625	0.630	0.596	0.649	0.632	0.723	0.718	0.706	0.696	0.727	0.724
SNAP25	0.021	0.037	0.531	0.314	–	–	0.396	0.265	0.615	0.606	0.624	0.622	0.609	0.597	0.702	0.674	0.638	0.628	0.703	0.688
FN1	0.232	0.126	0.554	0.154	–	–	0.443	0.113	0.612	0.564	0.604	0.540	0.619	0.575	0.688	0.666	0.692	0.674	0.699	0.679

Table S11: Per-protein PCC and SSIM on glioma (68 markers, 9 test samples), as in Tab. S5 ; DCX and DCXv2 are two markers of the panel for the same protein.

Virtual Staining

Virtual staining uses AI to predict histochemical or immunohistochemical appearances from unstained or differently stained tissue images.

See all 20 papers

spora: A Unified Multimodal Dataset and Platform for Spatial Proteomics
Oct 15, 2026Benedikt von Querfurth, Johann Wenckstern, Eeshaan Jain +9Multimodal Dataset Medical Image Benchmarks
EPFL · University Hospital Zurich · University of Zurich +2
CytoFormer: A Molecularly Supervised Cell Foundation Model for Histopathology Cell Classification
Aug 17, 2026Jialu Yao, Songhao Li, Alina Yu +1Histopathology Image Classification Pathology Foundation Models
Department of Pathology and Laboratory Medicine, University of Pennsylvania, Philadelphia, PA, USA · Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA, USA · Germantown Friends School, Philadelphia, PA, USA +1
Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images
Sep 29, 2026Zijun Gao, Chunbin Gu, Jinxi Xiang +2Spatial Transcriptomics Spatial Attention
The Chinese University of Hong Kong · Stanford University School of Medicine

Oct 15, 2026J/K move · Enter open · S save

spora: A Unified Multimodal Dataset and Platform for Spatial Proteomics

Benedikt von Querfurth, Johann Wenckstern, Eeshaan Jain, Lukas Klein+8

EPFL · University Hospital Zurich · University of Zurich+2

Multimodal Dataset Medical Image Benchmarks

Open paper AskPDF arXiv

Spatial proteomics (SP) captures the molecular composition and spatial organization of tissues at single-cell resolution, opening a direct view of the cellular ecosystems that drive disease. Protein panels, acquisition platforms, and tissue contexts differ from one cohort to the next, and a coherent picture of tissue biology can only emerge from models trained to bridge this heterogeneity. Foundation models offer the natural path forward, but their training depends on large, harmonized corpora that span platforms, panels, and cohorts. No public resource currently approaches that scale; available data sit in small, narrowly scoped releases with incompatible formats and conventions. We present

\texttt{spora}

, a large-scale multimodal SP dataset of 13,546 standardized multiplex images from 5,477 patients across 35 cohorts and five major SP technologies. Curated with biomedical experts,

\texttt{spora}

provides model training-ready formats together with cell and tissue segmentations, structured clinical metadata, and, where available, paired H&E enabling the training of virtual staining models. It is complemented by

\texttt{spora [io]}

, a unified interface for tiling, sampling, and standardization across all cohorts and modalities, and

\texttt{spora [bench]}

, a benchmark suite spanning cell phenotyping, diagnostic and therapeutic prediction as well as virtual staining tasks across twelve cohorts that places current SP foundation models on common ground against task specific baselines.

Abstract

Virtual Staining

Explore similar work

spora: A Unified Multimodal Dataset and Platform for Spatial Proteomics

CytoFormer: A Molecularly Supervised Cell Foundation Model for Histopathology Cell Classification

Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images