Virtual Multiplexed Immunostaining

Updated 9 July 2026

Virtual multiplexed immunostaining is a computational approach that digitally generates or restores multiple staining outputs from limited inputs like H&E, brightfield IHC, or label-free microscopy.
It employs advanced model architectures—such as diffusion models, conditional GANs, and attention U-Nets—to predict missing channels and preserve crucial cellular details.
The technique overcomes traditional IHC limitations by reducing reagent costs and conserving tissue through digital surrogate staining, improving downstream diagnostic analysis.

Virtual multiplexed immunostaining denotes a set of computational strategies that generate, restore, or infer multiplex immunostaining signals from more limited inputs, including hematoxylin and eosin (H&E), brightfield immunohistochemistry (IHC), label-free microscopy, or incomplete multiplex panels. In pathology and spatial biology, the concept addresses the fact that conventional IHC is limited by its single-marker approach, while multiplexed immunofluorescence (mIF), multiplex immunohistochemistry (mIHC), imaging mass cytometry (IMC), and related assays provide simultaneous visualization of multiple biomarkers but remain constrained by high reagent cost, multi-round staining protocols, specialized imaging platforms, and limited tissue availability (Omar et al., 2024). Current virtual approaches therefore target several closely related goals: generation of multiple IHC stains from a single H&E image, translation from brightfield IHC to mIF, restoration of missing channels in multiplex immunofluorescence, and prediction of arbitrary unmeasured marker subsets in high-dimensional multiplex imaging (Dubey et al., 2024, Lee et al., 17 Mar 2026, Bao et al., 2021, Możejko et al., 4 Feb 2026).

1. Conceptual scope and relation to physical multiplex imaging

Physical multiplex imaging methods were developed to capture molecular and spatial information that cannot be recovered from routine H&E alone. mIF uses panels of fluorophore-conjugated antibodies and advanced microscopy with spectral unmixing; cycling and stripping technologies such as CycIF and CODEX permit iterative staining and imaging for up to 40+ markers; IMC and MIBI use metal-tagged antibodies for ultra-high multiplexing without spectral overlap; and IBEX enables multiplexed imaging of more than 65 parameters, is compatible with over 250 commercially available antibodies and 16 unique fluorophores, and proceeds by iterative cycles of antibody labelling, imaging, and chemical bleaching that can be completed at relatively low cost in 2–5 days (Omar et al., 2024, Radtke et al., 2021).

Virtual multiplexed immunostaining does not replace the underlying biological assay in a literal sense. Rather, it constructs a digital surrogate of missing stains or missing channels. In one common formulation, a model generates multiple IHC stains from a single H&E-stained tissue section, thereby preserving tissue for subsequent molecular testing in small biopsies (Dubey et al., 2024). In another, a model synthesizes mIF channels from widely available brightfield IHC while attempting to preserve nuclei count, shape, and spatial arrangement because errors in those quantities directly affect endpoints such as the Ki67 proliferation index (Lee et al., 17 Mar 2026). A third formulation addresses the “missing stain” problem in MxIF imaging, where certain antibody rounds fail and absent channels must be reconstructed digitally rather than reacquired at further tissue cost (Bao et al., 2021).

This breadth of usage explains why the literature spans pathology image translation, multiplex channel completion, label-free stain synthesis, and feature learning for downstream cell phenotyping. A plausible implication is that “virtual multiplexed immunostaining” is better treated as a family of computational objectives than as a single narrowly defined algorithmic task.

2. Input modalities and task formulations

The field has diversified across several input-output regimes.

Task formulation	Input	Output
Virtual multiplex IHC	Single H&E image	Multiple IHC stains
Brightfield-to-fluorescence translation	Brightfield IHC	mIF channels
Missing-stain restoration	Incomplete MxIF panel	Missing marker channels
Label-free virtual multi-staining	Autofluorescence or PARS microscopy	H&E and multiple histochemical or IHC stains
Flexible marker reconstruction	Arbitrary measured IMC markers	Arbitrary target IMC markers

The H&E-to-IHC setting includes text-conditioned or prompt-guided systems trained on uniplex paired H&E and IHC images. VIMs generates multiple immunohistochemistry stains from a single H&E-stained tissue section using a large vision-language single-step diffusion model, CLIP-based prompt encoding, LoRA fine-tuning, and an adversarial training module, while requiring only uniplex paired data during training (Dubey et al., 2024). A marker-wise conditioned latent diffusion model instead represents each marker by a one-hot vector added to the time embedding of the U-Net, enabling marker-by-marker generation of up to 18 different marker types from a single H&E input using a single model (Oh et al., 20 Aug 2025). PGVMS also operates from H&E, but replaces generic prompting with a pathology-specialized visual LLM and adds explicit strategies for protein distribution consistency and spatial misalignment correction (Chen et al., 26 Feb 2026).

The brightfield-to-mIF setting is exemplified by Segmentation-before-Staining, which translates IHC to multiplex IF while conditioning the generator on a continuous nuclei probability map from a pretrained segmentation foundation model (Lee et al., 17 Mar 2026). In multiplex immunofluorescence itself, pixN2N-HD treats arbitrary missing stain scenarios across 11 markers through an “N-to-N” conditional GAN with a random gate mechanism, so a single model can cover all possible missing-stain permutations (Bao et al., 2021). In IMC, ImmuVis generalizes the problem further: real-world marker sets vary across studies, so it predicts missing markers from arbitrary measured subsets using marker-adaptive hyperconvolutions rather than a fixed channel definition (Możejko et al., 4 Feb 2026).

Label-free variants extend the idea beyond stained brightfield inputs. One framework uses four-channel autofluorescence microscopy images of unstained tissue and a digital staining matrix to generate virtual H&E, PanCK, and ERG on the same section for vascular invasion assessment (Zhang et al., 22 Aug 2025). Another uses dual-excitation Photon Absorption Remote Sensing microscopy with interlaced 266 nm and 355 nm excitation, then RegGAN-based image translation to generate H&E, Masson’s trichrome, PAS, and Jones’ silver from one scan (Tweel et al., 5 Sep 2025).

3. Model architectures and conditioning mechanisms

A recurring design question is how to inject stain identity, structural priors, and pathology semantics into a single generative model.

In cell-centric mIF analysis, a semi-supervised variational autoencoder was proposed for robust cell feature extraction directly from mIF image patches of size $48 \times 48 \times 9$ . The encoder maps each patch into a latent representation $z$ , while a classifier $C_\gamma$ operates only on a supervised latent subspace $z'$ , with the full latent used for reconstruction and the subspace used for phenotype prediction (Sandarenu et al., 2024). The objective was written as

$L(\theta, \phi, \gamma) \coloneqq - \sum_{(x, y)} \left[ \text{ELBO}(\theta, \phi; x) + \alpha \cdot \log \left( l_\text{CE}(y, C_\gamma(\mu_\phi(x)')) \right) \right].$

Although this model is not a stain generator, its stated motivation includes downstream virtual multiplexed immunostaining, because richer and more generalizable cell features support marker positivity inference and cell identity assignment when inputs are incomplete.

Diffusion-based virtual staining systems vary mainly in how marker identity is represented. VIMs uses a pre-trained latent diffusion model, a CLIP prompt encoder, LoRA adapters, skip connections, and single-step denoising. Its marker-specific generation equation is

$y_i = Q(x_i + z_t, c_i),$

where $Q$ is the denoising UNet and $c_i$ is the prompt for IHC marker $i$ (Dubey et al., 2024). By contrast, the marker-wise conditioned diffusion model avoids text for scalability, uses one-hot marker conditioning, doubles the model input channels to concatenate H&E and marker latents, and then performs a second fine-tuning stage with pixel-level $L_1 + L_2$ loss for single-step deterministic inference (Oh et al., 20 Aug 2025). The data suggest a practical trade-off: text conditioning offers linguistic flexibility, whereas marker-wise one-hot conditioning scales more cleanly as the number of target markers increases.

PGVMS combines both semantic specificity and explicit pathology constraints. It introduces Pathological Semantics-Style Guided generation using CONCH, a pathology-specialized vision-LLM trained on 1.17M IHC image–text pairs; Protein-Aware Learning Strategy, which constrains optical-density-derived protein distributions at global, histogram, and block levels; and Prototype-Consistent Learning Strategy, which aligns semantically corresponding protein regions across misaligned serial sections (Chen et al., 26 Feb 2026). This framework was presented specifically to address inadequate semantic guidance, inconsistent protein distribution, and spatial misalignment across stain modalities.

Structure-aware translation models add priors before generation rather than only regularizing outputs afterward. Segmentation-before-Staining concatenates a continuous soft cell probability map $z$ 0 with the RGB IHC input: $z$ 1 The same work adds a variance-preserving loss to match local second-order intensity statistics and preserve cell-level heterogeneity in synthesized fluorescence channels (Lee et al., 17 Mar 2026). Label-free virtual mIHC uses a conditional GAN with an Attention U-Net generator and a Digital Staining Matrix encoding the target stain as an extra channel, with $z$ 2 for ERG, $z$ 3 for PanCK, and $z$ 4 for H&E (Zhang et al., 22 Aug 2025). PARS-based virtual multi-staining similarly couples label-free acquisition with a registration-aware GAN because real and target images are not perfectly aligned after tissue processing (Tweel et al., 5 Sep 2025).

4. Fidelity criteria, supervision regimes, and evaluation

A central methodological divide concerns what counts as “fidelity.” Several papers argue that pixel-level realism alone is insufficient.

VIMs was trained on uniplex paired datasets in which each H&E image is paired with only one IHC stain, yet the same model generated both CDX2 and CK8/18. It used standard image metrics such as MSE, SSIM, and FID, biological metrics such as DICE, IoU, and Hausdorff distance on DAB-masked regions, downstream gland segmentation, and blinded review by two board-certified pathologists on 50 challenging cases (Dubey et al., 2024). The pathologists ranked VIMs top for overall image fidelity and false negative rates, while Pix2Pix performed slightly better in minimizing false positives for CDX2 only.

The marker-wise conditioned diffusion model was evaluated on HEMIT and Orion-CRC. On HEMIT, it reported average SSIM $z$ 5, average Pearson correlation $z$ 6, and average PSNR $z$ 7, outperforming pix2pix, pix2pixHD, HEMIT, and Marigold. On Orion-CRC, which contains 18 IF markers from 41 colorectal cancer whole-slide images and 170,000 patches, it reported average SSIM $z$ 8, average $z$ 9 $C_\gamma$ 0, and average PSNR $C_\gamma$ 1 (Oh et al., 20 Aug 2025). The same study also reported that text-based conditioning collapses as the number of markers increases, with Orion-CRC SSIM approximately $C_\gamma$ 2 under text conditioning, whereas one-hot conditioning remained strong and improved further after fine-tuning.

Segmentation-before-Staining explicitly reframed evaluation around clinical quantifiability. On DeepLIIF and HNSCC, adding the soft prior increased SSIM across all tested architectures; for example, Pix2Pix U-Net on DeepLIIF improved from $C_\gamma$ 3 to $C_\gamma$ 4, and in HNSCC the regression U-Net improved from no prior SSIM $C_\gamma$ 5 to binary prior $C_\gamma$ 6 and soft prior $C_\gamma$ 7 (Lee et al., 17 Mar 2026). The paper emphasized nuclei count fidelity, reduction of blurred or merged nuclei, and maintenance of Ki67 positivity error within clinically robust intervals. This directly challenges the misconception that visually plausible fluorescence synthesis is adequate if nuclear morphology is distorted.

PGVMS further pushed evaluation toward pathology-specific agreement. On MIST and IHC4BC, it reported the lowest integrated optical density errors and the highest Pearson correlation with true protein expression across stains. One example given was HER2 Pearson- $C_\gamma$ 8: $C_\gamma$ 9 for PGVMS versus $z'$ 0 for the best 1-to-1 method and $z'$ 1 for the best prior 1-to-many method (Chen et al., 26 Feb 2026). Mean subjective scores from pathologists exceeded $z'$ 2 across localization, heterogeneity, and intensity, and downstream classification on virtual IHC was comparable to ground truth.

In missing-stain reconstruction, pixN2N-HD formalized the combinatorial burden of channel absence. With 11 markers, $z'$ 3 possible combinations arise; training separate models was estimated at roughly four years of single-GPU time, whereas a single pixN2N-HD model required about 20 hours and showed no significant performance difference from task-specific baselines under a Wilcoxon signed-rank test (Bao et al., 2021). In label-free virtual mIHC, 85 PanCK and ERG image pairs were blindly evaluated by three board-certified pathologists, who found nearly perfect agreement between virtual and histochemical staining, and inference was reported as less than 2 seconds per $z'$ 4 patch (Zhang et al., 22 Aug 2025). For dual-excitation PARS virtual multi-staining, average diagnostic quality was $z'$ 5 for chemical stains and $z'$ 6 for virtual stains, and pathologists could not reliably distinguish real from virtual (Tweel et al., 5 Sep 2025).

5. Downstream analysis, toolchains, and multiplex image infrastructure

Virtual multiplexed immunostaining is typically only one stage in a larger analytical workflow. The broader multiplex imaging stack includes preprocessing, spectral unmixing, quality control, segmentation, feature extraction, phenotyping, graph construction, and slide-level visualization (Omar et al., 2024).

Feature extraction remains especially important because virtual staining quality is often operationalized through downstream cell-level tasks. In the semi-supervised VAE study, six cell phenotypes were classified from more than 44,000 curated and balanced mIF patches extracted across 1,093 tissue microarray cores, and the proposed semi-supervised VAE achieved accuracy $z'$ 7, precision $z'$ 8, and recall $z'$ 9, outperforming ResNet50 with PCA, a standard VAE, handcrafted morphological features from QuPath, and a semi-supervised autoencoder baseline (Sandarenu et al., 2024). The paper explicitly linked improved latent representations to better virtual multiplexed immunostaining through better cell type annotation.

Graph and MIL models operate on multiplex outputs after image synthesis or direct acquisition. Mew constructs a multiplex network with a Voronoi layer for geometric information and a Cell-type layer for capturing cell-wise homogeneity, then uses a scalable SIGN-inspired GNN with interpretable attention for patient-level prediction on large mIF graphs (Yun et al., 2024). A separate graph-based neural network for immune profiling combined tissue morphology and five-marker expression data in hierarchical cell-graph and tile-graph representations, exceeding 60% weighted F1-score for tumour staging at the region-of-interest level (Martin et al., 2022). Fluoroformer adapted multiple instance learning to multiplexed whole-slide images through scaled dot-product attention for channel fusion, reporting prognostic $L(\theta, \phi, \gamma) \coloneqq - \sum_{(x, y)} \left[ \text{ELBO}(\theta, \phi; x) + \alpha \cdot \log \left( l_\text{CE}(y, C_\gamma(\mu_\phi(x)')) \right) \right].$ 0-indices of $L(\theta, \phi, \gamma) \coloneqq - \sum_{(x, y)} \left[ \text{ELBO}(\theta, \phi; x) + \alpha \cdot \log \left( l_\text{CE}(y, C_\gamma(\mu_\phi(x)')) \right) \right].$ 1 with ResNet50 embeddings and $L(\theta, \phi, \gamma) \coloneqq - \sum_{(x, y)} \left[ \text{ELBO}(\theta, \phi; x) + \alpha \cdot \log \left( l_\text{CE}(y, C_\gamma(\mu_\phi(x)')) \right) \right].$ 2 with UNI embeddings on 434 non-small cell lung cancer samples, while its channel attention maps recapitulated immuno-oncological hallmarks such as CD8 and FOXP3 activity at the tumor margin (Harary et al., 2024).

Supporting infrastructure has expanded in parallel. PathML was described as an open-source Python package for scalable and modular analysis of multiplexed imaging, with tile stitching, inference APIs, a model zoo, and graph APIs (Omar et al., 2024). SpatialVisVR provides a VR interface for CODEX and related multiplexed images, supports manipulation of up to 100 protein channels, and couples VAE-based embeddings, dynamic time warping, and cosine similarity for contextual similar-patient search (Veerla et al., 2024). Synplex offers a synthetic simulator of highly multiplexed histological images with user-defined numbers of phenotypes, marker expression levels, morphology, neighborhood interactions, and technical artifacts, thereby supplying annotated data for training and benchmarking (Jiménez-Sánchez et al., 2021). ImmuVis extends this ecosystem to IMC by offering a foundation model pretrained on IMC17M—28 cohorts, 24,405 images, 265 markers, and over 17M patches—with self-supervised masked reconstruction, arbitrary marker subset handling, and calibrated per-pixel uncertainty via a heteroscedastic likelihood objective (Możejko et al., 4 Feb 2026).

6. Limitations, misconceptions, and emerging directions

Several misconceptions recur in the literature. One is that virtual multiplexing is simply a variant of image colorization. Recent work instead treats it as a pathology-constrained translation problem in which nuclei count, spatial arrangement, compartment specificity, optical density distributions, and cell-level heterogeneity are all clinically consequential (Lee et al., 17 Mar 2026, Chen et al., 26 Feb 2026). Another misconception is that multiplex-paired training data are always necessary; VIMs, for example, was trained solely on uniplex paired H&E and IHC images, and PGVMS also uses only uniplex training data (Dubey et al., 2024, Chen et al., 26 Feb 2026). A third is that generic prompting automatically scales to large marker panels; the 18-marker diffusion study reported strong degradation for text-based conditioning as marker count increased, motivating one-hot marker conditioning instead (Oh et al., 20 Aug 2025).

Limitations are also consistent across papers. Supervised virtual staining depends strongly on training data quality and alignment. The label-free virtual mIHC study noted that weak chemical staining, especially for ERG, limited ground-truth quality and dataset size, and emphasized that clinical validation is ongoing (Zhang et al., 22 Aug 2025). PGVMS identified residual challenges in encoding biomarker-specific spatial priors such as membrane or nuclear specificity (Chen et al., 26 Feb 2026). The one-shot color deconvolution CD-UNET work showed strong generalization across at least nine stains but also stated that further work is needed for entirely unseen staining and tissue types (Lahiani et al., 2018). Only a minority of methods currently provide explicit uncertainty estimates; ImmuVis is distinctive in making calibrated uncertainty a first-class output rather than an afterthought (Możejko et al., 4 Feb 2026).

Current research directions therefore converge on a few themes. One is pathology-native conditioning, using pathological visual LLMs rather than generic natural-image embeddings (Chen et al., 26 Feb 2026). Another is structure-aware synthesis, in which segmentation priors, registration-aware losses, and variance constraints are injected directly into translation pipelines (Lee et al., 17 Mar 2026, Tweel et al., 5 Sep 2025). A third is scalable unified modeling: single models for many stains, many markers, or arbitrary marker subsets, rather than one model per stain (Oh et al., 20 Aug 2025, Możejko et al., 4 Feb 2026). Taken together, these developments suggest that the field is moving from proof-of-concept virtual restaining toward integrated, quantitatively constrained, and deployment-oriented multiplex pathology systems.