Pseudo-Visium SP Benchmark

• Pseudo-Visium SP is a virtual Visium-like dataset derived from high-resolution imaging proteomics that enables continuous protein field reconstruction from sparse seq-SP measurements.
• It is constructed from glioblastoma CODEX data using a hexagonal grid and circular ROI integration with simulated capture efficiency to mimic 10x Genomics sampling.
• The benchmark provides a standardized evaluation protocol for methods like Neural Proteomics Fields, demonstrating improved performance through rigorous cross-validation using MSE and PCC metrics.

Pseudo-Visium SP is an open-source benchmark dataset and evaluation protocol for spatial super-resolution in sequencing-based spatial proteomics (seq-SP). It was introduced together with Neural Proteomics Fields (NPF) to address a specific methodological gap: real seq-SP data are typically sampled at sparse spot locations and therefore do not provide dense ground truth for rigorous evaluation of super-resolved prediction at unsampled positions. Pseudo-Visium SP resolves this by constructing a virtual Visium-like dataset from high-resolution multiplexed imaging proteomics, yielding paired low-resolution spot measurements and dense target protein maps suitable for continuous-space protein reconstruction (Zhao et al., 24 Aug 2025).

1. Definition and benchmark role

Pseudo-Visium SP is a “pseudo” or virtual Visium-like dataset derived from high-resolution multiplexed imaging proteomics data and designed specifically for the new task of spatial super-resolution in seq-SP. The supported task is defined as predicting protein expression at unsampled locations from sparse spot measurements and histology, with a continuous-space rather than discrete-grid-only formulation (Zhao et al., 24 Aug 2025).

The dataset exists because dense evaluation is otherwise unavailable in native seq-SP settings. Real sequencing-based spatial proteomics are sparse by design, which makes fair comparison of super-resolution methods difficult, particularly when the target is prediction away from observed spots. The benchmark therefore functions as the core evaluation environment for methods that attempt continuous protein field reconstruction, rather than merely local interpolation.

A common misunderstanding is to treat Pseudo-Visium SP as a generic spatial omics benchmark. The paper instead presents it as a benchmark tightly coupled to seq-SP super-resolution and introduced to fill a field-level gap: existing spatial transcriptomics benchmarks are not directly transferable because protein expression is described as more tissue-specific and more variable across organs, donors, and conditions than gene expression. The paper further states that methods used in ST, such as STNet and istar, do not directly solve seq-SP protein reconstruction, and that standardized open datasets and cross-validation protocols had been lacking (Zhao et al., 24 Aug 2025).

2. Data provenance and construction pipeline

Pseudo-Visium SP is derived from publicly available glioblastoma CODEX data from Greenwald et al. (sample = 12). The CODEX data include 40 imaging channels, each representing a protein distribution across the tissue. Construction proceeds by converting these imaging proteomics measurements into simulated spot-based observations that mimic 10x Genomics Visium sampling geometry (Zhao et al., 24 Aug 2025).

Component	Specification
Source data	Publicly available glioblastoma CODEX data from Greenwald et al. (sample = 12)
Imaging channels	40
Grid geometry	Hexagonal grid with 100 μm inter-spot spacing
Quantification ROI	Circular region of interest with radius 55 μm
Capture efficiency	$\eta = 0.1$
Registration	H&E rigidly aligned to CODEX maximum-intensity projection using ANTs

The pipeline is explicitly described in four stages. First, H&E-stained images are rigidly aligned to the CODEX maximum-intensity projection using ANTs registration. Second, virtual spots are generated by placing a hexagonal grid with 100 μm inter-spot spacing across the whole tissue section, mimicking the 10x Genomics sampling geometry. Third, protein expression for each virtual spot is quantified by integrating per-channel intensities inside a circular region of interest of radius 55 μm around the spot center, and the result is scaled by a simulated capture efficiency:

$$\hat{P}_i = \eta \sum I(x,y), \quad x,y \in ROI_i,\ \eta = 0.1.$$

Finally, the authors create three systematic spatial shifts, which produce four mutually exclusive datasets used for 4-fold cross-validation (Zhao et al., 24 Aug 2025).

The resulting dataset is synthetic in sampling geometry but biologically grounded in underlying protein distributions. This suggests that its main epistemic role is not to replace native seq-SP collections, but to provide a controlled benchmark where dense targets are available and continuous-space super-resolution can be quantified.

3. Task formulation and evaluation protocol

The benchmark formalizes seq-SP super-resolution as continuous protein regression. Let $I \in \mathbb{R}^{3\times H\times W}$ denote the spot image from the whole-slide image, and let

$P = \{p_1,p_2,\dots,p_k\}$

be the observed protein expression vector at a spot, where $k$ is the number of proteins. The model takes the normalized spatial coordinate $(x,y)$ of a spot and the corresponding image patch as input, and predicts

$P' = \mathrm{NPF}(I,[x,y] \mid \theta).$

Training minimizes mean squared error between predicted and observed protein vectors:

$\min_\theta \mathrm{MSE}(P', P).$

All protein expression values are log-transformed before training, and input image patches are $224\times224$ pixels centered at normalized spot coordinates (Zhao et al., 24 Aug 2025).

The benchmark table organizes the dataset into six subject-based subsets. These subsets collectively cover 12 samples, with per-subset sample counts of 1, 1, 2, 2, 3, and 3. For each sample, evaluation uses 70% of spots in one subset for training, 30% for validation, and the remaining three subsets for testing. The paper states that the benchmark is therefore designed to test generalization across held-out spatial regions and across samples, not merely interpolation within a single dense map (Zhao et al., 24 Aug 2025).

Evaluation uses two main metrics: mean squared error (MSE, lower is better) and Pearson correlation coefficient (PCC, higher is better). This metric choice follows directly from the fact that the task is continuous protein regression rather than classification. Training uses Adam with an initial learning rate of 0.001, cosine annealing, linear warmup from $10^{-6}$ to epoch 5, batch size 32, and 100 epochs (Zhao et al., 24 Aug 2025).

4. Relation to Neural Proteomics Fields

Pseudo-Visium SP is tightly coupled to Neural Proteomics Fields. The paper states that the benchmark exists specifically to evaluate NPF and related methods, and that NPF is the paper’s first deep learning model for seq-SP super-resolution. NPF is designed per tissue slice to capture tissue-specific protein distributions and morphology, a design choice motivated by inter-tissue variability and tissue-specific morphology–protein relationships (Zhao et al., 24 Aug 2025).

NPF comprises two major branches. The Spatial Modeling Module (SMM) learns tissue-specific protein spatial distributions by encoding coordinates into continuous protein fields. The Morphology Modeling Module (MMM) extracts tissue-specific morphological features and fuses features from a frozen pathology foundation model, UNI, with a tissue-specific feature extractor (TSFE) using cross-attention. The continuous-space super-resolution idea is borrowed conceptually from NeRF: instead of learning a mapping only on discrete sampled spots, NPF learns an implicit function over spatial coordinates so that protein expression can be reconstructed at arbitrary unsampled positions in tissue (Zhao et al., 24 Aug 2025).

Within this formulation, the SMM uses a frequency-aware positional encoding

$$\hat{P}_i = \eta \sum I(x,y), \quad x,y \in ROI_i,\ \eta = 0.1.$$0

with $$\hat{P}_i = \eta \sum I(x,y), \quad x,y \in ROI_i,\ \eta = 0.1.$$1, and then refines the coordinate embedding through $$\hat{P}_i = \eta \sum I(x,y), \quad x,y \in ROI_i,\ \eta = 0.1.$$2 residual MLP blocks into an $$\hat{P}_i = \eta \sum I(x,y), \quad x,y \in ROI_i,\ \eta = 0.1.$$3-dimensional latent representation. The paper identifies this as the mechanism by which “super-resolution in continuous space” is implemented (Zhao et al., 24 Aug 2025).

A plausible implication is that the benchmark and the model are methodologically co-designed. Pseudo-Visium SP supplies the dense supervision needed to make continuous protein-field reconstruction measurable, while NPF supplies an architecture whose inductive biases explicitly target that benchmarked problem.

5. Comparative performance and ablation evidence

On Pseudo-Visium SP, NPF is reported as the best-performing method in mean results across 12 samples. It achieves MSE 0.1590 and PCC 0.8748, compared with the strongest listed non-NPF baseline, ResNet50, at MSE 0.2142 and PCC 0.8371. The paper summarizes this as at least a 3.8% PCC improvement and an MSE reduction of 0.06 over the baselines. Across individual subsets, the gains are also reported as consistent: in the MGH258 subset, NPF reaches MSE 0.1271 and PCC 0.8789; in ZH1041, MSE 0.1790 and PCC 0.8746; and in ZH916, MSE 0.1628 and PCC 0.9042 (Zhao et al., 24 Aug 2025).

The benchmark compares NPF against several baseline families: interpolation methods (Nearest, KNN), image-feature-based predictors (ResNet50, ViT-B, Swin-T), and ST-inspired predictors (istar, STNet). The table also reports learnable parameter counts, and the paper presents this as part of a fairness narrative: NPF achieves better performance with a parameter count close to ResNet50 despite explicit spatial modeling and a dual-branch architecture (Zhao et al., 24 Aug 2025).

The ablation studies are central to the benchmark’s interpretive value. Adding SMM to different image backbones improves performance consistently on Pseudo-Visium SP: ResNet50 improves from MSE 0.2142/PCC 0.8371 to 0.1826/0.8547, Swin-T from 0.2265/0.8303 to 0.1904/0.8555, and ViT-B from 0.4056/0.6952 to 0.2261/0.8238. The architecture ablation further shows that the full NPF model is best overall, with MSE 0.1590 and PCC 0.8748, outperforming single-module and partial dual-module variants. The paper notes that TSFE is particularly important, while also concluding that the full combination of SMM, TSFE, and UNI yields the best results (Zhao et al., 24 Aug 2025).

These results support two benchmark-specific claims. First, the benchmark is discriminative: it separates methods with and without explicit coordinate modeling. Second, spatial modeling is not merely a generic add-on, but materially improves performance when paired with morphology.

6. Interpretation, external validation, and limitations

The paper validates the framework on real 10x Visium spatial proteomics data from human tonsil using two publicly available datasets, “Human Tonsil” and “Human Tonsil Add-on Antibodies,” with a 7:1:2 train/validation/test split. On Human Tonsil, NPF achieves MSE 0.0648 and PCC 0.8125, versus ResNet50’s 0.1241/0.7247 and STNet’s 0.1168/0.7316. On the add-on antibodies dataset, NPF reaches 0.1184/0.7104, versus 0.1588/0.6007 and 0.1572/0.6140. The paper presents these results to show that the framework and benchmark are not limited to the pseudo-data setting (Zhao et al., 24 Aug 2025).

At the same time, Pseudo-Visium SP remains a simulated benchmark. Its sparse inputs are generated from multiplexed imaging proteomics rather than observed directly from a sequencing assay. A common misconception is therefore to interpret it as native seq-SP ground truth. The paper instead characterizes it as a synthetic but biologically grounded dataset. That distinction matters: the benchmark is intended to enable standardized dense evaluation under controlled conditions, not to erase assay-specific differences between CODEX-derived virtual spots and real seq-SP measurements.

Its main significance lies in enabling a central technical claim to be tested quantitatively: protein expression can be reconstructed as a continuous spatial field from sparse coordinates and tissue images. In that sense, Pseudo-Visium SP is not auxiliary to NPF; it is the benchmark protocol that makes continuous-space seq-SP super-resolution operationally measurable. A plausible implication is that future work on seq-SP super-resolution, especially work emphasizing held-out spatial generalization rather than within-grid interpolation, will require datasets and protocols of this type even when alternative model classes are used (Zhao et al., 24 Aug 2025).