SNR-ST-Mix: Geometry-Aware Regression Augmentation

• SNR-ST-Mix is a data augmentation method for spatial transcriptomics regression that uses local neighborhood mixup constrained by spatial proximity and expression similarity.
• It enhances imputation by interpolating histology and gene expression data with KNN-based neighbor selection and symmetric beta sampling to reduce noise and preserve tissue structure.
• Empirical results across cancer and healthy tissue datasets demonstrate improved MSE, MAE, and PCC, confirming its effectiveness without requiring changes to the backbone model.

SNR-ST-Mix, introduced as Sample-specific Neighborhood Regression ST Mixup, is a data augmentation framework for spatial transcriptomics (ST) imputation that is designed for regression rather than classification. It targets the setting in which histology patches are used to predict continuous gene-expression vectors, and it addresses the mismatch between standard augmentation schemes and ST biology by constraining interpolation to a spot’s spatial neighborhood and biasing partner selection toward transcriptionally similar spots. In its reported formulation, SNR-ST-Mix is geometry- and expression-aware, requires no architectural change to the backbone model, and is intended to preserve local biological structure while improving generalization and prediction stability under sample-specific training (Yu et al., 7 Jun 2026).

1. Problem formulation and motivation

Spatial transcriptomics measures gene expression while preserving tissue coordinates, but the measurements are described as sparse, noisy, low-resolution, and heterogeneous across tissues and samples. These characteristics are particularly restrictive in sample-specific imputation, where a model must be learned from a single tissue section or a small number of sections. Under such conditions, augmentation becomes attractive, but the standard augmentation repertoire is misaligned with the target task (Yu et al., 7 Jun 2026).

The motivation for SNR-ST-Mix is rooted in the claim that most existing augmentation strategies are designed for classification tasks. In the ST setting, however, the prediction target is a continuous gene-expression vector, and naive interpolation can be biologically implausible. The reported critique is threefold. First, Mixup assumes categorical labels or simple interpolable targets, whereas ST labels are continuous. Second, CutMix and related image augmentations can produce unrealistic combinations when applied directly to ST. Third, random mixing can combine spots from distant tissue regions, distinct microenvironments, or transcriptionally dissimilar states, thereby generating artificial targets that violate local spatial biology. The paper positions C-Mixup and RC-Mixup as closer to the regression setting, but still insufficient because they do not explicitly encode the spatial organization of ST data (Yu et al., 7 Jun 2026).

The central design problem is therefore to construct augmentation samples that are simultaneously spatially local and expression-consistent. The method answers this by coupling two restrictions: spatial neighborhood selection and expression-aware weighting. This combination is presented as the core mechanism by which interpolation remains on a local biological manifold rather than drifting into implausible mixtures.

2. Data model and neighborhood construction

The data representation used by SNR-ST-Mix is

$\mathcal{D}=\{(x_i, y_i, c_i)\}_{i=1}^N$

where $x_i \in \mathbb{R}^{3 \times H \times W}$ is the histology patch centered at spot $i$, $y_i \in \mathbb{R}^{G}$ is the gene-expression vector over $G$ genes, and $c_i \in \mathbb{R}^{2}$ is the spatial coordinate of spot $i$. Spatial distance is defined by

$d_{ij} = \|c_i - c_j\|_2 .$

The first structural constraint is spatial-neighborhood restriction. For each anchor spot $i$, SNR-ST-Mix defines a $K$-nearest-neighbor set in coordinate space:

$x_i \in \mathbb{R}^{3 \times H \times W}$0

This moves mixup from a global to a local operation. The rationale is formalized using a smooth spatial field model

$x_i \in \mathbb{R}^{3 \times H \times W}$1

together with a Lipschitz-type prior

$x_i \in \mathbb{R}^{3 \times H \times W}$2

Under these assumptions, the expected expression discrepancy between two spots satisfies

$x_i \in \mathbb{R}^{3 \times H \times W}$3

The explicit implication in the formulation is that smaller spatial distance limits expected label mismatch and helps preserve local tissue structure (Yu et al., 7 Jun 2026).

The second structural constraint is expression-aware neighbor selection inside the spatial neighborhood. Expression distance is defined as

$x_i \in \mathbb{R}^{3 \times H \times W}$4

and similarity is computed by a Gaussian kernel,

$x_i \in \mathbb{R}^{3 \times H \times W}$5

where $x_i \in \mathbb{R}^{3 \times H \times W}$6 controls the decay rate. These similarities are normalized into a conditional sampling distribution,

$x_i \in \mathbb{R}^{3 \times H \times W}$7

This construction ensures that spatially adjacent but expressionally dissimilar spots remain available in principle, yet are downweighted in partner selection. The intended effect is to reduce mixing across sharp biological boundaries or heterogeneous regions without abandoning local spatial context.

3. Mixup construction and statistical rationale

Given an anchor spot $x_i \in \mathbb{R}^{3 \times H \times W}$8, SNR-ST-Mix samples a partner $x_i \in \mathbb{R}^{3 \times H \times W}$9 from the categorical distribution $i$0, with $i$1, and then samples a mixup coefficient from a symmetric beta law,

$i$2

In the reported experiments, $i$3, making $i$4 approximately uniform on $i$5 (Yu et al., 7 Jun 2026).

The mixed input patch and mixed target are defined by linear interpolation:

$i$6

and

$i$7

The method applies this interpolation jointly in image space and expression space. Because the partner spot is both spatially close and transcriptionally similar, the synthetic pair is intended to stay within a local biological regime rather than representing an arbitrary convex combination of unrelated tissue states (Yu et al., 7 Jun 2026).

The reported statistical argument emphasizes noise reduction. If spot noise is independent and zero-mean with covariances $i$8 and $i$9, then the covariance of the mixed label satisfies

$y_i \in \mathbb{R}^{G}$0

for $y_i \in \mathbb{R}^{G}$1. The stated interpretation is that expression-aware mixup reduces variance by averaging similar noisy labels, while keeping bias small because $y_i \in \mathbb{R}^{G}$2 is constrained to remain small. A plausible implication is that the method is particularly suited to heteroscedastic ST measurements, where variance suppression without strong manifold distortion is valuable.

4. Objective function, regularization, and training workflow

SNR-ST-Mix augments a regression predictor $y_i \in \mathbb{R}^{G}$3 with a composite objective containing four loss terms (Yu et al., 7 Jun 2026).

The primary term is the vicinal regression loss,

$y_i \in \mathbb{R}^{G}$4

This is the direct MSE objective on mixed examples.

The second term is the mixup consistency loss,

$y_i \in \mathbb{R}^{G}$5

Its role is to enforce approximate linearity of the predictor along the interpolation segment.

The third term is an edge alignment / graph smoothness loss,

$y_i \in \mathbb{R}^{G}$6

with $y_i \in \mathbb{R}^{G}$7. For similar neighboring spots, this behaves like a graph smoothness penalty. When true biological boundaries produce large expression differences, the alignment form permits correspondingly large prediction differences and therefore does not force indiscriminate smoothing.

The fourth term is a Pearson-correlation regularizer,

$y_i \in \mathbb{R}^{G}$8

The total objective is

$y_i \in \mathbb{R}^{G}$9

The reported weights are $G$0, $G$1, and $G$2 (Yu et al., 7 Jun 2026).

Algorithmically, the procedure consists of: computing $G$3 for each spot, converting expression similarities within that neighborhood into $G$4, sampling $G$5, sampling $G$6, forming $G$7, and optimizing the composite loss. The paper states that the method is architecture-agnostic and adds only lightweight preprocessing: KNN computation in coordinate space and expression-similarity weighting.

5. Backbone model, datasets, and empirical results

The reported implementation uses a Vision Transformer (ViT) pretrained on ImageNet. Input patches are resized to $G$8, the original classification head is removed, and a single linear projection is appended to predict the $G$9-dimensional gene vector. The model is fine-tuned end-to-end with dropout rate 0.2, AdamW, learning rate $c_i \in \mathbb{R}^{2}$0, $c_i \in \mathbb{R}^{2}$1, $c_i \in \mathbb{R}^{2}$2, weight decay $c_i \in \mathbb{R}^{2}$3, cosine annealing, and training for up to 200 epochs with early stopping; model selection is based on validation MSE (Yu et al., 7 Jun 2026).

The evaluation uses 10x Genomics Visium data from HEST-1K, with eight samples: TENX13 and TENX14 (breast cancer), MISC72 and MISC73 (bowel cancer), TENX65 (ovarian cancer), TENX46 (prostate cancer), and MISC128 and MISC129 (healthy heart). Preprocessing includes $c_i \in \mathbb{R}^{2}$4 patch extraction, ImageNet normalization, log-normalization of gene counts, filtering of genes and spots, and selection of the top 250 genes by mean expression. The split is 30% training, 20% validation, and 50% testing. The baseline set comprises no augmentation / baseline ViT, basic augmentation with random flips and random cropping, vanilla Mixup, and CutMix. Evaluation uses MSE, MAE, and PCC (Yu et al., 7 Jun 2026).

SNR-ST-Mix is reported to achieve the best overall performance across all eight datasets. Representative comparisons include the following. On TENX13, the baseline records MSE 0.2074, MAE 0.3081, PCC 0.4180; CutMix gives 0.1978 / 0.3051 / 0.4415; SNR-ST-Mix gives 0.1886 / 0.2911 / 0.4748. On TENX14, the baseline gives 0.1718 / 0.2940 / 0.4229, whereas SNR-ST-Mix gives 0.1463 / 0.2710 / 0.4802. On MISC72, the baseline is 0.2895 / 0.3789 / 0.4728, while SNR-ST-Mix reaches 0.2542 / 0.3436 / 0.4925. On MISC73, the baseline is 0.2456 / 0.3510 / 0.4798, and SNR-ST-Mix reaches 0.2227 / 0.3254 / 0.5064 (Yu et al., 7 Jun 2026).

Dataset	Tissue	SNR-ST-Mix (MSE / MAE / PCC)
TENX13	breast cancer	0.1886 / 0.2911 / 0.4748
TENX14	breast cancer	0.1463 / 0.2710 / 0.4802
MISC72	bowel cancer	0.2542 / 0.3436 / 0.4925
MISC73	bowel cancer	0.2227 / 0.3254 / 0.5064
TENX65	ovarian cancer	0.2045 / 0.3222 / 0.5568
TENX46	prostate cancer	0.1456 / 0.2851 / 0.3684
MISC128	healthy heart	0.4734 / 0.4867 / 0.4559
MISC129	healthy heart	0.5163 / 0.5268 / 0.4385

The general trend reported is that SNR-ST-Mix improves MSE and MAE in every dataset and achieves the highest PCC across all datasets. The gains are stated to be largest in more heterogeneous tumor tissues and smaller, though still consistent, in more homogeneous tissues such as healthy heart (Yu et al., 7 Jun 2026).

Beyond aggregate metrics, the method is reported to improve per-gene predictions. For representative genes including ISG15, MT-ND3, S100A6, and HLA-A, the method is described as better recovering high-expression regions, preserving sharp boundaries, suppressing noise in low-expression areas, and producing smoother and more coherent expression maps. In per-gene MSE scatter plots, almost all points lie below the identity line relative to baselines. The paper reports that in TENX13, 100% of genes improved over no-mixup, basic augmentation, and CutMix, while 99.2% improved over vanilla Mixup; similar near-universal improvement is stated for MISC72 and MISC129 (Yu et al., 7 Jun 2026).

Ablations are used to isolate the role of each component. In TENX13, the reported MSE values are 0.2025 for vanilla mixup, 0.1984 after adding KNN restriction, 0.1989 after adding label similarity, 0.1944 when both are combined, and 0.1886 for the full loss design. The interpretation presented is that spatial locality and expression similarity are complementary, and that the additional loss terms further improve linearity, spatial structure preservation, and PCC (Yu et al., 7 Jun 2026).

6. Interpretation, limitations, and conceptual position

The biological interpretation of SNR-ST-Mix is explicit. Spatial proximity is taken as a proxy for anatomical continuity and shared microenvironment, while expression similarity is used to ensure that mixed samples come from compatible molecular states. Together, these constraints are intended to preserve tissue domains, smooth gradients, and local boundaries. This is particularly relevant in ST because neighboring spots often share cell populations, yet tumor margins or niche boundaries can produce sharp transitions. SNR-ST-Mix is therefore positioned not merely as a regularization device, but as a biologically informed vicinal sampling strategy (Yu et al., 7 Jun 2026).

Several limitations are also stated. The gains vary with tissue heterogeneity, so performance improvement is more pronounced in tumors than in homogeneous tissues. The use of a fixed neighborhood size means that one value of $c_i \in \mathbb{R}^{2}$5 may not capture variable local heterogeneity across an entire section. The method is developed at Visium resolution, where spot-level aggregation can obscure sharp biological boundaries; the authors note that the framework may work even better at single-cell or subcellular resolution. Proposed extensions include region-level mixup, adaptation to CutMix/PuzzleMix-style methods, and integration with deconvolution or cell morphology cues (Yu et al., 7 Jun 2026).

A common misconception arises from the acronym itself. In this context, SNR does not denote signal-to-noise ratio; it denotes Sample-specific Neighborhood Regression. This distinguishes SNR-ST-Mix from unrelated SNR-based methods in other domains, such as speech enhancement approaches that explicitly control the signal-to-noise ratio of remixed pseudo-data through an SNR control module (Li et al., 2024). The ST method is instead a regression-specific augmentation strategy grounded in local spatial geometry and transcriptomic similarity.

Within the broader augmentation literature, SNR-ST-Mix can be situated as a specialization of mixup for structured biomedical regression. Its defining contribution is to replace global random interpolation with a constrained local interpolation scheme that is coupled to structure-aware regularization. The reported evidence indicates that this formulation expands the effective training manifold, improves predictive stability, and does so without increasing model complexity or requiring architectural redesign (Yu et al., 7 Jun 2026).