---
title: 'STARK: Adaptive Denoising in Spatial Transcriptomics'
url: https://www.emergentmind.com/topics/spatial-transcriptomics-via-adaptive-regularization-and-kernels-stark
type: topic
---

# STARK: Adaptive Denoising in Spatial Transcriptomics

Spatial Transcriptomics via Adaptive Regularization and Kernels (STARK) is a denoising methodology for spatial transcriptomics data, specifically designed to perform robust gene expression recovery in settings characterized by ultra-low sequencing depths. The approach synergistically combines kernel ridge regression with a dynamically adaptive graph Laplacian regularizer. This enables effective noise suppression, sharp boundary preservation between cell types, and supports gene expression interpolation at arbitrary spatial coordinates. The STARK framework delivers closed-form solutions for each alternating minimization subproblem, securing both algorithmic and statistical convergence guarantees. STARK’s performance, evaluated on real biological datasets, demonstrates superior label transfer accuracy and cell-type geometry preservation relative to contemporary graph-based and manifold-denoising techniques [2512.10994].

## 1. Motivation and Background

Spatial transcriptomics platforms (e.g. Stereo-seq, 10x Visium, Slide-seq) generate high-dimensional gene-expression vectors \(F^\star(q)\in\R^d\) at spatial locations \(q\in\Q\subset\R^2\). Although modern technologies increase the field of view, sequencing cost per pixel scales with the total number of reads. Operation within the ultra-low regime (\(\mathcal O(10^2)\) reads/pixel) precipitates severe multinomial or Poisson noise and dropouts, which results in very sparse observations \(Y_i\in\R^d\). The spatial sampling is irregular, undermining use of standard convolutional filters. Generic denoisers (e.g. Total Variation, non-local means) presuppose scalar images and regular grids; single-cell imputation techniques (MAGIC, SAVER, scImpute) neglect spatial structure; and manifold-learning methods such as SPROD build a one-shot graph from noisy data, often failing when input is highly corrupted.

A viable solution must integrate spatial smoothing for denoising, maintain sharp inter-cell-type boundaries, incrementally adapt the underlying image model, offer provable guarantees, and support interpolation for arbitrary \(q\).

## 2. Mathematical Formulation

Let observed pixels be \(\{q_i\}_{i=1}^m \subset \Q\) with measurements \(Y_i = F^\star(q_i) + V_i\), where \(V_i\) represents noise. STARK seeks a function \(F: \Q \to \R^d\) in a vector-valued RKHS \(\H^d\), alongside row-stochastic weights \(W \in \R_+^{m \times m}\) defined on a neighbor graph \(\E_\tau = \{(i,k) : \|q_i-q_k\|\le\tau\}\). The objective function \(J(F,W)\) incorporates:

- **Data-fit**: \(\frac{1}{m}\sum_{i=1}^m \|Y_i - F(q_i)\|^2\)
- **RKHS Ridge**: \(\lambda\|F\|_{\H^d}^2,\ \lambda > 0\)
- **Graph Laplacian regularizer**: \(\frac{\omega}{2m}\sum_{(i,k)\in \E_\tau} W_{ik} \|F(q_i)-F(q_k)\|^2\)
- **Entropy term for \(W\)**: steers regularization using spatial and gene-expression proximity

The joint minimization problem is block-convex and given by:
\[
\min_{F \in \H^d, W \in \C_\tau} J(F,W)
\]
where \(\C_\tau\) is the convex set ensuring neighborhood, normalization, and non-negativity constraints. The representer theorem permits reduction to finite-dimensional optimization. For \(\bK\) the kernel matrix and \(\Theta\) parameter coefficients,
\[
F(\cdot) = \frac{1}{\sqrt{m}}\sum_{i=1}^m \K(q_i, \cdot)\theta_i, \quad
\bF = \sqrt{m}\,\bK\Theta
\]
The augmented loss reads:
\[
J(F,W)= \frac{1}{m} \|\bY-\sqrt{m}\,\bK\Theta\|_F^2 + \lambda \Tr(\Theta^\top\bK\Theta)
+ \omega\Bigl\{\Tr\!\bigl[\Theta^\top\bK\bar L_W\bK\Theta\bigr]+E(W)\Bigr\}
\]
with \(\bar L_W\) denoting the Laplacian.

## 3. Incrementally Adaptive Graph Laplacian

At initialization, edges are purely spatial:
\[
\tilde W^{(0)}_{ik} = \exp(-\|q_i-q_k\|^2/s_2^2)\,1_{\{\|q_i-q_k\|\le\tau\}}, \quad
W^{(0)}=\mathrm{RowNormalize}(\tilde W^{(0)})
\]

For each iteration \(t\), given the current estimate \(F^t\),
\[
\tilde W^{(t+1)}_{ik} =
\exp(-\|F^t(q_i)-F^t(q_k)\|^2 / s_1^2) \exp(-\|q_i-q_k\|^2 / s_2^2) 1_{\{\|q_i-q_k\|\le\tau\}}
\]
Rows are normalized to sum to 1. This formulation is derived as the closed-form minimizer of the entropic graph Laplacian regularizer component, balancing spatial and expression-based affinity.

## 4. Alternating Minimization and Algorithmic Structure

STARK employs an alternating minimization (block coordinate descent) scheme:

1. **F-update (Kernel Ridge Regression)**:
   \[
   (\bK^2 + \lambda \bK + \omega \bK\bar L_{W^t}\bK) \Theta^{t+1} = \frac{1}{\sqrt{m}} \bK\bY, \quad \Theta \in \Range(\bK)
   \]
   The solution:
   \[
   \Theta^{t+1} = \frac{1}{\sqrt{m}}(\bK^2 + \lambda\bK + \omega\bK\bar L_{W^t}\bK)^+ \bK\bY
   \]

2. **W-update (Entropic Graph Fitting)**:
   \[
   \hat W^{t+1}_{ik} = \frac{\tilde W^{(t+1)}_{ik}}{\sum_\ell \tilde W^{(t+1)}_{i\ell}}
   \]

3. **Iteration**: Repeat for \(N\) steps; typically \(N=7\) suffice.

4. **Complexity**: Dominated by \(O(m^3)\) operations per iteration owing to matrix inversion or factorization. For \(m \approx 5,000\), this is tractable on modern hardware, with additional efficiency available via Laplacian sparsity and iterative linear solvers.

## 5. Theoretical Guarantees

Under modeled noise regimes and total reads \(R\), with regularization parameters \(\lambda=O(R^{-1})\), \(\omega=O(R^{-1})\):

- **Statistical Convergence**: Any stationary point \((\bar F, \bar W)\) of \(J\) achieves
  \[
  \mathbb{E}[\|\bar F - F^\star\|_{L^2_m}] = O(R^{-1/2})
  \]
  by comparison to oracle graph minimization and leveraging concentration for multinomial noise.

- **Algorithmic Convergence**: Alternating updates converge to a stationary point via block-convexity, unique block minimizers, and use of properties from coordinate descent theory (Kurdyka–Łojasiewicz property, Attouch–Bolte framework).

## 6. Empirical Evaluation and Benchmarking

Empirical assessment is conducted on the Mouse Organogenesis Spatiotemporal Transcriptomic Atlas (MOSTA) E9.5 snapshot via Stereo-seq, with 15,717 genes across 5,503 pixels (\(R_0\simeq 7.6\times 10^7\) reads).

**Evaluation Metrics**:
- **Label transfer accuracy**: PCA and \(k\)-NN on low-dimensional embeddings of cell-type labels (fraction correct).
- **kNN overlap**: Fractional match between directed kNN graphs on original and denoised data.
- **Relative error**: \(\|\bar\bF - \bF_0\|_F / \|\bF_0\|_F\).

Comparison is performed against SPROD (Wang et al.), GraphPCA (Yang et al.), and STAGATE (Dong et al.), each optimally tuned for label accuracy.

**Performance**:
- STARK exceeds competing methods for label transfer accuracy and kNN overlap, especially pronounced at ultra-low read depths (\(R/m \in [14, 200]\) reads/pixel).
- At \(R/m\approx 100\) reads/pixel: STARK \(\approx 0.85\) accuracy vs SPROD \(\approx 0.80\), GraphPCA \(\approx 0.74\), STAGATE \(\approx 0.70\).
- GraphPCA and STAGATE yield lower relative error but less effective geometry preservation of cell types.
- Interpolation capabilities to unseen locations are demonstrated through subsampling experiments.

**Visual Analysis**:
- Denoised cell-type spatial maps and variant comparisons illustrate performance (Figure 2, Figure 4 in source).

## 7. Practical Implementation Considerations

**Hyperparameter Selection**:
- Kernel length scale (\(l\)): set radius for ≈7 neighbors/pixel; \(s_2 = l, \tau = 1.5l\).
- Post-first iterate, \(s_1\) is set to the 75th percentile of pairwise gene-expression distances.
- Regularization: \(\lambda = \alpha\|\bK\|_{\text{op}}\), \(\omega = \alpha \cdot 6\|\bK\|_{\text{op}}\), \(\alpha\) optimized by matching residuals to read count statistics.
- Iterations: \(N=7\) is empirically sufficient.

**Validation**:
- Hyperparameters are optimized by downsampling real counts to validation sets for each read depth.

**Computational Resource Notes**:
- For sample sizes (\(m \approx 5,000\)), direct matrix resolution (\(O(m^3)\)) requires a few seconds on multi-core architectures; further efficiency possible through Laplacian sparsity and iterative solvers.

STARK’s synthesis of kernel ridge regression and incrementally adaptive Laplacian regularization establishes it as a practical, edge-aware denoising tool for sparse, noisy spatial transcriptomics data, supported by strong theoretical guarantees and demonstrably superior empirical performance in the ultra-low sequencing depth regime [2512.10994].

Source: https://www.emergentmind.com/topics/spatial-transcriptomics-via-adaptive-regularization-and-kernels-stark