scProfiterole: Clustering of Single-Cell Proteomic DataUsing Graph Contrastive Learning via Spectral Filters

The paper introduces scProfiterole, a computational framework that leverages Arnoldi orthonormalization to implement polynomial interpolations of spectral graph filters within a graph contrastive learning paradigm, thereby enhancing the robustness and accuracy of cell type identification in noisy single-cell proteomic data.

The Gist

The Big Picture: The "Protein Puzzle"

Imagine you are trying to sort a massive pile of mixed-up LEGO bricks into their correct sets. In the world of biology, these "bricks" are cells, and the "colors" on the bricks are proteins (the tiny machines that make cells work).

For a long time, scientists could only look at the instructions for making these proteins (RNA). But looking at the instruction manual doesn't always tell you what the final machine actually looks like or how it's behaving. Now, we have new technology that lets us look directly at the proteins themselves (Single-Cell Proteomics).

The Problem:
Looking at these proteins is like trying to sort the LEGO bricks in a dark room where the lights flicker, some bricks are missing, and the colors are blurry.

• Missing Data: Some proteins are hard to detect, leaving gaps in the picture.
• Noise: The measurements are shaky and full of static.
• The "Over-Smoothing" Trap: If you try to use standard computer tools to sort these cells, they tend to "blur" everything together. It's like using a heavy-handed blender: you mix the red bricks and blue bricks so thoroughly that you can't tell them apart anymore. This is called over-smoothing.

The Solution: scProfiterole

The authors created a new tool called scProfiterole (a playful name combining "sc" for single-cell and "profiterole," a cream-filled pastry, perhaps implying it fills the gaps with something sweet!).

Instead of just blindly mixing the data, scProfiterole uses a smart strategy called Graph Contrastive Learning combined with Spectral Filters. Let's break that down.

1. The Map (The Graph)

First, the tool builds a map. Imagine every cell is a person at a giant party.

• If two people are wearing similar outfits (have similar proteins), they stand close together.
• If they are different, they stand far apart.
• This creates a giant web of connections (a graph).

2. The Problem with the Map

Because the data is noisy, some people who should be standing together are standing apart, and some strangers are standing too close. The map is messy.

3. The Magic Glasses (Spectral Filters)

This is where the paper gets clever. Standard tools look at the map using a "wide-angle lens" that sees everything equally. If you zoom in too much (add more layers to the neural network), everything blurs into a gray soup (over-smoothing).

scProfiterole puts on a special pair of Spectral Glasses. These glasses don't just look at who is standing next to whom; they look at the vibrations or patterns of the whole party.

• They act like a Low-Pass Filter: Think of this like a noise-canceling headphone. It lets the deep, steady bass notes (the real, strong similarities between cells) pass through, but it blocks out the high-pitched, scratchy static (the noise and missing data).

The Three Types of Glasses

The paper tests three different styles of these "glasses" to see which one sorts the cells best:

1. Random Walk with Restart (RWR): Imagine a drunk person wandering the party. They take a step to a neighbor, then maybe a step back, then a step forward. They keep doing this, but occasionally they get bored and teleport back to where they started. This helps them explore the neighborhood without getting lost.

   • Result: It works okay, but it's a bit rigid.

2. Heat Kernels: Imagine dropping a hot stone into a pool of water. The heat spreads out smoothly and evenly over time. This filter looks at how "heat" (information) would diffuse through the cell network. It's very flexible and smooths out the rough edges of the data perfectly.

   • Result: The Winner. This method sorted the cells better than anything else.

3. Beta Kernels: This is a pre-packaged, mathematical recipe for smoothing. It's like using a specific, pre-mixed cake batter. It's simple and direct, but not as adaptable as the "Heat" method.

The Secret Sauce: Arnoldi Orthonormalization

Here is the technical magic trick. To make these "glasses" work, the computer has to do some heavy math to figure out exactly how to smooth the data. Usually, this math is unstable and breaks easily (like trying to balance a house of cards in a windstorm).

The authors used a technique called Arnoldi Orthonormalization.

• Analogy: Imagine you are trying to build a tower of blocks, but the blocks are slippery. Instead of stacking them one by one (which might make the tower fall), you use a special scaffolding system that locks the blocks in place perfectly before you let go. This allows the computer to calculate the perfect smoothing recipe without the math crashing.

What Did They Find?

1. Better Sorting: scProfiterole sorted the cells (identified cell types) much better than old methods. It was like going from sorting LEGO bricks by eye in the dark to using a smart robot that knows exactly which piece goes where.
2. Heat Kernels Rule: The "Heat Kernel" glasses were the best at handling the messy, noisy data.
3. Don't Guess, Calculate: The tool doesn't just guess how to smooth the data; it uses a precise mathematical recipe (polynomial interpolation) to get it right every time.
4. No Extra Cost: Even though this sounds complicated, it doesn't take much longer to run than the standard methods. It's fast and efficient.

The Takeaway

scProfiterole is a new, smarter way to organize the messy data of single-cell proteins. By using "spectral glasses" (specifically Heat Kernels) and a special math trick (Arnoldi) to handle the noise, it helps scientists clearly see the different types of cells in our bodies. This is a huge step forward for understanding diseases and how our bodies work at the most microscopic level.

Technical Summary

1. Problem Statement

The emergence of single-cell proteomics (scProteomics) technologies offers direct insights into cellular function and signaling, distinct from transcriptomic data (scRNA-seq). However, analyzing scProteomics data presents unique computational challenges:

• Data Quality: scProteomic data suffers from high levels of missing values (drop-outs), noise, and batch effects due to technical limitations in mass spectrometry and sample preparation.
• Algorithmic Limitations: Existing tools for scRNA-seq (e.g., standard Graph Convolutional Networks or GCNs) perform poorly on proteomic data.
  • Over-smoothing: Deep GCN architectures cause "over-smoothing," where node features become indistinguishable as information propagates through too many layers. This limits the ability to capture long-range dependencies essential for biological systems.
  • Graph Imperfections: Cell-to-cell similarity graphs derived from proteomic data are sparse and noisy. Standard GCNs using adjacency matrices treat all $k$-hop neighbors equally, making them vulnerable to graph imperfections.
  • Lack of Specialized Tools: There is a scarcity of algorithms specifically designed for clustering and cell type identification in scProteomics.

2. Methodology: scProfiterole

The authors propose scProfiterole (Single Cell Proteomics Clustering via Spectral Filters), a framework that integrates Graph Contrastive Learning (GCL) with Spectral Graph Theory to overcome the limitations of standard GCNs.

Core Components:

1. Spectral Graph Convolution:
   Instead of using the raw adjacency matrix $\hat{A}$ for feature propagation, scProfiterole replaces it with a spectral filter $g(\hat{A})$. The propagation equation is:
   $$H^{(l+1)} = \sigma(g(\hat{A})H^{(l)}W^{(l)})$$
   This allows the model to operate in the eigenspace of the graph, emphasizing specific frequency components (low-pass filtering) to smooth node features while preserving cluster structure.

2. Polynomial Interpolation via Arnoldi Orthonormalization:
   Direct computation of spectral filters (eigen-decomposition) is computationally prohibitive for large graphs. Filters are typically approximated using polynomials: $g(\hat{A}) = \sum \theta_k \hat{A}^k$.

   • Challenge: Standard methods (truncation or Taylor series approximation) often yield poor approximations, especially for complex filters like Heat Kernels. Furthermore, solving for polynomial coefficients via Vandermonde systems is numerically unstable (ill-conditioned).
   • Solution: scProfiterole uses Arnoldi orthonormalization to compute polynomial coefficients. This method constructs an orthonormal basis for the Krylov subspace, providing a numerically stable and accurate interpolation of the target filter function without requiring explicit eigen-decomposition.

3. Three Families of Homophilic Filters:
   The framework implements three types of low-pass filters, initialized via the Arnoldi method:

   • Random Walk with Restart (RWR): Truncated versions are common but limited. scProfiterole uses interpolation to create a more accurate polynomial representation.
   • Heat Kernels (HK): Modeled as continuous-time random walks ($e^{-T\mathcal{L}}$). These offer a flexible, broad family of filters. The paper argues that Taylor approximations of HK are suboptimal; interpolation provides superior fidelity.
   • Beta Kernels (BK): A family of filters defined directly as polynomials, offering a straightforward low-pass design without needing truncation or approximation.

4. Graph Contrastive Learning (GCL) Integration:
   The spectral filters are used within a GCL encoder (based on the scPROTEIN framework). The model learns cell embeddings by maximizing agreement between augmented views of the graph while minimizing noise, utilizing the spectral filters to guide the initial weights of the polynomial coefficients.

3. Key Contributions

• Framework Development: Introduction of scProfiterole, the first framework to combine spectral graph filters with GCL specifically for single-cell proteomic data.
• Numerical Stability: Development of a stable polynomial interpolation method using Arnoldi orthonormalization, overcoming the ill-conditioning of Vandermonde systems to accurately approximate complex spectral filters.
• Filter Analysis: Systematic comparison of RWR, Heat, and Beta kernels, demonstrating that Heat Kernels combined with interpolation yield the best performance.
• Initialization Strategy: Demonstration that the choice of spectral filter for initializing learnable polynomial coefficients is critical. Proper initialization guides the learning process toward robust solutions, reducing sensitivity to data fluctuations.

4. Experimental Results

The framework was evaluated on three scProteomics datasets: Scope2_Specht (1,490 cells), N2, and nanoPOTS (integrated).

• Performance Metrics: Adjusted Rand Index (ARI), Average Silhouette Width (ASW), Normalized Mutual Information (NMI), and Purity Score (PS).
• Key Findings:
  1. Superiority over Baselines: scProfiterole significantly outperformed classical clustering methods (K-means, Louvain) and standard GCL baselines (GCN with raw adjacency matrix).
  2. Heat Kernels are Optimal: Interpolated Heat Kernels provided the best clustering performance. Compared to the standard GCN baseline, the Interpolated Heat Kernel improved ARI by 29.7%, ASW by 9.1%, and NMI by 8.1%.
  3. Interpolation vs. Approximation: For both RWR and Heat Kernels, the interpolated versions significantly outperformed their truncated or Taylor-approximated counterparts. Interpolation allowed the model to capture higher-order relationships with lower polynomial degrees.
  4. Robustness to Initialization: Models initialized with specific spectral filters (especially Heat Kernels) converged to better solutions than those with random initialization. The learned filters retained the shape of the initial spectral filter, confirming that initialization guides the learning trajectory.
  5. Sparsity and Noise: Heat Kernels demonstrated robustness across varying levels of graph sparsity (controlled by correlation thresholds). While performance generally dropped with extreme sparsity, Heat Kernels remained the most stable.
  6. Computational Efficiency: The polynomial interpolation step adds negligible computational cost (milliseconds) compared to the overall GCL training time.

5. Significance

• Bridging the Gap: scProfiterole addresses the critical lack of tools for scProteomics analysis, enabling effective cell type identification in a domain previously dominated by scRNA-seq tools.
• Overcoming Over-smoothing: By utilizing spectral filters, the method achieves deep neighborhood aggregation without the over-smoothing issues associated with deep GCN layers, allowing for the modeling of long-range biological dependencies.
• Scalability and Stability: The use of Arnoldi orthonormalization makes the application of complex spectral filters scalable and numerically stable, paving the way for applying advanced graph representation learning to noisy, high-dimensional biological data.
• Future Impact: The framework provides a principled foundation for next-generation proteomics, suggesting that careful design of spectral filters is more impactful than simply increasing network depth.

In conclusion, scProfiterole demonstrates that Graph Contrastive Learning guided by spectrally interpolated low-pass filters (specifically Heat Kernels) is a powerful approach for clustering single-cell proteomic data, offering significant improvements in accuracy and robustness over existing state-of-the-art methods.