SR2P: an efficient stacking method to predict protein abundance from gene expression in spatial transcriptomics data

The paper introduces SR2P, a high-performance stacking-based machine learning framework that accurately predicts spatial protein abundance from RNA-only gene expression data, thereby overcoming the limitations of current spatial transcriptomics technologies to enable deeper analysis of tumor immunology and immune cell signaling.

The Gist

Imagine you are trying to understand a bustling city. You have two ways to look at it:

1. The Blueprint (RNA): You have a detailed list of every instruction manual in every building. This tells you what the building plans to do.
2. The Reality (Proteins): You walk the streets and see what the buildings are actually doing. Are they open? Are they crowded? Are they under construction?

In biology, RNA is the blueprint, and Proteins are the reality. Usually, scientists can only see the blueprint (RNA) in high-resolution maps of tissues (Spatial Transcriptomics). They can't easily see the proteins because measuring them is expensive, slow, and technically difficult.

This creates a problem: Just because a building has a blueprint for a "Fire Station," doesn't mean the fire station is actually open and staffed. Sometimes the instructions are there, but the protein (the actual fire station) is missing or different.

Enter SR2P: The "Super-Translator"

The paper introduces a new tool called SR2P. Think of SR2P as a super-smart translator or a crystal ball that can look at the blueprints (RNA) and accurately predict what the buildings are actually doing (Proteins) in 3D space.

Here is how it works, broken down simply:

1. The Problem: The "Missing Link"

Most modern maps of our bodies (like those from 10x Genomics) only show the RNA. They miss the proteins. This is like trying to diagnose a sick patient by only reading their medical school textbooks, without ever seeing their actual symptoms. This is especially bad for studying the immune system, where the "surface markers" (proteins) tell us if immune cells are fighting or hiding.

2. The Solution: A "Team of Experts" (Stacking)

The researchers didn't just build one model to guess the proteins. They built a team of 11 different AI experts.

• Some experts are good at spotting patterns in lists (Tree-based models like XGBoost).
• Some experts are good at understanding neighborhoods and how spots on a map talk to each other (Graph Neural Networks).
• Some are good at simple math (Linear models).

Instead of picking just one expert, SR2P uses a "Stacking" method. Imagine a panel of judges. Each of the 11 experts makes a guess about the protein levels. Then, a "Head Judge" (a meta-learner) looks at all their guesses, weighs them, and combines them into one final, highly accurate prediction. This is why it's called a "stacking" method—it stacks the strengths of many models to get the best result.

3. The Magic Ingredient: "Neighborhood Watch"

One of the key tricks SR2P uses is understanding neighborhoods.
In a tissue, a cell doesn't exist in a vacuum. It talks to the cells next to it.

• Non-Spatial Models: Look at a single cell's blueprint and guess its protein.
• SR2P (Spatial Models): Looks at the cell and its four immediate neighbors (North, South, East, West). It knows that if a cell is surrounded by "immune cells," it's likely an immune cell too. This "neighborhood context" makes the prediction much more accurate.

4. The Results: What Did They Find?

The team tested SR2P on real data from breast cancer, tonsils, and head/neck cancers.

• Accuracy: SR2P was better at guessing the proteins than any single method alone. It could recreate the "map" of where immune cells were hiding, even though it only had the RNA data to start with.
• The Catch (Tissue Specificity): They found that a model trained on a "Breast Cancer" blueprint works great for other breast cancers, but it struggles if you try to use it on a "Brain Tumor" blueprint. It's like a translator who is fluent in French but gets confused when you switch to Japanese. You need to train your translator on the specific "language" of the tissue you are studying.
• Real-World Win: They used SR2P on patients with head and neck cancer who were treated with immunotherapy. By predicting the protein levels, they could see which patients had "hot" tumors (full of fighting immune cells) and which had "cold" tumors (full of suppressive cells). This helped explain why some patients responded to treatment and others didn't.

The Bottom Line

SR2P is a cost-effective superpower.

Instead of spending thousands of dollars and weeks of lab time to measure proteins in every single tissue sample, scientists can now use SR2P to look at the RNA data they already have and predict the protein landscape with high accuracy.

It turns a "flat" map of instructions into a "3D" map of reality, helping doctors and researchers understand the tumor's immune environment better, faster, and cheaper. It's like upgrading from a black-and-white sketch to a full-color, high-definition movie of what's happening inside your body.

Technical Summary

1. Problem Statement

Spatial transcriptomics (ST) technologies (e.g., 10x Genomics Visium) have revolutionized the mapping of gene expression within tissue architecture. However, most widely available ST datasets contain RNA expression data only, lacking direct measurements of surface protein abundance.

• The Gap: Proteins are the functional executors of cellular activity and are often the primary therapeutic targets. Relying solely on RNA is insufficient because of post-transcriptional regulation, leading to significant RNA-protein abundance discordance, particularly for immune cell surface markers.
• The Challenge: While emerging multi-omics technologies (e.g., Visium CytAssist, DBiT-seq) can measure both RNA and protein simultaneously, they are costly, technically demanding, and not yet widely adopted. There is a critical need for a cost-effective computational method to infer protein abundance from existing RNA-only spatial datasets to enable retrospective analysis and immune profiling in the tumor microenvironment (TME).

2. Methodology: The SR2P Framework

The authors developed SR2P, a stacking-based machine learning framework designed to predict spatial protein abundance from spatial transcriptomic profiles.

Core Architecture

SR2P utilizes an ensemble learning (stacking) approach that integrates 11 distinct predictive models as "base learners" to leverage their complementary strengths:

1. Linear Models: Partial Least Squares (PLS) regression.
2. Tree-Based Ensemble Models: XGBoost, LightGBM, and CatBoost.
3. Graph Neural Networks (GNNs): Graph Attention Networks (GAT), GraphSAGE, and Dual Graph Attention Networks (DGAT).

Spatial Feature Engineering

To address the unique nature of spatial data, the authors implemented two strategies:

• Native Spatial Modeling: GNNs inherently model spatial relationships by constructing graphs where spots are nodes connected by edges based on Euclidean distance (k-nearest neighbors).
• Spatial Augmentation for Non-GNNs: For linear and tree-based models, the authors explicitly constructed spatial features. For each spot, they concatenated the gene expression vectors of its four cardinal neighbors (North, South, East, West). If neighbors were missing due to tissue boundaries, the closest available unselected neighbors were used. This created "Spatial" versions of the models (e.g., CatBoost-Spatial).

Stacking Mechanism

• Base Learners: The 11 models (including spatial and non-spatial variants) generate predictions.
• Meta-Learner: The out-of-fold predictions (OOFPs) from all base learners are aggregated to form a meta-feature matrix. An Extremely Randomized Trees (ExtraTrees) model is trained on this matrix to produce the final prediction. This meta-learner was selected based on its superior performance in preliminary experiments.

Evaluation Strategy

The framework was rigorously benchmarked using three validation settings across six spatial genomics datasets (Breast Cancer, Glioblastoma, Tonsil, and HNSCC):

1. Within-Sample Validation: 10-fold spatial cross-validation (using k-means clustering on coordinates to ensure spatial coherence).
2. Within-Tissue Validation: Training on one sample and testing on another of the same tissue type.
3. Cross-Tissue Validation: Training on one tissue type and testing on a completely different tissue type to assess generalizability.

3. Key Results

Performance Benchmarking

• Superior Accuracy: SR2P consistently outperformed all individual base learners and competing methods across all validation settings. It achieved the highest Spearman correlations and lowest Root Mean Squared Error (RMSE).
• Impact of Spatial Information: Incorporating spatial features significantly improved the performance of non-GNN models (e.g., CatBoost-Spatial and LightGBM-Spatial), often surpassing GNN-based models. This highlights that explicit spatial context is crucial for accurate protein prediction.
• GNN Limitations: While GNNs (DGAT, GAT) showed promise, they generally exhibited higher variance and lower performance compared to spatially augmented tree-based models, suggesting current GNN architectures may not fully capture the complex structure-function relationships in spatial proteomics.

Generalizability and Tissue Specificity

• Tissue Dependence: Prediction accuracy was highly tissue-dependent. Models trained within the same tissue type performed best.
• Cross-Tissue Transfer: Performance dropped significantly when models were trained on one tissue and tested on another (e.g., training on Breast Cancer and testing on HNSCC). However, SR2P maintained the most robust performance in these cross-tissue scenarios compared to other methods, though absolute accuracy decreased.
• Marker Specificity: The framework successfully predicted high-abundance, spatially coherent markers (e.g., CD45, CD163, EPCAM) but struggled with low-abundance or weakly structured markers (e.g., specific myeloid markers).

Computational Efficiency

• Despite the complexity of stacking 11 models, SR2P demonstrated high computational efficiency. The inference time for the meta-learner alone was 2.19 seconds on a standard server. Even including base learner inference, the total time remained practical (seconds per sample), making it suitable for large-scale studies.

4. Biological Application: HNSCC Immunotherapy

The authors applied SR2P to a cohort of Head and Neck Squamous Cell Carcinoma (HNSCC) patients treated with immune checkpoint blockade (anti-PD-1) therapy.

• Recovering Immune Regions: In RNA-only datasets, spatial clustering often missed immune-rich regions. By integrating predicted protein abundance with RNA data, SR2P identified 9.7% more macrophage-enriched spatial spots compared to RNA-only analysis.
• Therapeutic Response Biomarkers: The predicted protein profiles successfully distinguished between Responders and Non-Responders:
  • Responders: Showed higher abundance of T-cell markers (CD8A, CD45) and "immunologically hot" phenotypes.
  • Non-Responders: Enriched for immunosuppressive myeloid/macrophage markers (CD68, CD14, ITGAX).
• Validation: The predicted protein maps correlated strongly with deconvoluted cell-type proportions and known marker gene expression, validating the biological relevance of the predictions.

5. Significance and Conclusion

• Unlocking Retrospective Data: SR2P enables researchers to extract protein-level insights from the vast existing archives of RNA-only spatial transcriptomics data, bypassing the need for expensive multi-omics re-profiling.
• Enhanced TME Characterization: By inferring surface protein abundance, the method significantly improves the resolution of immune cell mapping (e.g., macrophages, T-cells) within the tumor microenvironment, which is critical for understanding immunotherapy resistance.
• Methodological Advancement: The study demonstrates that stacking diverse model architectures (linear, tree-based, and graph-based) combined with explicit spatial feature engineering is a superior strategy for spatial omics prediction compared to single-model approaches or GNNs alone.
• Practical Utility: SR2P is computationally efficient and provides a robust, generalizable framework for spatial multi-omics integration, offering a practical solution for tumor immunology studies where direct protein measurements are unavailable.

Availability: The SR2P software is open-source and available on GitHub.