Single-cell spatial multi-omics molecular pathology enabled by SuperFocus

SuperFocus is a modality-agnostic computational platform that integrates histopathology with single-cell spatial multi-omics by accurately projecting genome-scale molecular data onto tissue morphology without external references, thereby enabling scalable, cell-resolved molecular pathology analyses across diverse disease contexts.

The Gist

Imagine you are trying to understand a bustling city. You have two very different ways of looking at it:

1. The Aerial Photo (Histology): You have a high-resolution satellite image of the whole city. You can see the buildings, the parks, the streets, and exactly where every house is. But you can't hear what the people inside are saying, what they are eating, or what their jobs are.
2. The Street Survey (Spatial Omics): You send out a team of reporters to interview people. They are very good at gathering deep data (genes, proteins, chemicals), but they can only stand in specific, large squares on the map. They can't interview everyone, and their reports are a bit blurry because they are averaging the answers of everyone standing in that big square.

The Problem:
For a long time, doctors and scientists had to choose between the clear picture of the city (the photo) or the deep data (the survey). They couldn't easily combine them to say, "That specific person in that specific house is a baker who is currently stressed."

The Solution: SuperFocus
The paper introduces a new AI tool called SuperFocus. Think of it as a super-smart translator and detective that can take the blurry street survey and project it perfectly onto the high-resolution aerial photo, cell by cell.

Here is how it works, using simple analogies:

1. The "Zoom-In" Ladder (Cascading Imputation)

Imagine you have a blurry photo of a crowd. You want to know what each person is wearing.

• Step 1: You look at a big group of 100 people (a "spot") and guess the average outfit.
• Step 2: You zoom in to a group of 25 people. You use your knowledge of the big group to make a better guess for this smaller group.
• Step 3: You zoom in again to 5 people, then finally to one single person.

SuperFocus does this mathematically. It starts with the big, blurry data spots and uses a "ladder" of AI models to step down, getting sharper and sharper until it can predict the molecular data for every single cell in the tissue, not just the ones the reporters interviewed.

2. The "Trust Score" (Quality Control)

One of the biggest fears with AI is that it might just "hallucinate" or make things up when it's guessing.
SuperFocus is different because it carries a Trust Score for every single prediction.

• If the AI sees a cell that looks very similar to the cells it was trained on, it gives a High Trust Score (Green light: "I'm pretty sure this is right").
• If the AI sees a weird, rare cell type it hasn't seen before, it gives a Low Trust Score (Yellow/Red light: "I'm guessing here, be careful").

This is like a weather forecaster saying, "It will rain tomorrow (90% confidence)" versus "It might rain, but I'm not sure (20% confidence)." This prevents doctors from making bad decisions based on bad guesses.

3. The "Universal Adapter" (Modality-Agnostic)

Usually, these tools only work for one type of data (like just RNA). SuperFocus is like a universal power adapter. It works with:

• RNA (instructions for making proteins)
• Epigenetics (switches that turn genes on/off)
• Proteins (the actual workers in the cell)
• Metabolites (chemicals and nutrients)

It can even take data from two different maps that don't line up perfectly (like a map of the city's traffic and a map of its power grid) and merge them into one perfect picture.

Real-World Examples from the Paper

The authors tested SuperFocus on four different "cities" (diseases) and found amazing things:

• The MALT Lymphoma (Stomach Cancer): They found that inside the tumor, there were different neighborhoods. Some areas had "sleepy" immune cells, while others had "angry" ones. SuperFocus showed exactly where these cells were interacting, which helps explain why the cancer is growing.
• The Human Hippocampus (Brain): They mapped out the brain's "wiring" (chromatin accessibility). They could see which genes were "switched on" in specific neurons, helping us understand how the brain is organized at a microscopic level.
• The Liver (MASH): They found a specific group of liver cells that were "poisoned" by fat (lipotoxic). These cells were stressed and inflamed, a key step in liver disease that was previously hard to spot.
• The Parkinson's Mouse Brain: They combined a map of brain chemicals (metabolites) with a map of brain genes. They discovered that in the diseased part of the brain, there was a lack of a protective chemical called taurine in the white matter, and an overabundance of immune cells (microglia) trying to clean up the mess.

Why This Matters

Before SuperFocus, scientists had to choose between seeing the whole picture (the whole slide) or seeing the fine details (single cells). They couldn't have both.

SuperFocus changes the game. It allows us to take a cheap, fast, spot-based test and turn it into a high-definition, whole-slide, single-cell movie of what is happening inside a patient's body. It bridges the gap between the pathologist looking at a slide under a microscope and the geneticist looking at a spreadsheet of data, finally letting them speak the same language.

In short: SuperFocus takes a blurry, low-resolution map of a city and uses AI to fill in the missing details, giving us a crystal-clear view of every single citizen (cell) and what they are doing, while telling us exactly how confident we should be in that view.

Technical Summary

1. Problem Statement

Current spatial omics technologies (e.g., Visium, DBiT-seq, MALDI-MSI) offer scalable, genome-scale molecular profiling but typically measure signals aggregated over "spots" or pixels containing multiple cells. This limits the resolution of cellular heterogeneity and whole-slide molecular architecture. Conversely, histopathology (H&E) provides high-resolution cellular morphology but lacks molecular depth.

• The Gap: There is a lack of a unified, modality-agnostic computational framework that can project genome-scale molecular information onto histopathology images at single-cell resolution across entire tissue sections without requiring external reference data. Existing methods often lack robust quality control, fail to generalize beyond transcriptomics, and struggle with unreliable predictions in unprofiled regions.

2. Methodology: The SuperFocus Framework

SuperFocus is a computational platform designed to infer single-cell omics features from spot-based measurements using a constrained cascading imputation strategy integrated with histology. It operates in three stages:

Stage 1: Input Curation & Feature Encoding

• Inputs: A high-resolution digitized H&E image (20x or 40x) and spot-level spatial omics data (transcriptomics, epigenomics, proteomics, or metabolomics) from the same or an adjacent section.
• Cell Segmentation: A computational pathology foundation model (default: CellViT) segments cells and encodes morphological features.
• Dyadic Partitioning: The original omics spots are recursively partitioned into smaller "subspots" (dyadic splitting) until the majority of subspots contain only one cell. This creates a hierarchy of resolution levels (Level 0 = original spot; Level $L$ = single-cell).
• Feature Aggregation: Image encoding vectors are computed for each subspot by aggregating the embeddings of cells within and immediately surrounding the subspot.

Stage 2: Constrained Cascading Imputation & Training

• Cascade Architecture: A series of deep neural networks are trained, one for each resolution level.
• Trans-Resolution Learning: Models are trained recursively. The parameters learned at a coarser resolution (e.g., 64 $\mu$m) serve as a "warm start" for the next finer resolution (e.g., 32 $\mu$m). This stabilizes predictions as resolution increases.
• In-Sample Quality Control (QC): Two diagnostic metrics are computed for each feature to quantify prediction reliability:
  1. Relative Sum-of-Squared-Errors (RSSE): Measures global prediction error by comparing the sum of predicted subspot values against the original observed spot value.
  2. Error Spatial Locality (ESL): Measures the spatial concentration of errors (ratio of $\ell_2$ to $\ell_1$ norm of the error vector). High ESL indicates clustered errors, flagging unreliable regions.

Stage 3: Whole-Slide Inference & Credibility Scoring

• Prediction: The finest-resolution model predicts omics profiles for every cell in the entire tissue section.
• Credibility Score: For cells outside the original profiled Field of View (FOV), a prediction-credibility score is assigned. This score decreases as the morphological distance between the cell's local neighborhood and the training data increases, flagging extrapolative predictions that may be unreliable.

3. Key Contributions

1. Modality Agnosticism: Unlike previous tools focused solely on transcriptomics, SuperFocus works across multi-omics modalities (RNA, ATAC-seq, Proteins, Metabolites) and can integrate mismatched spatial grids (e.g., Visium + MALDI-MSI).
2. Built-in Reliability: Introduces RSSE, ESL, and Credibility Scores to explicitly flag unreliable predictions, moving beyond simple visualization to actionable, confidence-aware molecular pathology.
3. Scalability: Enables whole-slide, single-cell resolution analysis from limited spot-based data, bridging the gap between scalable experimental acquisition and cellular-resolution inference.
4. No External Reference Required: The method infers cell states directly from the paired histology and spot data without needing external single-cell reference atlases (though it can utilize them for annotation).

4. Results & Validation

Benchmarking (Ground Truth Validation)

• Dataset: Human kidney VisiumHD data (64 $\mu$m spots for training, 8 $\mu$m spots as ground truth).
• Performance: SuperFocus improved key accuracy metrics by 28–73% over state-of-the-art methods (iStar and Thor).
  • MS-SSIM (Structural Similarity): Improved from 0.59/0.67 (baselines) to 0.86.
  • NRMSE (Error): Improved from 0.117/0.131 to 0.035.
• QC Validation: The in-sample RSSE and ESL metrics showed high correlation ($r > 0.85$) with ground-truth errors, proving their utility in predicting model performance.

Application Case Studies

1. MALT Lymphoma (Patho-DBiT):
   • Resolved heterogeneous cell compositions and cell-cell interactions within tumor microenvironments.
   • Identified specific B-cell subgroups (e.g., TLR4+, FOXP3+) and activated T-helper cells driving the tumor niche, revealing pathways (NF-$\kappa$B, JAK-STAT) not visible at spot resolution.
2. Human Hippocampus (Spatial ATAC-RNA):
   • Inferred single-cell chromatin accessibility and gene regulatory networks (GRNs).
   • Successfully reconstructed cell-type specific TF activities (e.g., NFIA in neurons, SOX10 in oligodendrocytes) and linked them to spatial gene expression.
3. Human Liver MASH (Spatial CITE-seq):
   • Identified a lipotoxic hepatocyte subpopulation characterized by high LOX-1, CD142, and TIGIT protein expression.
   • Validated this finding independently via transcriptomics, showing upregulation of metabolic stress genes (ACACA, HPR) in the same cluster.
4. Parkinsonian Mouse Brain (Visium + MALDI-MSI):
   • Integrated transcriptomic and metabolomic data from mismatched grids (55 $\mu$m vs. 100 $\mu$m).
   • Revealed spatially localized metabolic signals (e.g., dopamine depletion, taurine reduction) associated with microglial activation and white-matter remodeling in the corpus callosum.

5. Significance

• Paradigm Shift: SuperFocus shifts the operating point of spatial molecular pathology from "spot-level" to "whole-slide, cell-resolved" analysis, making high-resolution molecular profiling accessible without the prohibitive cost of single-cell spatial assays for entire slides.
• Clinical Utility: By providing confidence scores and reducing spurious predictions, it makes single-cell spatial data actionable for clinical diagnostics and translational research.
• Future Integration: The framework is designed to work in tandem with digital pathology foundation models (e.g., UNI, Prov-GigaPath) and ROI selection tools, creating a closed loop where histology guides assay placement and computational reconstruction fills in the molecular landscape.

In summary, SuperFocus represents a significant advancement in computational pathology, enabling the reconstruction of high-fidelity, single-cell multi-omics maps from scalable, spot-based experiments while rigorously quantifying the reliability of those reconstructions.