Binary-SPA: A Reference-Free Method for Cell Annotation in High-Resolution Spatial Transcriptomics

Binary-SPA is a robust, reference-free computational framework that achieves high-accuracy, full-coverage cell type annotation in high-resolution spatial transcriptomics by combining initial marker-based classification with internal anchor-based label transfer, thereby eliminating the need for external single-cell RNA sequencing references.

The Gist

Imagine you have a massive, high-resolution photograph of a bustling city. Every single person in the photo is a cell, and your job is to figure out exactly who they are: are they a baker, a police officer, a teacher, or a construction worker?

In the world of biology, this is called cell annotation. Scientists use a powerful new technology called Spatial Transcriptomics to take these "photos" of tissues, seeing not just where cells are, but what genes they are "speaking" (expressing).

However, there's a big problem: How do you label everyone correctly?

The Old Ways: Two Flawed Approaches

1. The "Guest List" Method (Label Transfer):
   Imagine trying to identify the people in your photo by comparing them to a guest list from a different party you attended last year.

   • The Problem: If the people at your current party are wearing different clothes, acting differently, or if the guest list is from a completely different city, the comparison fails. In biology, this means if you try to match your tissue sample to a reference dataset from a different person or a different disease state, the results are often wrong. Plus, if you don't have a guest list at all (like with old, archived medical samples), this method is useless.

2. The "Spot the Badge" Method (Marker-Based):
   Imagine looking at each person and saying, "If I see a police hat, they are a cop."

   • The Problem: Sometimes a police officer isn't wearing their hat. Sometimes a construction worker is wearing a hat that looks like a police hat. If you rely on just one or two "badges" (genes), you might miss people or misidentify them. Also, this method often leaves many people in the photo unlabeled because they don't have a clear "badge" visible.

The New Solution: Binary-SPA

The authors of this paper created a new tool called Binary-SPA. Think of it as a smart, self-teaching detective that solves the problem without needing an outside guest list.

Here is how it works, using a simple two-step analogy:

Step 1: The "Confident Crowd" (Binary Step)

First, the detective looks at the crowd and asks: "Who is wearing multiple clear badges?"

• Instead of caring about how loud someone is shouting (gene expression levels), Binary-SPA just cares if they are speaking or not (On/Off).
• If a cell has a few clear "badges" (genes) that say "I am a baker," it gets labeled immediately.
• These are the "Clear Cells." They are the ones the computer is 100% sure about.

Step 2: The "Peer Group" (SPA Step)

Now, what about the people who aren't wearing clear badges? Maybe their hat is tilted, or they are standing in the shadows.

• Instead of calling in an outside expert (a reference dataset), the detective asks the "Clear Cells" to help.
• "Hey, you look like a baker. You are standing right next to this confused person. Does this person look like they belong in your group?"
• Because everyone is in the same photo (the same tissue sample), they all share the same lighting, the same background noise, and the same style. The "Clear Cells" act as a perfect internal reference to teach the computer what the "Unclear Cells" are.

Why This is a Game-Changer

The paper tested this new method on all sorts of tricky situations:

• Different Cameras: It worked perfectly whether the "photo" was taken with a high-end camera (Xenium) or a slightly different model (Visium HD).
• Old Photos: It worked on fresh tissue and on old, archived tissue (like formalin-fixed samples) where the "image quality" (RNA) is often degraded.
• The "Bone Marrow" Challenge: Bone marrow is like a crowded subway station where everyone looks very similar and is constantly changing. Old methods struggled here, but Binary-SPA successfully identified the different cell types and even spotted the subtle changes that happen as a disease (like multiple myeloma) progresses.
• The Gold Standard: When they compared their results to a protein-based "truth" (like checking ID cards directly), Binary-SPA was almost perfectly accurate (96.8% match), beating every other method.

The Bottom Line

Binary-SPA is like a self-sufficient translator. It doesn't need a dictionary from another language (external reference data). Instead, it finds the people who are speaking clearly, uses them to understand the context, and then translates the rest of the crowd.

This means doctors and researchers can now analyze any tissue sample—even old, archived ones from the hospital basement—with high accuracy, without needing to find a matching "reference" sample first. It turns a messy, confusing crowd into a perfectly organized directory.

Technical Summary

1. Problem Statement

Accurate cell type annotation is a critical bottleneck in high-resolution spatial transcriptomics (ST). Existing methods face significant limitations:

• Label Transfer Methods: These rely on transferring annotations from external single-cell RNA sequencing (scRNA-seq) reference datasets. They suffer from:
  • Reference Dependency: Performance degrades significantly when the reference does not perfectly match the tissue (e.g., diseased vs. healthy, different individuals, or archived clinical samples).
  • Coverage Gaps: They often fail to annotate cells with ambiguous transcriptomic profiles, leaving a portion of the data unannotated.
• Marker-Based Methods: These rely on predefined marker genes but often suffer from:
  • Incomplete Coverage: They frequently fail to annotate cells that do not express high levels of specific markers due to stochastic gene expression or technical sensitivity limits.
  • Scoring Artifacts: Traditional scoring often relies on expression magnitude, which can be skewed by highly expressed single genes rather than the combinatorial presence of multiple markers.
• Clustering Mismatch: Unsupervised clustering groups cells by global transcriptomic similarity, which often conflicts with traditional cell typing based on specific marker proteins, leading to inconsistent or overly fragmented annotations.

2. Methodology: Binary-SPA

The authors developed Binary-SPA (Binary Self-referenced Projection Annotation), a two-stage, reference-free computational framework designed for high-resolution ST data (e.g., Xenium, Visium HD).

Stage 1: Binary Classification (The "Binary" Step)

This stage identifies high-confidence cells using a user-defined marker matrix without external references.

1. Quality Control: Unsupervised clustering is performed to detect unexpected populations (e.g., metastatic cells or novel disease states) and update the marker matrix accordingly.
2. Marker Matrix Construction: A user-defined matrix is created where rows are cell types and columns are marker genes. Entries are set to 1 if the gene is a known positive marker for that cell type, and 0 otherwise.
3. Platform Adaptation: The marker matrix is intersected with the genes actually detected in the specific spatial dataset to ensure compatibility.
4. Binarization: The cell-by-gene expression matrix is converted to binary: 1 for detectable expression and 0 for non-detectable. This prioritizes the presence of multiple markers over the magnitude of expression, mimicking classical immunophenotyping logic.
5. Cell Type Score (CTS) Calculation: A matrix multiplication is performed between the marker matrix and the transposed binarized expression matrix. This yields a CTS for each cell type per cell, representing the count of detected markers.
6. Normalization & Confidence Filtering: CTS values are min-max normalized. The $\Delta$CTS (difference between the highest and second-highest normalized scores) is calculated.
   • Clear Cells: Cells with $\Delta$CTS > threshold (e.g., 0.15) are confidently annotated.
   • Unclear Cells: Cells with low $\Delta$CTS are left unannotated for the next stage.

Stage 2: Self-Referenced Projection Annotation (The "SPA" Step)

This stage resolves the "unclear" cells using an internal reference.

1. Internal Reference Generation: The "Clear Cells" from Stage 1 serve as the reference dataset.
2. Anchor-Based Label Transfer: Using the MapQuery function in Seurat, annotations are transferred from the Clear Cells to the Unclear Cells.
3. Advantage: Because both sets of cells originate from the same sample, batch effects and domain shifts are eliminated. The method leverages the fact that even cells lacking strong marker expression retain global transcriptomic signatures of their true identity.

3. Key Contributions

• Reference-Free Framework: Eliminates the need for external scRNA-seq references, making it applicable to archived clinical specimens (FFPE) and rare tissues where matched references do not exist.
• 100% Annotation Coverage: Achieves complete cell annotation by combining high-confidence marker-based calls with self-referenced label transfer, solving the coverage gap of pure marker-based methods.
• Robustness to Technical Variability: The binary scoring approach is resilient to differences in sequencing depth, platform sensitivity, and RNA degradation (common in decalcified bone marrow).
• Biological Interpretability: By relying on user-defined marker sets and individual cell scoring rather than global clustering, the output aligns better with traditional biological cell type definitions.

4. Results & Validation

The authors validated Binary-SPA across multiple platforms (Xenium, Visium HD), tissue types (Colon, Liver, Ovary, Bone Marrow), and preservation methods (Fresh-Frozen, FFPE).

• Benchmarking against Consensus (COAD, HCC, OV):
  • Compared to a "voting-based" consensus of five methods using matched scRNA-seq references, Binary-SPA achieved 100% annotation coverage (vs. ~90% for the voting method).
  • In tile-based Pearson correlation analysis against CODEX protein imaging ground truth, Binary-SPA showed higher or comparable accuracy (e.g., median $r=0.87$ vs. $0.85$ in Colon Adenocarcinoma) while requiring no external data.
• Comparison to Other Algorithms:
  • Outperformed label-transfer tools (Tangram, TACCO, SELINA, SPOINT, CellTypist) when external references were mismatched or unavailable.
  • Significantly outperformed marker-based tools (TACIT, ScType) in both coverage (100% vs. 42–84%) and the ability to recover all predefined cell types.
• Bone Marrow Validation (Challenging Tissue):
  • Successfully annotated bone marrow biopsies (including decalcified samples with degraded RNA) where reference-dependent methods (SingleR) failed to detect clinically relevant progression in plasma cell abundance.
  • Binary-SPA showed a strong correlation ($r=0.894$) with clinical manual plasma cell counts, significantly outperforming SingleR ($r=0.731$).
• Archival Clinical Specimens:
  • Applied to a normal bone marrow clot biopsy (standard clinical sample, no decalcification). Binary-SPA achieved 100% coverage and 19 cell types.
  • Validated against Lunaphore COMET multiplexed protein imaging, achieving near-perfect concordance ($r=0.968$), whereas other methods showed significantly lower correlations.

5. Significance

Binary-SPA represents a paradigm shift in spatial transcriptomics analysis by decoupling high-accuracy cell annotation from the requirement of matched external reference datasets.

• Clinical Applicability: It enables robust analysis of archival FFPE samples and clinical biopsies (like bone marrow clot biopsies) where matched scRNA-seq references are rarely available.
• Standardization: It provides a scalable, reproducible workflow that bridges the gap between classical marker-based immunophenotyping and modern high-throughput spatial omics.
• Reliability: By validating against orthogonal protein data (CODEX, COMET) and clinical counts, it establishes a new standard for accuracy in complex, heterogeneous tissues, facilitating both research discovery and clinical diagnostics.