Evidence of off-target probe binding affecting 10x Genomics Xenium gene panels compromise accuracy of spatial transcriptomic profiling

This study introduces the Off-target Probe Tracker (OPT) tool to identify and validate off-target probe binding in 10x Genomics Xenium panels, demonstrating that such non-specific interactions can significantly distort spatial gene expression profiles and compromise data accuracy.

The Gist

The Big Picture: A Case of Mistaken Identity

Imagine you are a detective trying to find a specific suspect in a crowded city. You have a "Wanted" poster with a very specific description of the suspect's face (the probe). Your job is to scan the crowd and point out anyone who matches that description.

In the world of biology, scientists use a high-tech tool called Xenium (made by 10x Genomics) to do exactly this. They want to find specific genes (the suspects) inside cells to see where they are and how active they are. The tool uses tiny molecular "posters" (probes) that are designed to stick only to the specific gene they are looking for.

The Problem:
This paper discovered that sometimes, the "Wanted" poster isn't specific enough. The molecular poster might stick to a look-alike (a different gene that looks very similar) instead of the real suspect. When this happens, the machine gets confused. It thinks it found the target gene, but it's actually seeing a mix of the target and the look-alike. This leads to a "false alarm" or a distorted map of where genes are actually working.

The Detective's New Tool: OPT

The authors of this paper, a team from Johns Hopkins University, built a new software tool called OPT (Off-target Probe Tracker).

Think of OPT as a super-strict background check system. Before scientists even start their expensive experiments, they can run their "Wanted posters" (probe sequences) through OPT. OPT scans the entire library of human genes to see: "Hey, does this poster match anyone else in the crowd besides the person we are looking for?"

What They Found

The team tested a popular "Wanted Poster" set used for studying breast cancer (the Human Breast Gene Panel).

1. The Mistaken Identities: They found that for at least 14 out of 313 genes, the probes were likely sticking to the wrong genes.
   • Analogy: Imagine you are looking for a person named "John Smith." Your poster is so generic that it also sticks to "Jon Smythe," "John Smythe," and "Jonathan Smith." Your machine counts all of them as "John Smith," giving you a wrong headcount.
2. The Proof: To prove this wasn't just a computer glitch, they compared the Xenium results with two other independent methods (like checking with a second detective and a third witness).
   • The "Visium" Test: They looked at the same tissue slice using a different technology. For the gene APOBEC3B, Xenium said it was highly active in certain areas. But the other two methods said, "That gene isn't even there!"
   • The "Aha!" Moment: When the researchers added up the activity of the real target gene plus the look-alike genes, the numbers finally matched what Xenium was seeing. This confirmed that Xenium was seeing a "sum" of the target and the impostors.

Why Does This Matter?

1. It's Expensive: These tests cost about $5,000 per slide. If the data is wrong because of "impostor" genes, researchers might waste months of work and thousands of dollars drawing the wrong conclusions.
2. It Skews Science: If a drug is designed to target a specific gene, but the data says that gene is active when it's actually a look-alike, the drug might fail in clinical trials.
3. The "Paralog" Problem: Many of these mistakes happen with paralogs—genes that are like siblings. They share the same family DNA (sequence) but have different jobs. If you can't tell them apart, you can't understand what each one is actually doing.

The Solution and The Future

The authors aren't saying "Stop using Xenium!" They are saying, "Be smarter about how you use it."

• Check Your Work: They recommend using tools like OPT to check probe designs before buying them.
• Look for the Look-Alikes: If a gene has a known "sibling" with a similar sequence, scientists should be extra careful and perhaps use a different method to confirm their findings.
• Transparency: The paper urges companies to share the exact "Wanted Posters" (probe sequences) they use. Currently, some companies keep these secret, which makes it hard for scientists to check for these mistakes.

The Takeaway

Science is a process of constant improvement. Just as a GPS might occasionally get confused by a similar-looking street name, these high-tech gene scanners can get confused by similar-looking genes.

This paper is like a mechanic telling us, "Hey, your GPS has a bug where it confuses Main Street with Main Avenue. Here is a tool to check your route before you drive, and here is how to fix the map so you don't end up in the wrong neighborhood."

By catching these errors early, scientists can ensure that the maps of our bodies they are building are accurate, reliable, and truly helpful for curing diseases.

Technical Summary

1. Problem Statement

The paper addresses a critical limitation in probe-based spatial transcriptomics, specifically the 10x Genomics Xenium platform. Xenium relies on padlock probes that hybridize to target RNA sequences, followed by ligation and rolling circle amplification (RCA) to generate a detectable fluorescent signal.

• The Core Issue: The technology assumes probe specificity. However, if a probe binds to an unintended RNA sequence (off-target binding), the resulting signal cannot be distinguished from on-target binding. This leads to the quantification of a composite signal (target + off-target), distorting the true spatial expression profile of the intended gene.
• The Gap: While 10x Genomics provides pre-designed gene panels, the specific probe sequences are often not fully disclosed or rigorously validated against all possible off-target interactions in current transcriptome annotations. Previous validation studies (e.g., Janesick et al., 2023) relied on orthogonal comparisons that showed high overall correlation but missed specific gene-level discrepancies caused by off-target binding.

2. Methodology

The authors developed a computational pipeline and applied it to validate predictions using orthogonal experimental data.

A. Software Tool: Off-target Probe Tracker (OPT)

• Function: A Python-based tool designed to identify putative off-target binding.
• Mechanism:
  • Takes probe sequences (FASTA) and aligns them against reference transcriptomes using nucmer (a maximal exact match aligner).
  • Parameters: Users can define binding strictness.
    • Perfect Homology: Requires 100% identity across the full 40bp probe.
    • Pad Mode: Allows mismatches at the terminal ends (e.g., 10bp on either side) while requiring a perfect match in the central 20bp (the ligation site), simulating biological conditions where hybridization can occur despite terminal mismatches.
  • Annotation Agnostic: Runs against multiple reference annotations (GENCODE v47, RefSeq v110, CHESS v3.1.3) to account for discrepancies in gene models, particularly regarding pseudogenes and non-coding RNAs.
  • Output: A summary of target genes and their predicted off-target binding partners.

B. Validation Strategy

The authors applied OPT to the 10x Genomics Xenium Human Breast Gene Panel (313 genes, 2,582 probes) and validated findings using three orthogonal datasets derived from the same breast cancer tissue block (Janesick et al., 2023):

1. Visium CytAssist: A sequencing-based spatial transcriptomics platform (55µm spots) used to compare spatial expression patterns.
2. 3' scRNA-seq: Single-cell RNA sequencing used to compare cell-type specific expression patterns.
3. HuBMAP Custom Panels: Applied OPT to custom panels (Placenta; Multi-organ: Kidney, Lung, Heart) and cross-referenced with tissue-specific bulk/scRNA-seq data from the Human BioMolecular Atlas Program to assess tissue-dependent impact.

C. Analysis Metrics

• Spatial Comparison: Used STalign to register Xenium and Visium images. Aggregated Xenium data to Visium resolution. Calculated Root Mean Square Error (RMSE) and Pearson Correlation between technologies.
• Single-Cell Comparison: Used Harmony for batch correction and UMAP for visualization. Calculated cluster-level expression correlations.
• Hypothesis Testing: If a gene showed high discrepancy between Xenium and orthogonal methods, the authors aggregated the expression of the target gene plus its predicted off-target genes in the orthogonal data. If the aggregated signal matched the Xenium signal better than the target alone, it confirmed off-target binding.

3. Key Results

A. Identification of Off-Target Binding

• Perfect Homology: Using GENCODE v47, OPT identified 37 genes with probes exhibiting perfect 40bp off-target matches.
• Protein-Coding Focus: When restricting analysis to protein-coding genes (excluding pseudogenes/lncRNAs) and taking the union across GENCODE, RefSeq, and CHESS, 14 genes were identified as having high-confidence off-target binding to protein-coding genes. These include: ADH1B, AKR1C1, APOBEC3A, APOBEC3B, AQP1, C1QA, CD79B, CD8B, CEACAM6, POLR2J3, S100A4, TOMM7, TPD52, and TPSAB1.
• Imperfect Matching: Allowing 10bp terminal mismatches ("pad mode") identified an additional 18 genes with potential off-target binding, including ACTG2 and TUBB2B.

B. Experimental Validation (The "APOBEC3B" Case Study)

• Discrepancy: APOBEC3B showed high expression in Xenium but was undetected (zero expression) in both Visium and scRNA-seq datasets.
• Resolution: The predicted off-targets for the APOBEC3B probe were APOBEC3D and APOBEC3F.
  • In Visium and scRNA-seq, APOBEC3D and APOBEC3F were expressed.
  • The spatial and single-cell expression patterns of the aggregated signal (APOBEC3B + APOBEC3D + APOBEC3F) in the orthogonal data closely matched the APOBEC3B pattern in Xenium.
  • Quantitative Improvement: Correlation between Xenium and Visium for APOBEC3B alone was NaN (due to zero Visium expression). When comparing Xenium to the aggregated Visium signal, the Pearson correlation became positive (0.160) and RMSE decreased, confirming the signal was likely a sum of the target and off-targets.
• General Trend: For genes with predicted off-targets, the "Target + Off-target" aggregate consistently showed higher concordance with Xenium data than the target gene alone.

C. Impact on Custom Panels

• Application of OPT to HuBMAP custom panels revealed significant off-target risks:
  • Placenta Panel: 30 of 49 genes with predicted off-targets had off-targets that were protein-coding.
  • Multi-organ Panel: 11 of 24 genes had protein-coding off-targets.
• Tissue Specificity: The impact is variable; off-target binding only distorts data if the off-target gene is expressed in the specific tissue being profiled.

D. Annotation Variability

• The study highlighted that off-target predictions vary significantly depending on the reference annotation (GENCODE vs. RefSeq vs. CHESS).
• Some probes designed based on older annotations (e.g., GENCODE v28) mapped to exonic regions in older versions but to intronic/intergenic regions in newer versions (v47), suggesting probe design may rely on outdated genomic knowledge.

4. Key Contributions

1. Development of OPT: An open-source tool for researchers to audit probe specificity before or after experiments, supporting multiple reference annotations and mismatch tolerance.
2. Empirical Evidence: Provided the first concrete evidence that specific 10x Genomics Xenium probes (e.g., APOBEC3B) are detecting off-target paralogs, leading to false-positive spatial signals.
3. Methodological Framework: Established a rigorous workflow for validating spatial transcriptomics data by integrating orthogonal spatial (Visium) and single-cell (scRNA-seq) data to deconvolve off-target effects.
4. Resource Provision: Released a comprehensive list of predicted off-target interactions for all publicly available 10x Genomics pre-designed Xenium panels.

5. Significance and Recommendations

• Data Interpretability: The findings suggest that spatial expression patterns for certain genes in Xenium data may represent a "mixture" of paralogs rather than a single gene, potentially leading to incorrect biological conclusions (e.g., misattributing function to a specific gene family member).
• Probe Design: The authors recommend that companies disclose probe sequences and that future probe design must account for the most current transcriptome annotations and tissue-specific expression of off-targets.
• Best Practices:
  • Researchers should use tools like OPT to screen panels for off-target risks.
  • For existing datasets, genes with high-confidence off-target predictions should be treated with caution or excluded from training foundation models to prevent error propagation.
  • Integrative analyses should consider aggregating target and off-target signals when comparing Xenium to other technologies.
• Transparency: The paper calls for greater transparency from commercial vendors regarding probe sequences to enable independent validation and reproducibility in the scientific community.

In summary, this paper demonstrates that off-target probe binding is a non-negligible source of error in high-resolution spatial transcriptomics, capable of fundamentally altering the interpretation of gene expression data, and provides the necessary tools and framework to mitigate these risks.