Hybrid untargeted and targeted RNA sequencing facilitates genotype-phenotype associations at single-cell resolution

This paper proposes a hybrid single-cell RNA sequencing strategy that combines short-read whole-transcriptome amplification for broad coverage with long-read targeted sequencing for deep variant detection, thereby overcoming coverage limitations to enable robust genotype-phenotype associations at single-cell resolution.

The Gist

Imagine you are a detective trying to solve a mystery inside a bustling city. The city is a human body, the buildings are cells, and the people living inside are the genes. Your goal is to understand two things about every single person in the city:

1. Who are they? (What job do they do? Are they a baker, a doctor, or a firefighter?)
2. What is their secret? (Do they have a specific genetic mutation that makes them unique?)

For a long time, scientists had to choose between two very different tools to investigate, and neither was perfect on its own. This paper introduces a brilliant new strategy that combines both tools to get the full picture.

The Problem: Two Flawed Tools

Tool 1: The "Short-Read" Camera (Illumina)
Think of this as a high-speed camera that takes thousands of photos of a city, but each photo is just a tiny, blurry snippet of a street corner.

• The Good: It takes so many photos that you can count almost every person in the city. You can tell exactly which neighborhood (cell type) they live in.
• The Bad: Because the photos are so small and cut up, you can't see the whole person. If someone has a unique tattoo (a genetic mutation), this camera often misses it because the tattoo might be on the part of the body the photo didn't capture.

Tool 2: The "Long-Read" Telescope (PacBio)
This is a powerful telescope that can see the entire person from head to toe in one go.

• The Good: It sees the whole picture! If someone has a tattoo, this telescope spots it easily. It's great for finding those genetic secrets.
• The Bad: It's slow and expensive. You can only take a few photos of the city. You might miss the quieter neighborhoods entirely, and you don't get enough data to count everyone accurately.

The Solution: The "Hybrid" Strategy

The authors of this paper realized that trying to use just one tool was like trying to solve a murder mystery with only a magnifying glass or only a telescope. You need both.

They proposed a Hybrid Strategy that works like this:

1. Step 1: The Broad Sweep (Short-Read)
   First, they use the "Short-Read" camera to scan the entire city. This gives them a massive list of everyone in the city and tells them exactly who they are (bakers, doctors, etc.). It's like taking a census.

2. Step 2: The Deep Dive (Targeted Long-Read)
   Next, they take that same city and use the "Long-Read" telescope, but they don't look at the whole city. Instead, they use a special filter (a "targeted panel") to focus only on the 50 most important buildings (genes) related to their specific mystery (in this case, how the body makes hormones).

   • Because they are only looking at 50 buildings, they can zoom in incredibly deep. They get so much detail that they can find the tiniest genetic "tattoos" (mutations) even in people who are very quiet or hard to see.

The Magic of the Mix

By combining these two, they get the best of both worlds:

• From the Short-Read: They know exactly who the cells are and how many there are.
• From the Long-Read: They know exactly what genetic secrets those specific cells are holding.

The Result:
They can now say, "Ah, this specific group of 'hormone-making' cells has a specific genetic mutation that changes how they work." Before, they might have seen the mutation but not known which cell type it belonged to, or seen the cell type but missed the mutation.

Why This Matters

Think of it like trying to understand why a specific car in a traffic jam is moving strangely.

• Old way: You could either count all the cars (but miss the engine details) OR look closely at one engine (but not know if it's a Ferrari or a Toyota).
• New way: You count all the cars to see the traffic flow, and then you zoom in on the engines of the specific cars that matter to see exactly what's broken.

This paper proves that by mixing these technologies, scientists can finally link a person's genetic code directly to their behavior and function at the most detailed level possible. It's a game-changer for understanding diseases, especially those caused by rare genetic glitches in specific types of cells.

Technical Summary

1. Problem Statement

Single-cell RNA sequencing (scRNA-seq) is a powerful tool for characterizing cellular identity and heterogeneity. However, integrating genotype (genetic variants) with phenotype (transcriptional profiles) at single-cell resolution faces significant technical hurdles:

• Short-read limitations: Standard short-read scRNA-seq (e.g., Illumina) suffers from 5' and 3' biases due to library fragmentation, making it difficult to detect variants outside the tagged regions and limiting the ability to link mutations to expression.
• Long-read limitations: Long-read scRNA-seq (e.g., PacBio, Oxford Nanopore) sequences full-length transcripts, overcoming fragmentation biases. However, it typically suffers from low sequencing depth, which restricts variant calling sensitivity, particularly in lowly expressed genes, and limits the number of cells that can be genotyped.
• Targeted sequencing trade-offs: While targeted enrichment improves depth, previous approaches often relied on platforms with higher error rates (like ONT) or failed to provide a unified workflow for both cell typing and genotyping.

The authors aim to develop a strategy that combines the broad coverage of short-read sequencing with the full-length, high-depth capabilities of targeted long-read sequencing to robustly link mutational profiles with transcriptional programs.

2. Methodology

The study utilized a single adrenal gland from a patient to generate a shared single-cell cDNA library (using the 10x Genomics Chromium 3' kit). This single source material was split and processed through three distinct sequencing strategies to allow for direct comparison:

1. SR-WTA (Short-Read Whole-Transcriptome Amplification):

   • Platform: Illumina NovaSeq X Plus.
   • Method: Standard 10x Genomics library preparation with paired-end 150 bp sequencing.
   • Goal: High-throughput cell typing and broad transcriptome coverage.

2. LR-WTA (Long-Read Whole-Transcriptome Amplification):

   • Platform: PacBio Revio (HiFi reads).
   • Method: PacBio Kinnex single-cell RNA kit. This involves removing 10x TSO artifacts, incorporating biotin tags, and assembling cDNA into linear arrays for full-length sequencing.
   • Goal: Full-length transcript sequencing for unbiased variant detection across the whole transcriptome.

3. LR-Twist (Long-Read Targeted Sequencing):

   • Platform: PacBio Revio (HiFi reads).
   • Method: Hybridization-based capture using a Twist custom panel targeting 50 genes relevant to steroidogenesis (e.g., CACNA1H, CYP11B2, KCNJ5).
   • Workflow: Pre-capture PCR $\rightarrow$ Twist hybridization $\rightarrow$ Streptavidin capture $\rightarrow$ Post-capture PCR $\rightarrow$ Kinnex library preparation.
   • Goal: Deep sequencing of specific genes to maximize variant detection sensitivity.

Computational Pipeline:

• A Snakemake pipeline was developed to integrate data from all three strategies.
• SR-WTA: Processed via CellRanger (alignment, UMI counting, cell calling).
• LR-WTA/LR-Twist: Processed via Iso-Seq CLI (demultiplexing, deduplication) and aligned to both transcriptome (minimap2) and genome (for variant calling).
• Variant Calling: Combined calls from Clair3-RNA and DeepVariant on pseudobulk data, followed by filtering for coding regions and removal of common dbSNP variants.
• Cell Genotyping: Cells were classified as "mutated" if at least one read supported the alternative allele.

3. Key Contributions

• Hybrid Strategy Proposal: The authors propose a novel workflow combining SR-WTA for comprehensive cell typing and LR-Twist for deep, targeted genotyping. This leverages the high cell recovery of short-reads and the variant sensitivity of targeted long-reads.
• Benchmarking Study: A systematic comparison of SR-WTA, LR-WTA, and LR-Twist using the same cDNA source, providing a rare direct evaluation of their complementary strengths and weaknesses.
• Software Tool: Development of a modular Snakemake pipeline that accepts Illumina FASTQ and PacBio BAM files to generate expression matrices and identify variant-carrying cells.
• Target Panel Design: Demonstration of a 50-gene panel designed for steroidogenesis, showing that even small, focused panels can achieve deep saturation for rare variant detection.

4. Key Results

A. SR-WTA vs. LR-WTA (Cell Typing Performance)

• Cell Recovery: SR-WTA recovered significantly more cells (e.g., +7,683 cells in Sample A) than LR-WTA. This is attributed to deeper sequencing depth and the CellRanger algorithm's ability to recover low-RNA-content cells (e.g., steroidogenic cells) that fall below the "knee" threshold in long-read data.
• Transcriptional Consistency: Despite differences in cell counts, the gene expression profiles were highly correlated (Spearman $\rho \approx 0.825$). Cell type annotations (using CellTypist) were highly consistent (ARI = 0.977) for shared cells.
• Conclusion: SR-WTA is superior for recovering a broad range of cell types, especially those with low RNA content, while LR-WTA provides comparable expression profiles for the cells it does detect.

B. LR-WTA vs. LR-Twist (Genotyping Performance)

• Enrichment: LR-Twist achieved an approximate 70-fold enrichment for on-target genes compared to LR-WTA (38.5% of UMIs on-target vs. 0.5% in WTA).
• Variant Detection:
  • LR-Twist enabled the detection of variants in lowly expressed genes (e.g., CACNA1H) where LR-WTA failed due to insufficient read depth (3 UMIs vs. 9–15 variants detected with LR-Twist).
  • LR-Twist significantly increased the number of genotyped cells per variant, improving statistical power for downstream differential expression analysis between mutated and reference cells.
• Limitations: LR-Twist had a lower on-target rate (~30-40%) compared to some bulk targeted methods, likely due to the "off-target paradox" in low-complexity bait libraries. However, it still provided sufficient depth for robust variant calling.

C. The Hybrid Workflow

• The study demonstrates that using SR-WTA for cell typing and LR-Twist for genotyping allows researchers to link specific mutations to transcriptional states in single cells.
• LR-Twist alone is insufficient for cell typing because targeted enrichment skews expression profiles, leading to misclassification of cell types.

5. Significance

• Overcoming Depth vs. Breadth Trade-off: This work solves the critical bottleneck in single-cell genomics where researchers must choose between identifying many cell types (short-read) or detecting rare variants (long-read). The hybrid approach allows for both.
• Clinical and Biological Relevance: The ability to genotype lowly expressed genes in specific cell types (like steroidogenic cells in the adrenal gland) is crucial for understanding diseases driven by somatic mutations (e.g., adrenal tumors).
• Cost-Effectiveness: By using targeted enrichment, LR-Twist requires fewer sequencing reads than whole-transcriptome long-read sequencing, allowing multiple samples to be multiplexed on a single PacBio SMRT Cell, whereas LR-WTA typically requires one cell per sample.
• Future Applications: The framework is adaptable to other disease contexts by simply changing the targeted gene panel (e.g., cancer hotspots or specific pathway regulators), facilitating the discovery of novel variants and their impact on cellular phenotypes.

In summary, the paper establishes that a hybrid SR-WTA + LR-Twist strategy is the optimal approach for high-resolution genotype-phenotype mapping, offering the cell recovery of short-read sequencing with the variant detection power of targeted long-read sequencing.