Sample-specific haplotype-resolved protein isoform characterization via long-read RNA-seq-based proteogenomics

This paper presents an end-to-end workflow that integrates long-read RNA sequencing with mass spectrometry to construct haplotype-resolved, sample-specific proteome databases, enabling the detection of allele-specific protein isoforms and linked variants that are missed by traditional reference-based approaches.

The Gist

The Big Picture: The "Recipe" Problem

Imagine your body is a massive, bustling kitchen. Inside this kitchen, there are millions of chefs (cells) making dishes (proteins). To know exactly what dishes are being made, scientists use a technique called Mass Spectrometry. Think of this like a food critic who takes a bite of a dish, breaks it into tiny crumbs (peptides), and tries to guess the full recipe based on those crumbs.

The Problem:
Usually, the food critic uses a single, generic "Master Recipe Book" (the Reference Proteome) to guess the dishes. But here's the catch:

1. Everyone's kitchen is different: Your DNA has tiny typos (variants) compared to the Master Recipe Book.
2. Chefs improvise: Sometimes, chefs skip a step or add an extra ingredient (alternative splicing), creating a dish that isn't in the book at all.
3. The "Twin" Confusion: You have two copies of every recipe (one from Mom, one from Dad). Sometimes, the "Mom" copy has a typo, and the "Dad" copy is perfect. The Master Recipe Book doesn't know which copy is which, so it just lists the "average" dish.

Because the Master Recipe Book is incomplete and generic, the food critic often misses unique dishes or misidentifies them.

The Solution: A Custom, "Haplotype-Resolved" Cookbook

The authors of this paper built a new tool that creates a custom, personalized recipe book for the specific sample being tested. They call this a "haplotype-resolved proteome."

Here is how they did it, step-by-step:

1. Reading the Full Instructions (Long-Read RNA-seq)

Instead of reading the recipe in tiny, confusing snippets, they used a new technology called Long-Read RNA sequencing.

• The Analogy: Imagine trying to read a novel by looking at a few scattered words on a page (short-read sequencing). It's hard to know which sentence the words belong to. Long-read sequencing is like holding the whole page in your hand. You can see the entire sentence structure and, crucially, you can see exactly which typos belong to that specific sentence.

2. Sorting the "Mom" and "Dad" Copies (Phasing)

Once they have the full sentences, they need to figure out which typos belong to the "Mom" copy and which belong to the "Dad" copy. This is called Phasing.

• The Analogy: Imagine you have two identical-looking suitcases (the two chromosomes). One has a red sticker on the handle, and the other has a blue one. If you see a red sticker on a shirt inside the suitcase, you know that shirt belongs to the "Red Suitcase." The paper tested different algorithms (like WhatsHap) to see which one was best at sorting these suitcases. They found that WhatsHap was the most accurate sorter.

3. Building the Custom Database

They took the sorted "Mom" and "Dad" recipes, combined them with the actual dishes the chefs were making (the RNA data), and built a Personalized Database.

• The Result: This database doesn't just say "We have a protein called Insulin." It says, "We have Insulin-Mom (which has a slight typo) and Insulin-Dad (which is perfect)."

4. The Taste Test (Mass Spectrometry Search)

Finally, they took the real food crumbs from the lab and searched them against this new, custom database.

• The Discovery: Because the database was so specific, they found "flavors" (peptides) that the old Master Recipe Book completely missed. They could identify:
  • Variant Peptides: Dishes with the specific "Mom" or "Dad" typos.
  • Splice Peptides: Dishes where the chef skipped a step.
  • Linked Variants: Even if they couldn't taste a specific typo directly, they could infer it was there because it was "linked" to another typo they did taste on the same "suitcase."

Why Does This Matter? (The Real-World Impact)

The authors tested this on two scenarios:

1. A Stem Cell Line (WTC11): They mapped out the complex "menu" of this cell line, showing that most of the complexity comes from genetic typos, not just different cooking styles.
2. Stem Cells Turning into Bone Cells: They watched a stem cell turn into a bone cell over time. They could see how the "Mom" and "Dad" versions of proteins changed as the cell differentiated.

The Takeaway:
This paper proves that we can stop using a generic, one-size-fits-all recipe book for biology. By using long-read technology to sort out the "Mom" and "Dad" versions of our genes, we can build a custom menu for every individual. This allows scientists to see the true, unique dishes being made in our bodies, which is crucial for understanding diseases, developing personalized medicines, and figuring out why some people react differently to treatments than others.

In short: They built a better map. Instead of a blurry, generic map of the world, they created a high-definition, GPS-enabled map that shows exactly which roads (genes) are open and which are closed for your specific journey.

Technical Summary

1. Problem Statement

Traditional bottom-up mass spectrometry (MS) proteomics relies on searching peptide spectra against reference protein databases. These databases typically assume a single reference sequence per isoform, failing to capture:

• Genetic Variation: Individual-specific variants (SNVs, indels) that alter amino acid sequences.
• Allele-Specificity: The co-occurrence of multiple variants on the same chromosome (haplotypes), which creates distinct protein sequences not represented in standard references.
• Transcriptomic Complexity: Alternative splicing events that generate novel isoforms not present in reference annotations.

Existing proteogenomic workflows often address these issues independently or rely on population-based phasing (using reference panels) rather than sample-specific data. Furthermore, short-read RNA-seq cannot resolve full-length transcript structures or link distant variants on the same molecule, limiting the ability to construct accurate, sample-specific, haplotype-resolved proteomes.

2. Methodology

The authors developed an end-to-end Snakemake pipeline that integrates long-read RNA sequencing (lrRNA-seq) with MS data to construct and search haplotype-resolved, sample-specific proteomes.

Workflow Steps:

1. Data Input: Requires matched lrRNA-seq (PacBio Iso-Seq) and MS data, along with a reference genome and transcriptome.
2. Preprocessing & Alignment:
   • lrRNA-seq reads are trimmed (polyA, adapters) and aligned to the genome using minimap2.
3. Variant Calling & Phasing:
   • Variants are called using Clair3-RNA.
   • Phasing: Variants are phased using WhatsHap (identified as the top performer in benchmarks) to determine which variants reside on the same allele.
   • Filtering: Variants are filtered to ensure they form contiguous phase blocks across Coding Sequences (CDS).
4. Transcript Discovery:
   • Novel transcripts are discovered using Bambu.
   • Open Reading Frames (ORFs) are predicted for novel isoforms using ORFanage.
5. Proteome Construction:
   • Haplosaurus is used to map phased genetic variants onto reference and novel transcripts to generate protein haplotypes.
   • This creates a database containing:
     • Reference isoforms.
     • Novel splice isoforms.
     • Allele-specific variants (homozygous and heterozygous).
     • Linked variants inferred through phasing.
   • Decoy sequences are generated using DecoyPYrat for False Discovery Rate (FDR) estimation.
6. MS Search & Inference:
   • The custom database is searched against MS data using sage.
   • Protein inference is performed using a greedy set-cover approach (RcppGreedySetCover).
   • Post-search annotation links identified peptides back to specific haplotypes and splice structures.

Benchmarking:
The authors benchmarked phasing algorithms (WhatsHap, HapCUT2, Margin, phASER, HiPhase) on Genome-in-a-Bottle (GIAB) HG002 and HG005 PacBio lrRNA-seq data. WhatsHap was selected for the pipeline due to its superior balance of low switch error rates and high completeness in phasing full CDS regions.

3. Key Contributions

• First End-to-End Workflow: The first pipeline to construct and search haplotype-resolved proteomes directly from sample-matched lrRNA-seq and MS data.
• Integration of Phasing and Splicing: Uniquely combines allele-specific genetic variation with full-length transcript structures, allowing for the detection of protein isoforms defined by both specific splice patterns and linked genetic variants.
• Practical Framework: Provides a modular, reproducible Snakemake workflow that leverages existing tools (Clair3, WhatsHap, Bambu, Haplosaurus, Sage) while integrating them into a cohesive proteogenomic pipeline.
• Demonstration of Feasibility: Proves that lrRNA-seq data can be used not just for variant calling, but for accurate read-based phasing suitable for proteogenomics.

4. Key Results

Application to WTC11 Cell Line:

• Database Composition: The custom database contained ~66,000 protein isoforms.
  • 84.5% matched GENCODE reference.
  • 15.2% contained non-silent genetic variants.
  • 0.3% exhibited alternative splicing relative to the reference.
  • Only 0.1% of isoforms contained both alternative splicing and genetic variation, suggesting these processes are largely independent in this sample.
• MS Search Performance:
  • The sample-specific database identified 10,757 protein groups, slightly more than GENCODE (10,702) and UniProt (10,585).
  • Unique Discoveries: The sample-specific search identified 364 unique protein groups not found in GENCODE, primarily driven by novel sequences containing genetic variants.
• Variant Detection:
  • Direct Identification: 538 variants were directly confirmed by peptide evidence (258 homozygous, 280 heterozygous).
  • Linked Identification: Phasing enabled the inference of 1,072 homozygous variants and 201 heterozygous variants that were not directly spanned by a peptide but were linked to a directly identified variant on the same haplotype.
  • Indels: Linked identification significantly improved indel detection (32 linked indels vs. 3 direct indels), as linkage requires only one SNV on the haplotype to be covered, whereas direct indel detection requires full peptide coverage of the indel.

Application to iPSC-to-Osteoblast Differentiation:

• The workflow successfully tracked allele-specific expression changes during differentiation.
• Example: The DSP gene showed increased expression in iPSCs, with both haplotypes (A and B) detected. The method confirmed the presence of specific heterozygous SNVs in the protein products, demonstrating the ability to resolve allele-specific dynamics in time-course experiments.

5. Significance and Future Outlook

• Precision Proteomics: This approach moves proteomics from a "reference-centric" view to a "sample-centric" view, enabling the detection of disease-relevant or individual-specific protein isoforms that are invisible to standard databases.
• Feasibility of Read-Based Phasing: The study validates that PacBio lrRNA-seq provides sufficient read length and accuracy to phase variants within coding regions, a critical step for constructing accurate allele-specific proteomes.
• Clinical & Biological Applications: The framework is applicable to studying:
  • Neoantigens: Identifying patient-specific tumor antigens derived from haplotypes.
  • Rare Diseases: Characterizing protein products of rare variants in undiagnosed patients.
  • Dynamic Systems: Tracking how allele-specific expression changes during development or in response to stimuli.
• Limitations & Future Work: The authors note limitations regarding the depth of lrRNA-seq (affecting detection of rare heterozygous combinations) and the current lack of support for altered start codons. Future work will focus on extending this to neoantigen prediction, allele-specific expression (ASE) quantification, and refining protein inference algorithms for the increased complexity of haplotype-resolved searches.

In conclusion, this paper establishes a robust, practical framework for lrRNA-seq-based proteogenomics, demonstrating that integrating sample-specific genetic phasing with transcriptomics significantly enhances the resolution and accuracy of protein isoform characterization.