MAJEC: unified gene, isoform, and locus-level transposable element quantification from RNA-seq

MAJEC is a unified, fast Expectation-Maximization framework that accurately quantifies genes, isoforms, and individual transposable element loci from RNA-seq data by leveraging splice junction evidence to resolve read overlaps, thereby significantly reducing false attribution artifacts compared to existing tools like TEtranscripts and Telescope.

The Gist

Imagine your cell's DNA as a massive, ancient library. Inside this library, there are two types of books:

1. The Instruction Manuals (Genes): These are the active, useful books that tell the cell how to build proteins and keep you alive.
2. The "Junk" Books (Transposable Elements or TEs): These are ancient, repetitive copies of old stories scattered everywhere. They used to be able to copy-paste themselves around the library, but most are now silent. However, in diseases like cancer or during aging, these "junk" books sometimes wake up and start shouting, which can cause chaos.

The Problem: The Messy Overlap

The problem scientists face is that these "Junk" books are often printed inside the pages of the "Instruction Manuals."

Imagine a librarian trying to count how many times a specific story is being read.

• The Old Way (Tool A - TEtranscripts): This librarian is very strict. If a page has both an instruction and a junk story, they assume everything on that page belongs to the instruction. They ignore the junk story completely. This is safe, but if the junk story actually was being read, the librarian misses it.
• The Other Way (Tool B - Telescope): This librarian is the opposite. They only care about the junk stories. They don't look at the instruction manuals at all. If a page has both, they assume everything is a junk story. This leads to a huge problem: they think the junk stories are screaming loudly, when in reality, the instruction manual was just reading a sentence that happened to overlap with the junk.

The Result: Scientists were getting confused. They thought "Junk" was waking up when it was actually just the "Instructions" reading, or vice versa. They had to run two different, slow, and complicated tools to get the full picture, and even then, the answers didn't always match.

The Solution: MAJEC (The Smart Librarian)

The authors of this paper created a new tool called MAJEC. Think of MAJEC as a super-smart, unified librarian who looks at the whole library at once.

Here is how MAJEC works, using a simple analogy:

1. The "Clue" System (Splice Junctions)
When a book is read, the librarian doesn't just see the words; they see how the pages are connected.

• If a book is a standard "Instruction Manual," the pages are connected in a very specific, complex way (like a zipper).
• If a "Junk" book is being read on its own, the pages are connected differently or not at all.

MAJEC looks for these "zippers" (called splice junctions).

• If it sees the complex zipper of an Instruction Manual, it knows: "Ah, this reading belongs to the Instruction Manual, even if it's sitting on top of a Junk story."
• If it sees no zipper and just a messy pile of pages, it knows: "This is definitely the Junk story waking up on its own."

2. The "Tug-of-War" (Probabilistic Modeling)
Instead of making a rigid rule (like "Always give it to the Gene"), MAJEC runs a Tug-of-War for every single piece of evidence.

• It asks: "Does the evidence (the zipper, the strand direction) point more strongly to the Gene or the Junk?"
• It assigns the reading to whichever one has the strongest proof.
• This allows it to catch the rare moments when a Junk story is actually waking up inside a gene, without falsely blaming the gene for the noise.

Why This Matters (The Results)

The paper tested MAJEC against the old tools and found three major wins:

1. It's Accurate: MAJEC stopped the "False Alarms."

   • Example 1: The old "Junk-only" tool (Telescope) thought a specific Junk story was screaming loudly because it was sitting inside a gene that was being read. MAJEC realized, "No, that's just the gene reading," and corrected the count.
   • Example 2: The old "Gene-only" tool (TEtranscripts) missed a Junk story that was actually waking up inside a gene. MAJEC saw the lack of "zippers" and said, "This is the Junk story!" and counted it correctly.

2. It's Unified: You don't need two tools anymore. MAJEC gives you the count for the Genes, the specific versions of the Genes (Isoforms), and the specific Junk locations all in one go. It's like getting a single report card instead of three different ones.

3. It's Faster: Because it does everything in one pass through the data, it runs much faster than running the old tools separately. It's like doing your laundry, dishes, and vacuuming in one efficient trip around the house, rather than doing them one by one with different machines.

The Bottom Line

MAJEC is the new standard for reading the "messy" parts of our genetic library.

It solves the confusion caused by the fact that "Junk" and "Instructions" overlap. By using clues about how the books are stitched together, it can tell the difference between a gene being read and a virus-like element waking up. This helps scientists understand diseases like cancer and aging much better, without getting tricked by false signals.

Technical Summary

1. Problem Statement

The study of transposable elements (TEs) is critical for understanding cancer, aging, and immunity, but accurate quantification from RNA-seq data is hindered by the pervasive overlap between TEs and protein-coding genes. Existing tools suffer from significant limitations:

• TEtranscripts: Provides robust subfamily-level quantification but lacks locus-level resolution. It uses a hard-coded heuristic that assigns uniquely mapped reads overlapping both a gene exon and a TE exclusively to the gene. While effective for subfamilies, this rule fails at the locus level, suppressing genuine TE signal embedded within genes.
• Telescope: Offers locus-level resolution but operates in a "TE-only" feature space, ignoring gene annotations. Consequently, it cannot distinguish between reads from an independently transcribed TE and reads from a host gene transcript that happens to overlap a TE annotation. The authors demonstrate that this leads to massive contamination: 43% of Telescope's total TE signal originates from exon-overlapping loci, which constitute only 1.1% of all TE features.
• Workflow Fragmentation: Researchers currently must run multiple incompatible tools (e.g., Salmon/Kallisto for genes, TEtranscripts/Telescope for TEs) with different alignment requirements and annotation formats, making it difficult to reconcile gene and TE expression data.

2. Methodology: MAJEC Framework

The authors present MAJEC (Momentum Accelerated Junction Enhanced Counting), a unified Expectation-Maximization (EM) framework that jointly quantifies genes, transcript isoforms, and individual TE loci from BAM alignments in a single pass.

Key Architectural Components:

• Joint Feature Space: MAJEC constructs a unified model containing all annotated transcript isoforms (from GTF) and all annotated TE loci (from RepeatMasker). Reads overlapping both a gene and a TE compete probabilistically within this shared space.
• Junction-Informed Priors: Before the EM algorithm begins, MAJEC extracts splice junction evidence from the BAM file to modulate initial counts. This includes:
  • Junction Boost: Transcripts with unique splice junction support receive a count boost.
  • Completeness Penalty: Transcripts lacking expected splice junctions (relative to their median junction depth) are down-weighted.
  • Subset Penalty: Transcripts identified as subsets of longer isoforms are penalized unless they have unique exonic territory or specific junction evidence.
  • TSL Penalty: Optional down-weighting based on Transcript Support Levels.
• Two-Phase EM Algorithm:
  • Phase 1: Assigns uniquely mapped reads based on the prior-adjusted estimates.
  • Phase 2: Iteratively allocates multi-mapping reads using the updated probabilities from Phase 1 and previous EM iterations.
  • Momentum Acceleration: Uses expression-level grouped momentum to accelerate convergence, reducing the number of iterations required (typically 15–17).
• Output: Generates count estimates at the transcript, gene, TE locus, and TE subfamily levels, along with confidence metrics (distinguishability scores, assignment entropy, and discord scores).

3. Key Contributions

• Unified Quantification: MAJEC simultaneously resolves gene isoforms and individual TE loci in a single analysis, replacing the multi-tool pipelines currently required for joint gene-TE RNA-seq analysis.
• Probabilistic Resolution of Gene-TE Overlap: By modeling genes and TEs jointly, MAJEC replaces rigid heuristics (TEtranscripts) and blind modeling (Telescope) with a data-driven probabilistic assignment. It correctly assigns reads to the feature best supported by evidence (e.g., splice structure for genes vs. lack thereof for TEs).
• Junction-Based Isoform Accuracy: The integration of splice junction evidence into EM priors significantly improves isoform quantification accuracy, particularly for complex genes with ambiguous structures.
• Computational Efficiency: MAJEC utilizes internal multiprocessing and momentum acceleration to run faster than the combination of specialized tools it replaces, operating on standard workstations without HPC requirements.

4. Results

Isoform Quantification Benchmarks:

• Synthetic Data (Sequins): MAJEC performed on par with Salmon and RSEM (median absolute log2 fold error ~0.88–0.95).
• Real Transcriptomes (LongBench): MAJEC outperformed Salmon and RSEM on 54.3% and 53.7% of transcripts, respectively. The advantage was concentrated in transcripts with structural ambiguity (incomplete junctions or subset isoforms), where MAJEC's penalty system effectively suppressed unsupported isoforms.
• Precision vs. Sensitivity: MAJEC traded slight sensitivity for significantly higher precision, reducing false-positive transcript calls by 7,000–9,000 per cell line compared to Salmon.

TE Quantification and Contamination Reduction:

• Subfamily Concordance: MAJEC showed near-perfect agreement with TEtranscripts at the subfamily level (Pearson $r = 0.995$), validating its EM redistribution of multi-mappers.
• Locus-Level Purity: MAJEC reduced the fraction of TE signal originating from exon-overlapping loci from 43% (Telescope) to 5%.
• Differential Expression (DE) Accuracy:
  • Vignette 1 (False TE Reactivation): In the L1TD1 gene, Telescope falsely reported massive TE upregulation at HAL1ME ($log2FC = +9.1$) driven by host gene reads. MAJEC's joint model redirected these reads to L1TD1 itself ($log2FC = +14.0$), leaving the HAL1ME locus with too few counts to test — correctly recognizing the signal as genic rather than TE-derived.
  • Vignette 2 (False Gene Upregulation): In the LINC01949 gene, TEtranscripts falsely reported gene upregulation ($log2FC = +5.35$) by absorbing reads from a genuinely reactivated embedded L1PA7 element. MAJEC correctly identified the TE reactivation ($log2FC = +6.29$) and showed the host gene was unchanged.
• DE Concordance: MAJEC and TEtranscripts showed high DE concordance ($r = 0.987$ for subfamilies), while Telescope diverged significantly due to exon-overlap contamination.

Performance:

• MAJEC processed six samples in 20 minutes (wall time) using 6 cores, compared to ~5 hours for TEtranscripts and ~3 hours for Telescope (serially).
• Memory usage (~50 GB) is comparable to standard STAR alignments and fits within standard workstation limits.

5. Significance

MAJEC addresses a critical gap in RNA-seq analysis by providing a unified, accurate, and efficient framework for studying the interplay between genes and transposable elements.

• Scientific Impact: It resolves the "gene-TE overlap problem," preventing the systematic inflation of TE signals at gene boundaries (a major flaw in Telescope) and the suppression of genuine TE signals within genes (a flaw in TEtranscripts). This is crucial for studies involving epigenetic therapies, cancer, and aging where TE reactivation is a key mechanism.
• Practical Impact: It simplifies the bioinformatics workflow, replacing multiple tools with a single pipeline that produces gene, isoform, and TE locus counts simultaneously.
• Methodological Innovation: The use of junction-informed priors to drive EM convergence represents a novel approach to resolving multi-mapping reads in complex genomic regions, offering a template for future quantification tools.

In conclusion, MAJEC demonstrates that joint modeling of genes and TEs is not only feasible but necessary for accurate locus-level resolution, offering a significant improvement in accuracy and speed over current state-of-the-art tools.