Many human pseudogenes, gene copies that have lost the ability to make protein, are still transcribed into RNA. That RNA can regulate the parent gene, for example by soaking up the microRNAs that would otherwise repress it. Measuring this RNA is hard for a specific reason. A pseudogene often differs from its parent at only one or two bases in a hundred, and long-read sequencers make errors at about that same rate. Any read that looks pseudogene-like at one position might really be a parent read carrying a sequencing error, and the reverse is also true. Current pipelines handle this in sequence. First they clean the reads or align them. Then they divide ambiguous reads among the candidate genes. Each step decides about each site, or each read, on its own.
Grammatical error correction ran into the same structural problem, and its lesson carries over. In grammar, errors come in coupled groups: a wrong noun number forces a wrong verb, and one error hides the evidence for its neighbour. Fixing them one at a time gives corrections that do not fit together. We propose to treat a read's pseudogene-versus-parent decision the same way. Every diagnostic position on a read is evidence about one shared hidden cause, the read's origin. Sequencing errors are modelled as bursts that are correlated along the read, not as independent per-base noise. The method also explicitly allows an origin that switches partway along the read, as happens with chimeras or gene conversion, where a stretch of one gene's sequence has been copied into the other. The claim is that coupled modelling of this kind, run inside quantification, cuts false pseudogene expression ("homology echoes") and stabilises low-abundance estimates, where existing methods are known to degrade.
Where the two directions meet, and what is already done
Three ideas from grammatical error correction motivate this work. The first is the masking effect: Fluency Boost Learning and Inference feeds a model's own corrected output back in for repeated passes, because fixing an obvious error exposes subtler ones. The second is that joint inference beats independent decisions. Joint Learning and Inference for GEC raised subject-verb agreement F1 from 39.64 to 52.25 on CoNLL-2013 by enforcing agreement constraints globally. GEC Using Integer Linear Programming cast the corrections in a sentence as one optimisation with pairwise interaction terms. The third idea is a negative result that constrains our design: the joint model of Joint Learning and Inference for GEC was reliable only across one to three words and needed an ad hoc reliability filter. Long reads are useful here because they deliver evidence linked across kilobases. Our model must therefore show that it does not degrade with the distance between diagnostic sites, as those grammar models did.
On the genomics side, four papers come closest to this proposal, and none of them does it.
Aldy separates pharmacogenes such as CYP2D6 from their pseudogenes. It uses integer linear programming over DNA copy numbers. Its integer, uniform-coverage assumptions do not carry over to RNA, where abundance is continuous and spans orders of magnitude. For that reason we keep a probabilistic, EM-style quantifier and do not port integer linear programming directly.
DeChat corrects nanopore DNA reads without collapsing repeats and paralogs. It sorts reads into variant-consistent piles over three passes. It is a pre-correction step built on local heuristic clustering, and it assumes even DNA coverage. It never asks which gene a read came from at the quantification stage.
Bambu quantifies long-read transcripts with EM and handles repeats such as HERVH. It corrects splice junctions as a separate preprocessing step, and its read-to-transcript compatibility does not ask whether a mismatch is an error or a real divergence.
Clair3-RNA shows that adding phase information, meaning which parental copy of a chromosome a read comes from, raises RNA variant-calling F1 by about 2%. It also reports nanopore indel recall below 60%, because errors cluster. It works on windows of stacked reads at each site and never assigns a read to a gene family member.
Our proposal differs from all four. Origin assignment, a correlated-error model, and abundance estimation are fit together as a single likelihood for each gene family, on RNA.
Core Claim
Consider a family $F$ made of a parent gene and its pseudogenes, aligned to one another. Let $s_1,\dots,s_m$ be the paralogous sequence variant positions covered by read $r$. These are the positions where family members differ, so they are where the diagnostic evidence lies. The read's likelihood under the family is
$$ p(r \mid F) = \sum_{z_{1:m},\, e_{1:m}} \pi(z_1)\prod_{j=1}^{m} p(o_j \mid z_j, e_j, c_j)\prod_{j=2}^{m} q_\rho(z_j \mid z_{j-1})\, q_\lambda(e_j \mid e_{j-1}, d_j). $$
The symbols are as follows. $z_j \in F$ is the family member the read derives from at site $j$. $e_j$ is a binary "error-burst" state, and $o_j$ is the base or indel observed at that site. $c_j$ is the local sequence context, such as homopolymer length. $d_j$ is the distance from the previous site. $\pi$ holds the family member abundances being estimated. $q_\rho$ allows a rare switch of origin with probability $\rho$, to represent chimeras or conversion tracts. $q_\lambda$ lets error bursts persist over short distances. The model is a small hidden Markov model with $|F|\times 2$ states, so exact inference grows linearly with the number of sites. This avoids the combinatorial blow-up of higher-order grammar ILPs.
Three ingredients are new compared with an alignment-score EM. The first is the burst state, which captures the clustered errors that Clair3-RNA's results expose. The second is the switching origin, which tells chimeric reads apart from erroneous ones. The third is a re-alignment step that copies the fluency-boost loop: once a read's most likely origin is known, it is realigned to that template and scored again, so an early indel no longer misplaces the downstream diagnostic sites. A weak shrinkage prior on $\pi$, shared across samples of the same tissue, steadies estimates below 2 TPM (transcripts per million).
Research Program
We will build gene families from the GENCODE pseudogene resource. It catalogues 11,216 pseudogenes, at least 876 of them transcribed, and links about 86% of them to parent genes. We align each family and list its diagnostic sites.
We will fit the emission model $p(o_j\mid z_j,e_j,c_j)$ for each platform. The training data are reads at positions with confident genotypes in HG002, as used by DeChat and Clair3-RNA. Known RNA editing sites from REDIportal are excluded so that editing is not learned as error.
The quantifier sits as a module after alignment. It works only on reads that touch a family, and it hands its abundances back to Bambu for the rest of the transcriptome. Because it works on whole families, it can be dropped in without changing anything else, and it can be compared directly with Bambu's own EM.
We will analyse the model along three axes. The first is accuracy as a function of the read's error rate relative to the family's divergence. The second is accuracy against the distance between diagnostic sites, which tests the distance-degradation failure seen in grammar correction. The third is accuracy at very low abundance.
Baselines
The naive reference counts only reads that map uniquely to the pseudogene with high mapping confidence and discards the rest. This is common practice. It should be conservative, but it is blind to pseudogenes with very high identity to the parent. Any method that cannot beat it on the decoy false-positive rate described below is worthless.
The strongest existing method is Bambu's EM quantification, run unchanged on the same alignments and annotation. For context, the LRGASP assessment reports Spearman correlations of 0.74–0.87 against cell-mixing ground truth for IsoQuant, FLAIR and Bambu, with failure below 2 TPM. Separately, the RNA-seq quantification benchmark reports standard deviations of about 0.5 in low-abundance proportions between identical replicates. No reported result was found on Lacuna for pseudogene-specific quantification accuracy. We will therefore produce Bambu's number on our own benchmarks and not borrow one. IsoQuant and FLAIR will be run as secondary comparisons. All tools get the same reads, reference and compute, and default settings plus an equal grid over each tool's main thresholds.
The sequential-correction baseline runs DeChat on the RNA reads and then Bambu. It represents the "clean first, then quantify" approach.
Ablations isolate each new ingredient. The first removes burst states, which leaves an EM restricted to diagnostic sites with independent errors. The second removes origin switching. The third removes the re-alignment pass. The fourth removes the shared prior.
Beating Bambu means three things. The false-positive pseudogene expression on decoys must be at least halved. Spearman correlation on simulated families must improve in the band below 2 TPM. The difference between technical replicates must fall below Bambu's in that same band. Mid- and high-abundance accuracy must not get worse.
Why not just…?
Why not just sequence with PacBio HiFi or ONT R10 and trust the aligner? Lower error rates help, but the remaining errors are not uniform. Nanopore RNA indel recall stays below 60% even in Clair3-RNA, and frameshift indels are exactly the changes that define many processed pseudogenes. Low abundance remains a problem whatever the error rate. We will test this directly by running every method on the LRGASP PacBio data alongside the nanopore data. If the gain over Bambu disappears on the accurate platform, we will report the method as a nanopore-specific tool.
Why not just run standard EM, which already allocates multi-mapping reads? This is the main threat. Bambu's EM already uses the whole read, and the first ablation (diagnostic-site EM with independent errors) shows how much of any gain comes merely from focusing on diagnostic sites. If that ablation matches the full model, then correlated errors and origin switching are not needed. We would then publish the simpler method and drop the coupling claim. The replicate-instability results suggest that EM run naively over whole alignments is fragile at low counts, but they do not show that coupling is the cure. Only our ablation can show that.
Why not just pre-correct reads with a repeat-aware tool like DeChat? The error-correction benchmark shows that k-mer correctors over-correct rare true variants. DeChat avoids this in DNA by clustering reads, but its clustering relies on coverage, and a pseudogene read that is a hundred times rarer than its parent may never form a pile of its own. The DeChat-then-Bambu baseline measures this directly through the recall of real pseudogene reads after correction.
Proposed Experiments
We will run three experiments, each with its own form of ground truth.
The first is a simulation. We generate reads from GENCODE families at controlled abundances, including chimeric reads. Errors are resampled from real SG-NEx and LRGASP reads at matched sequence contexts, so the clustering in the simulated errors is realistic. We report per-family Spearman correlation and false-assignment rates, binned by abundance, divergence and the distance between diagnostic sites.
The second, the decoy test, is the decisive one on real data. We add synthetic "pseudogenes" to the reference. Each is made by mutating a parent at that family's real divergence rate, so it cannot exist in the sample. Every read assigned to a decoy is a measured homology echo. We run this on SG-NEx and LRGASP human samples.
The third is a real-data consistency test. We check agreement between replicates and between platforms on the LRGASP samples, using the cell-mixing ground truth where it covers family members. We also check that per-strain expression patterns in the mouse data of Sisu et al. stay consistent across replicates.
The project succeeds if the decoy false-positive rate and the low-TPM replicate error both beat Bambu, and the burst-state ablation shows that the correlated-error model is responsible.
References
- Fluency Boost Learning and Inference for Neural Grammatical Error Correction
- Joint Learning and Inference for Grammatical Error Correction
- Grammatical Error Correction Using Integer Linear Programming
- Allelic decomposition and exact genotyping of highly polymorphic and structurally variant genes
- Repeat and haplotype aware error correction in nanopore sequencing reads with DeChat
- Context-aware transcript quantification from long-read RNA-seq data with Bambu
- Clair3-RNA: a deep learning-based small variant caller for long-read RNA sequencing data
- The GENCODE pseudogene resource
- Systematic assessment of long-read RNA-seq methods for transcript identification and quantification
- A benchmark for RNA-seq quantification pipelines
- Benchmarking of computational error-correction methods for next-generation sequencing data
- Transcriptional activity and strain-specific history of mouse pseudogenes