Letting cell identity decide where ambiguous single-cell reads belong

In single-cell RNA sequencing, a single cell typically yields only a few thousand short sequence fragments ("reads"), each tagged with a cell barcode and a molecule counter. Each read is matched back to the genome to decide which gene it came from. That step breaks down for genes that have near-identical copies elsewhere in the genome. These include duplicated gene families such as the Fc-gamma receptors and HLA genes, and the roughly 2,000 inactive copies of ribosomal protein genes, called pseudogenes. A read from one copy matches several places equally well. Current pipelines handle this in one of two ways. They throw such reads away, which makes these genes invisible in exactly the immune and housekeeping families biologists care about. Or they split the reads blindly, which invents expression of silent copies.

This proposal rests on a simple observation. The near-identical copies are rarely expressed in the same cells. A paralog (a duplicated gene copy) may be on in monocytes and off in B cells, and a pseudogene may be transcribed in one cell type of one tissue. If we know what kind of cell a read came from, we have a strong prior on which copy produced it. We propose a statistical method that allocates ambiguous reads using expression patterns pooled across cells of similar identity. We pair it with a ground-truth benchmark built by trimming full-length single-cell reads down to the short tags that droplet platforms produce. The method is judged on one question: can cell-type pooling recover copy-specific expression that is otherwise erased or fabricated, and when does the data genuinely not contain the answer?

Core Claim

Which copy a multi-mapping read came from is best estimated from the expression of each candidate copy within the read's cellular neighborhood. Estimating it per cell has too few reads. Estimating it over the whole dataset averages over cell types that use different copies. For each family of near-identical genes, let $L_r$ be the set of loci that read $r$ is compatible with, $c$ its cell, and $k(c)$ a soft neighborhood of transcriptionally similar cells. We model

$$ P(g \mid r, c) \propto \theta_{k(c),g}\, p(r \mid g), \qquad \theta_{k,g} = \lambda\, \hat\theta^{\text{nbhd}}_{k,g} + (1-\lambda)\, \hat\theta^{\text{global}}_{g}, $$

The terms in this model are defined as follows:

  • $p(r \mid g)$ is the alignment likelihood of the read under locus $g$. It reflects mismatches at the few positions where the copies differ.
  • $\hat\theta^{\text{nbhd}}_{k,g}$ is the share of the family's molecules attributed to locus $g$ in neighborhood $k$, estimated by expectation-maximization (EM). EM alternates between soft read assignment and re-estimating those shares, using both uniquely mapping and ambiguous molecules.
  • $\lambda$ is a shrinkage weight toward the global share. It is learned per family from how much the unique-read shares actually vary across neighborhoods.

There is one guard against circularity. Neighborhoods are built from a cell graph that excludes every gene in any homologous family being corrected. The allocation therefore cannot reinforce clusters that its own output created.

The second claim matters as much as the first. For each family and neighborhood, the method reports whether the data can resolve the allocation at all. A 3′-tag protocol may never sequence the positions that distinguish two copies, and then no prior can recover the truth. In that case the method should return the family-level total with an "unresolvable" flag rather than a confident split.

Relation to Prior Work

Two bulk studies motivate the gap. Tonner et al. detected ribosomal protein pseudogene transcription across 16 bulk tissues of the Illumina Human Body Map 2.0. They used a composite genome index, zero-mismatch alignment, discarding every multi-mapper, and "uniqueome masking" (masking regions that are not unique within four mismatches over 75 bp). They confirmed 17 pseudogenes with RPKM > 5, including an RPL21 pseudogene at RPKM 170 in thyroid. They also acknowledged that strict filtering underestimates how many pseudogenes are transcribed. Bulk tissue cannot say which cells carry that signal. Peng et al. needed hand-refined gene models to separate FCGR2A from FCGR2C (up to 98% identity) in GEUVADIS lymphoblastoid RNA-seq. They found protective SNPs associated with the final FCGR2A exon at p = 2.1×10⁻⁹ to 2.6×10⁻¹¹, but noted that the receptors act in myeloid subsets their bulk cell-line data could not see. Both papers show that the homology problem is real and that it is tied to cell type. Neither has the single-cell resolution needed to test that link.

Zhang et al. constrain the design. They found 258 ribosomal protein pseudogenes with no stop codons or frameshifts and near-full length relative to their parents. Any allocation method that leaks reads from a highly expressed parent gene will "find" these as expressed genes. Silent pseudogenes are therefore our main false-positive test. Aldy resolves CYP2D6/CYP2D7 structure by integer programming over read depth in bulk DNA. It shows that homology can be resolved combinatorially when coverage is deep and structures are catalogued. Single-cell transcript tags have neither property, so we use a probabilistic soft allocation instead.

The method itself borrows from ELLIPSIS. ELLIPSIS quantifies splicing in sparse single-cell data by pooling coverage across transcriptionally similar cells within a splice graph, with average exon errors generally below 5% on simulations. It resolves ambiguity only inside one gene, after reads have already been assigned to that gene. We carry its pooling principle to ambiguity between genes, which is the step ELLIPSIS takes as given.

The downstream motivation comes from deconvolution methods, which estimate cell-type composition from bulk tissue. POPPY aligns bulk and single-cell profiles through ontology-aware contrastive learning. Count Bridges deconvolves bulk counts into integer single-cell profiles. DECONbench benchmarks proportion estimation. All three take the gene count matrix as given. We do not propose a new deconvolution method. We only test whether single-cell references with corrected homolog counts change their outputs for marker genes that come from homologous families.

Research Program

The first component is the allocator. It takes reads already aligned with all multi-mapping hits retained. It groups loci into homology families by shared multi-mapping molecules, and runs the neighborhood-shrunk EM above at the level of molecules rather than reads, so that duplicate copies of the same molecule count once. A fully per-cell variant ($\lambda$ applied at cell level) and a fully global variant ($\lambda = 0$) fall out of the same code. They serve as the key ablation.

The second component is the identifiability report. For each family we compute, from the reference sequences and the protocol's read geometry, the number of distinguishing positions that reads can actually cover. We also estimate a posterior spread for each neighborhood's share by bootstrapping over cells. Families with no coverable distinguishing positions are reported only as totals.

The third component is the benchmark. Full-length Smart-seq2 data such as the Darmanis glioblastoma set used by ELLIPSIS lets most molecules from homologous families be placed uniquely using the full read. We take those placements as reference truth. We then rebuild the data as 3′-tag-like 50–100 bp reads anchored at transcript ends and rerun every method. Allocation error is measured against the full-length placement. Because the reference truth comes from the same cells, it keeps real cell-type structure. A purely simulated benchmark would have to invent that structure.

Baselines

The naive reference point is unique-only counting. Every molecule with more than one best locus is discarded, which is the single-cell analogue of Tonner et al.'s filtering. A second naive baseline splits each ambiguous molecule uniformly across its candidate loci.

The strongest existing-style competitor is EM allocation without cell-type pooling, as done by bulk quantifiers. We run it two ways. The first is per cell, as some single-cell quantifiers do. The second is global, pooling all cells into one set of shares.

A further strong competitor is cluster pseudobulk EM. Cells are hard-clustered on the non-homologous genes, and bulk-style EM is run within each cluster. This is the obvious simple version of our idea.

All methods receive identical alignments with all hits retained, identical cell neighborhoods or clusters, and identical molecule deduplication. The tuning budget is matched: our $\lambda$ and neighborhood width are chosen on one held-out family set, and cluster resolution for the pseudobulk baseline is tuned on the same set.

We searched for a reported number on this exact task. There is no reported result found on Lacuna for allocation accuracy of multi-mapping molecules between paralogs or pseudogenes in single-cell data. The closest reported figures are on different tasks: ELLIPSIS's below-5% average exon error for within-gene splicing, and Tonner et al.'s count of 17 confidently expressed ribosomal protein pseudogenes under strict filtering. We will therefore establish the baselines' numbers ourselves.

Beating global and per-cell EM means lower allocation error on the truncation benchmark, with gains concentrated in families whose unique-read shares differ across neighborhoods. It also means no increase in false expression of silent pseudogenes. Beating cluster pseudobulk EM means a gain on rare or continuous cell states, where hard clusters merge populations.

Why not just…?

Why not just discard multi-mapping reads? Tonner et al. showed that this avoids false echoes but confines detection to a small fraction of candidate pseudogenes. In 3′-tag data with 10–20% capture efficiency, families such as FCGR and HLA lose most of their molecules. The truncation benchmark measures this directly as the recall of unique-only counting against full-length truth, family by family.

Why not just split reads uniformly or at random? Zhang et al.'s 258 intact pseudogenes guarantee that this produces confident-looking expression of silent copies whenever the parent is highly expressed. We will show this explicitly with the silent-pseudogene false-positive test described below.

Why not just run EM per cell or globally, as bulk tools do? This is the real threat. If cell types do not differ in which copy they use, global EM is optimal and our pooling adds nothing. The learned $\lambda$ makes this testable. If $\lambda$ collapses toward zero for most families and error gains vanish, the core claim is false, and we will report it. Per-cell EM should fail on sparse cells, and the benchmark stratifies error by cell depth to show where.

Why not just cluster and run bulk EM in each cluster? This baseline may capture most of the gain. If it matches our method, the contribution shrinks to the benchmark and the identifiability report. We would accept that outcome rather than add complexity. The test is whether soft neighborhoods help for rare subsets and continuous states.

Why not just use long-read single-cell sequencing? Long reads resolve many such loci, as targeted long-read PKD1 sequencing showed for bulk DNA. They are costlier per cell, however, and cannot reprocess existing short-read cohorts such as Onek1k's 10⁶ cells or the 3CA atlas. Where matched long-read data exist, they serve as an additional reference truth rather than a competitor.

Proposed Experiments

The decisive experiment is the truncation benchmark. We report allocation error (mean absolute error of each locus's share per neighborhood against full-length truth) and per-cell copy-specific count correlation. Results are stratified by family, sequence identity, and cell depth. Every result is also split by the identifiability flag. If error is low on flagged-unresolvable families, the flag is wrong. If error is high on resolvable ones, the allocator is wrong.

The second check is false positives on pseudogenes, run on 10x PBMC data and on Onek1k. Ribosomal protein pseudogenes that showed no expression in any of the 16 Human Body Map tissues should receive near-zero allocated counts. We report the fraction of parent-gene molecules leaked into them by each method. Tonner et al.'s 17 expressed pseudogenes serve as candidate positives. In blood, their absence is not an error, so these are used only where a matching tissue is available.

The third check is biological plausibility in the FCGR cluster on Onek1k. Allocated FCGR2A, FCGR2B, and FCGR2C expression should separate B cells from monocyte subsets in the directions known from cell-surface biology. Unique-only counts serve as the comparison. If genotypes are available for the cohort, we will test whether the Peng et al. SNPs associate with allocated FCGR2A expression specifically in B cells and myeloid cells. This would extend their bulk lymphoblastoid result to the cell types they could not examine.

The final check addresses downstream consequences. We build DECONbench-style in silico mixtures from single cells, with known proportions. References are built two ways, from unique-only counts and from allocated counts. We then run a standard reference-based deconvolution and a reference-free NMF method on both, measuring proportion error. DECONbench's best RNA method reached a mean absolute error of 0.024 on its PDAC mixtures. That is the scale at which a homology-driven change would need to matter. If the corrected references do not change proportion error for cell types with homologous markers, we will report that the upstream correction matters for gene-level biology but not for composition estimates.