A bulk RNA sequencing profile of a tumor is a smoothie of many cell types. "Deconvolution" methods unmix it and estimate what fraction of the sample each cell type makes up. These methods work at the level of whole genes, and that hides an important event. One gene can produce several transcript versions, called isoforms, through alternative splicing (cutting and joining different exons). Different isoforms can make proteins that end up in different places in the cell and do opposite things. In lung tumors, myeloid cells switch FLT-1 from a soluble decoy receptor to a membrane-bound signaling receptor. Bulk tumor sequencing missed this switch; only physically sorting the cells revealed it (Durrans et al.).
We propose to recover such switches computationally from ordinary bulk tissue RNA. The method estimates, for each cell type and each disease state, which isoforms of a gene that cell type expresses. It then checks whether the exons that differ between those isoforms carry the signals that decide where a protein goes in the cell: membrane anchors, nuclear import or export signals, and modification sites such as the phosphorylated serines and glycosylated asparagines that gate trafficking. The output is a short, ranked list of claims of the form "in tumor myeloid cells, gene G switches to an isoform that loses its membrane anchor." Biologists can test each one directly.
Core Claim
The aim is not better cell-fraction estimates. Cell fractions estimated from gene-level data are probably adequate, and we treat them as an input. The new object is a matrix of isoform proportions per cell type, allowed to differ by condition. We claim three things. First, it can be estimated from bulk short-read cohorts, given a small amount of long-read calibration and correction for cell size. Second, it recovers known cell-type-specific switches that gene-level analysis and naive bulk splicing analysis miss. Third, only a small fraction of the recovered switches change localization-relevant sequence, and that fraction is where biological follow-up should go.
The model for the count of isoform $i$ of gene $g$ in sample $s$ is
$$ \mathbb{E}[Y_{is}] = \sum_{c} w_{cs}\, r_c\, X_{gc}\, \pi_{ic}(z_s), $$
where $w_{cs}$ is the fraction of cell type $c$ in sample $s$, and $r_c$ is a total-RNA-per-cell factor for that cell type. The term $X_{gc}$ is the gene-level reference expression in cell type $c$, and $\pi_{ic}(z_s)$ is the share of gene $g$'s transcripts from isoform $i$ in cell type $c$ under condition $z_s$ (for example, tumor versus normal). The $\pi$ values for each gene and cell type sum to one. The fractions $w$ come from a gene-level deconvolution step. The size factor $r$ follows the correction motivated by CSsingle, which showed that ignoring RNA content biases estimates toward large cells. The unknowns $\pi$ are fit gene by gene across samples by constrained Poisson regression. This works because cohorts vary in composition: if a gene's isoform ratio in bulk tracks the myeloid fraction, the model can attribute that ratio to myeloid cells.
Research Program
The work has three components.
The first is quantification. Short reads assign reads to isoforms poorly. MPAQT is a Poisson model that combines short-read depth with long-read connectivity. On exon inclusion, short reads alone reached only $r = 0.33$ against RT-qPCR, while the joint model reached $r = 0.81$. We reuse MPAQT's likelihood. We feed its per-isoform uncertainty into the regression above, so ambiguous isoforms are down-weighted rather than trusted. We also measure how few long-read samples a cohort needs before $\pi$ estimates stabilize.
The second is the localization annotation layer, applied only to switches that pass a significance threshold. For each switched isoform pair, we map the differing exons onto protein coordinates. We then record two kinds of change. One is the gain or loss of annotated localization features: transmembrane helices, signal peptides, and nuclear import or export signals. The other is the gain or loss of modification sites predicted by a MeToken-style model, which predicts modification sites from the local 3D neighborhood of each residue. This annotation layer comes from the literature on single-residue gating. Phosphorylated Best1 S358 must bind 14-3-3γ before the channel reaches the membrane (Oh et al.). IGF1R/IR needs N913 glycosylation to reach the surface (Kim et al.). Nrf2 S40 phosphorylation triggers its nuclear entry (Feng et al.). Splicing that removes such a residue should change where the protein goes, regardless of how much of it is made.
For isoforms without annotation, we additionally run a structure-based localization predictor trained on CAPSUL. CAPSUL showed that 3D geometry carries localization signal, but it contains only one canonical structure per gene. Applying a CAPSUL-trained predictor to predicted structures of non-canonical isoforms is an extrapolation, and we evaluate it as one.
The third is interpretation. Some localization changes act through transcription factors entering the nucleus, as with STAT3 in PMN-MDSCs, the immunosuppressive neutrophil-like myeloid cells studied in Fu et al.. For these, we check whether the cell type predicted to lose or gain a nuclear signal also shows a matching change in the target genes downstream of that factor. This gives an internal consistency check that does not require protein data.
Baselines
The naive reference point is doing nothing at the isoform level. That means gene-level CIBERSORT-style fractions plus a bulk differential-transcript-usage test between conditions, with no attribution to cell types. If bulk testing already finds the same switches and the cell type can be guessed from marker knowledge, our method adds little. This baseline is run on the same quantifications as our method.
The strongest simple competitor is fraction-interaction regression. It regresses each isoform's bulk proportion on estimated cell fractions and their interaction with condition, without weighting by cell-type gene expression, size factors, or quantification uncertainty. This is a fair test of whether our model's mixture structure matters. It is given the same fractions, quantifications, and regularization tuning budget, chosen by held-out likelihood.
For quantification, the strongest existing reference is MPAQT, at $r = 0.81$ against RT-qPCR for exon inclusion versus $r = 0.33$ for short reads only (Apostolides et al.). Cell-type-resolved isoform recovery, however, has no reported result on Lacuna, so there is no number to beat directly. For gene-level fractions, CSsingle is reported to outperform 10 methods, but the file gives no numbers. Likewise, no reported result was found on Lacuna for GNNDeconvolver on SpatialCTD.
For the localization layer, the baseline is lookup of annotated features alone. For the modification-site component, the reference is MeToken's reported F1 of 50.40% versus 43.94% for fine-tuned ESM-2 on its dataset. We reuse rather than try to beat that result.
Two ablations isolate the key ingredients: removing the size factor $r_c$, and replacing MPAQT uncertainty with point estimates. Beating the baselines means two things on the sorted-cell mixtures described below. Our per-cell-type $\pi$ correlation should be higher than fraction-interaction regression, with confidence intervals over mixture replicates that do not overlap. And the known switches should be recovered with the correct cell type and direction where the baselines fail.
Why not just…?
Why not just run CIBERSORT with transcripts instead of genes as rows? That changes the reference matrix but still outputs cell fractions, and fractions are not what we want. We expect transcript rows to change fraction accuracy little, and we will report that as an ablation. What gets lost is the per-cell-type, per-condition isoform usage. The FLT-1 switch is a change in $\pi$ within myeloid cells, not a change in how many myeloid cells there are.
Why not just regress bulk isoform ratios on cell fractions? This is a real threat, and it is the strongest baseline above. Our model differs in three ways. It weights each cell type's contribution by that cell type's gene-level expression, so a switch in a lowly expressing cell type is not swamped. It corrects for RNA content, which CSsingle showed biases mixtures. And it carries quantification uncertainty, which matters given the 0.33 versus 0.81 gap. If the sorted-cell mixture experiment shows no gain, we will report the simpler regression as sufficient and keep only the localization layer.
Why not just sort the cells or use long reads on every sample? Sorting is the ground truth Durrans et al. used, from 6 patients. That scale cannot cover the large archived short-read cohorts where clinical questions live. Long-read cohorts such as Yu et al., with 5 patients and 5 controls, remain small. Our long-read-budget curve states exactly how much long-read data is enough.
Why not just look up annotated domains instead of predicting modification sites and structures? For well-annotated proteins, lookup may be enough, so we report what the predictors add beyond it. The predictors matter mainly for the novel transcripts: the 13,593 found by Nanopore in the blood cohort are largely unannotated.
Proposed Experiments
The decisive test uses in-silico mixtures built from the sorted NSCLC profiles in Durrans et al.. These include CD14+ monocytic cells, neutrophils, and epithelial cells, each from both tumor and normal lung. We sample cell fractions from Dirichlet distributions spanning the reported fourfold rise in immune infiltration, mix reads at those fractions, and quantify. We then score three things: the correlation of estimated against true per-cell-type $\pi$, detection of tumor-versus-normal switches (AUROC against switches called in the sorted data), and whether the FLT-1 soluble-to-membrane switch is assigned to the correct cell type and direction. If our method cannot recover FLT-1 here, the project's core claim fails.
Quantification is checked on SIRV and MAQC, the synthetic spike-ins and reference mixtures with RT-qPCR ground truth, to confirm the MPAQT behavior carries over. A long-read-budget sweep subsamples long reads to find the point where $\pi$ accuracy plateaus.
On the Nanopore blood cohort of Yu et al., we have no isoform ground truth per cell type. Instead we run a biological sanity check. The model should attribute PTPRC (the gene encoding CD45) isoform structure to T cells. Its CD45 isoform shifts should also move with the reported depletion of CD8+ T cells and monocytes rather than being independent of it.
For the localization layer, we report what fraction of recovered switches alter a localization-relevant feature, and we check that the pipeline correctly flags the known cases in the file: the FLT-1 membrane anchor and the AR-V7 loss of the ligand-binding domain. CAPSUL-trained predictions on isoforms will be reported as hypotheses only, since no benchmark of isoform-pair localization exists. A small curated pair set is a deliverable, not an assumption.
References
- Identification of Reprogrammed Myeloid Cell Transcriptomes in NSCLC
- CSsingle: A Unified Tool for Robust Decomposition of Bulk and Spatial Transcriptomic Data Across Diverse Single-Cell References
- Accurate isoform quantification by joint short- and long-read RNA-sequencing
- MeToken: Uniform Micro-environment Token Boosts Post-Translational Modification Prediction
- Direct interaction with 14–3-3γ promotes surface expression of Best1 channel in astrocyte
- Heterodimerization of Glycosylated Insulin-Like Growth Factor-1 Receptors and Insulin Receptors in Cancer Cells Sensitive to Anti-IGF1R Antibody
- Tamoxifen activates Nrf2-dependent SQSTM1 transcription to promote endometrial hyperplasia
- CAPSUL: A Comprehensive Human Protein Benchmark for Subcellular Localization
- PKN2 enhances the immunosuppressive activity of polymorphonuclear myeloid-derived suppressor cells in esophageal carcinoma by mediating fatty acid oxidation
- Analysis of post-transcriptional regulatory signatures and immune cell subsets in premature ovarian insufficiency based on full-length transcriptome