Some cancers switch on a growth gene without mutating it. A chromosome break and rejoin moves a powerful control element (an "enhancer," a stretch of DNA that boosts transcription of nearby genes) next to that gene, and the gene is driven hard. This is called enhancer hijacking. Nearly all hijacking discoveries come from bulk tumor samples, where sequencing averages over every cell in the piece of tissue. If only a fifth of the tumor cells carry the rearrangement, and the sample is half stroma, then only about one cell in ten has the boosted gene. The expression jump, and the downstream program it switches on, is diluted roughly tenfold. Current detectors treat each tumor as either carrying the rearrangement or not, so a diluted subclonal event looks like noise.
Whole-genome sequencing already measures that dilution. The share of reads spanning a breakpoint tells us what fraction of cancer cells carry the rearrangement, and tumor purity tells us what fraction of the sample is cancer. We propose to use the product of these two, the fraction of all sampled cells that carry the event, as a continuous dose. Both the hijacked gene's expression and its downstream pathways should scale with this dose, and nothing else in the sample should. This gives a stronger association test than a carrier-versus-noncarrier comparison. It also gives a cleaner input for functional enrichment analysis (testing whether a gene list is overrepresented in curated pathways), because genes are ranked by how well they track the dose rather than by a cut-off on bulk differential expression.
Core Claim
The claim is that weighting by carrier fraction recovers subclonal hijacking events and their downstream pathways from ordinary bulk WGS plus RNA-seq, and it does so better than two alternatives: binary-carrier regression, and atlas-based expression deconvolution followed by enrichment. It needs no single-cell data and no reference atlas. For gene $g$ in tumor $i$, we model log expression as
$$ y_{ig} = \alpha_g + \gamma_g\,\pi_i + \delta_g\,\pi_i\,(c_{ig}-2)\,\phi^{\mathrm{CN}}_{ig} + \beta_g\,\pi_i\,\phi^{\mathrm{SV}}_{i} + \mathbf{z}_i^\top\boldsymbol{\eta}_g + \varepsilon_{ig}. $$
The symbols are as follows. $\pi_i$ is tumor purity. $\phi^{\mathrm{SV}}_i$ is the cancer cell fraction of the candidate rearrangement, which is zero in non-carriers. $c_{ig}$ is the local copy number and $\phi^{\mathrm{CN}}_{ig}$ is the cancer cell fraction of that copy-number change. $\mathbf{z}_i$ holds cancer type and estimated immune and stromal fractions. $\beta_g$ is the hijacking effect.
The form follows from where the effect lives. Hijacking acts only in carrier cells, so its contribution to a sample average is proportional to the share of cells that carry it. Copy-number dosage is handled by its own purity- and fraction-scaled term. This separation addresses the open problem of telling hijacking apart from amplification in mixed samples. The additive purity term absorbs the generic "more tumor, more tumor genes" signal, so $\beta_g$ is identified only by carrier status and how subclonal the carriers are. The same regression run over all genes, with $\phi^{\mathrm{SV}}$ fixed to one recurrent event, gives a genome-wide ranking by $t$-statistic for $\beta$. We feed that ranking into rank-based pathway enrichment.
Relation to Prior Work
Four bulk studies motivate the gap. High-coverage whole-genome analysis of 1220 cancers regressed expression on binary SV presence within 100 kb, corrected for copy number and cancer type, and found 1,575 deregulated genes, 95% of them upregulated. Its authors state they are underpowered below roughly 1–3% event frequency. Medulloblastoma CESAM analysis found PRDM6 hijacking. The ccRCC study (Structural variation drives enhancer hijacking via 3D genome disruption in ccRCC) found and validated SEMA5B hijacking by silencing the enhancer with CRISPR interference (CRISPRi). The lymphoma study (super-enhancer retargeting) traced a PAX5-to-ZCCHC7 switch. All four treat each tumor as clonal. We keep the first study's regression structure and change only how the rearrangement enters it.
Two lines of work supply ideas but leave the problem open. SubNetDL (network-based subclonal model) shows that subclonal architecture inferred from bulk allele frequencies carries predictive signal that aggregate counts miss. It uses only point mutations and a protein-interaction graph, with no rearrangements and no expression. Count Bridges (Count Bridges) and POPPY (cross-modal transfer) deconvolve bulk expression into cell-level profiles. Both learn from single-cell references, and the POPPY paper acknowledges that tumor states absent from the atlas cannot be resolved. A hijacked subclone is exactly such a state. For that reason we treat them as competitors, not components.
The subclonal reconstruction benchmark (Crowd-sourced benchmarking) constrains the design. It found accuracy degrades below about 32 reads per chromosome copy, and that algorithm choice explains more variance than tumor features. We will therefore propagate uncertainty in cancer cell fraction rather than plugging in point estimates, and report results stratified by that depth threshold. For pathway libraries we reuse the curated collection behind SEanalysis, which tests static, unrearranged super-enhancer catalogs and so cannot see relocated enhancers.
Research Program
The pipeline has four steps.
- Estimate the dose. We estimate SV cancer cell fraction from split and discordant read support relative to local copy number and purity. We fit it jointly with SNV-based clusters so that each SV is assigned to a subclone, with a posterior distribution rather than a single value.
- Fit the dose-scaled regression. We fit the regression above with the fraction treated as uncertain. We use multiple imputation from the posterior, or an errors-in-variables fit if that proves tractable.
- Rank genes and test pathways. For recurrent hijacked targets, we rank all genes by the dose-specific $\beta$ statistic and run preranked enrichment against KEGG, Reactome and the SEanalysis pathway collection.
- Correct for hub bias. Hubs such as TP53 sit in many pathways, so the enrichment null is built by permuting tumor labels within cancer type. This preserves gene–gene correlation and annotation structure, and avoids comparing against a naive hypergeometric background.
The deliverable is a calibrated test for subclonal hijacking plus pathway maps for targets such as TERT, MDM2, CD274 and SEMA5B at subclonal doses.
Baselines
Naive reference. We split tumors into carriers and non-carriers of an event, call differentially expressed genes with a negative-binomial model, and run overrepresentation analysis at a fixed fold-change and FDR cut-off. This is what most enrichment studies do, and any method that cannot beat it is not worth adopting.
Strongest existing hijacking detector. This is the binary-SV linear model of the 1,220-cancer study, reimplemented with its 100 kb window and its copy-number and cancer-type covariates. It is the strongest competitor because it is the only cohort-scale SV–expression model in the file and already handles copy number. That paper reports 1,575 significant genes. We found no reported result on Lacuna for its recall of known hijacking events as a function of subclonal fraction, so we will measure that ourselves. It will run on the identical PCAWG samples, SV calls and covariates, with the same FDR control. Our method beats it if it reaches higher recall at matched empirical FDR for events below 50% carrier fraction, while staying no worse on clonal events.
Strongest deconvolution competitor. Count Bridges is used to estimate the malignant-compartment profile, followed by enrichment on carrier-versus-noncarrier malignant profiles. It is trained on the same single-cell references it was evaluated with, and we give it a tuning budget equal to ours in GPU-hours. The file states it outperforms CIBERSORTx and MuSiC on cell-type proportion recovery. We found no reported numeric deconvolution result on Lacuna for Count Bridges.
Ablations. Two ablations isolate the key ingredient. The first fixes $\phi^{\mathrm{SV}}=1$ for all carriers, so dose depends on purity only. The second drops the copy-number fraction term.
Why not just…?
Why not just add purity as a covariate to the existing binary model? This is the most serious threat. An additive purity term removes the tumor-content trend but does not scale the hijacking effect by how many cells carry it. A 10%-carrier tumor and a 90%-carrier tumor are still weighted equally. Our first ablation (purity-scaled indicator, no cancer cell fraction) sits between this alternative and the full model. If it matches the full model, the honest conclusion is that purity interaction is enough and estimating SV cancer cell fraction is unnecessary. That would be a useful, cheaper result, and we will report it as such.
Why not just deconvolve with Count Bridges or POPPY and run ordinary enrichment? These methods reconstruct cell-level expression from references of known cell states. A hijacked subclone that massively overexpresses NR4A3 or SEMA5B is not in any atlas, which POPPY's own limitation section acknowledges. They also do not use the genomic evidence of who carries the event. The head-to-head on the titration data below settles this directly: if the deconvolution route recovers the downstream program at low carrier fractions as well as we do, our method adds little.
Why not just profile tumors with single-cell chromatin and RNA assays? For new prospective studies that is the right tool when affordable. It cannot be applied to archives like PCAWG, and the POPPY paper notes single-cell assays remain too costly for routine clinical use. Our method is meant for the bulk archives where nearly all hijacking discoveries already sit.
Proposed Experiments
Wet-lab titration (decisive). Using 786-O, the ccRCC line in which Dong et al. silenced the SEMA5B-hijacking enhancer by CRISPRi, we mix control and enhancer-silenced cells at carrier fractions of 100, 50, 25, 10 and 5%. We further dilute each mix with non-malignant kidney epithelial cells to vary purity, then run bulk RNA-seq in triplicate. Here the cancer cell fraction is set by design, which avoids the estimation noise flagged in Crowd-sourced benchmarking. The pure contrast at 100% defines the reference downstream program. At each lower fraction we measure how much of that program each method recovers, using the overlap of enriched pathways at matched FDR and the rank correlation of gene statistics. The prediction is that our dose-weighted ranking keeps the program detectable at roughly 10% carriers, where carrier-versus-noncarrier ORA and deconvolution-then-ORA lose it. If that does not happen, the central claim fails.
In silico mixing on PCAWG. For tumors carrying high-confidence clonal events at the named targets (TERT, MDM2, CDK4, ERBB2, CD274, PDCD1LG2, IGF2), we mix their RNA-seq counts with counts from same-type non-carriers, and correspondingly downsample breakpoint reads. This gives known subclonal fractions. We then measure recall at fixed empirical FDR for every baseline, stratified by reads per chromosome copy above and below 32.
Discovery run. On the full PCAWG cohort we report new candidates whose support comes mainly from subclonal carriers. These are events the binary model scores as non-significant, and their pathway maps are produced with the within-type permutation null. Each candidate is checked for consistency with breakpoint geometry, meaning the enhancer is relocated to the gene side of the break, before we interpret it.
References
- High-coverage whole-genome analysis of 1220 cancers reveals hundreds of genes deregulated by rearrangement-mediated cis-regulatory alterations
- The whole-genome landscape of medulloblastoma subtypes
- Structural variation drives enhancer hijacking via 3D genome disruption in ccRCC
- Noncoding mutations cause super-enhancer retargeting resulting in protein synthesis dysregulation during B cell lymphoma progression
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Crowd-sourced benchmarking of single-sample tumor subclonal reconstruction
- SEanalysis: a web tool for super-enhancer associated regulatory analysis