When a tumor acquires two mutations, the joint effect is rarely the sum of the parts. Cancer genetics has documented this cooperation — "stepping on the gas while cutting the brakes" — for protein-coding genes like c-CBL and EGFR in lung cancer Tan et al. and for PTEN loss followed by miRNA-143/145 loss in breast cancer Wang et al.. Meanwhile, more than 90% of disease-associated variants sit outside protein-coding regions, in regulatory DNA that controls when and where genes are turned on. Yet the two literatures never meet: models of oncogenic synergy evaluate only coding mutations, and regulatory-variant models such as AlphaGenome — a deep network that reads a megabase of DNA and predicts chromatin, transcription, and splicing tracks — score every variant in isolation against a healthy reference genome. There is a third blind spot stacked on top: bulk tumor sequencing blends malignant subclones with stroma and immune cells, so an observed expression change could be a shifting cell mixture, a cell-state change, or a regulatory mutation carried only by one subclone — and current methods cannot tell these apart.
This proposal builds a single model that joins all three pieces: it infers which coding and non-coding mutations physically co-occur in the same subclone, predicts each subclone's regulatory output from its combined mutational haplotype using a sequence foundation model, and explicitly tests whether variant pairs interact — that is, whether the combined regulatory effect deviates from the sum of the individual effects — while a mixture component absorbs cell-composition confounding.
Core Claim
The central claim is that the functional impact of a non-coding somatic variant depends on which other mutations ride in the same clone, and that this dependence is measurable today by combining three existing capabilities that have never been composed. Epistasis — the deviation of a joint effect from additivity — is quantified at the regulatory-sequence level rather than the fitness or expression level. Concretely, let $S^A$, $S^B$, and $S^{AB}$ denote the patient's regulatory sequence carrying mutation A alone, mutation B alone, and both, and let $\hat{y}_g(S)$ be a sequence model's predicted activity for target gene $g$ on haplotype $S$. The regulatory epistasis between A and B is
$$\varepsilon_{A,B} \;=\; \hat{y}_g(S^{AB}) - \hat{y}_g(S^{A}) - \hat{y}_g(S^{B}) + \hat{y}_g(S^{\text{ref}}),$$
and the bulk expression of gene $g$ in a tumor is modeled as a clone-weighted mixture
$$y_g \;=\; \sum_{c} \pi_c \, \hat{y}_g(S_c) \;+\; \sum_{k} \rho_k \, m_{g,k},$$
where $\pi_c$ is the fraction of clone $c$ (carrying haplotype $S_c$), $\rho_k$ is the fraction of non-malignant cell type $k$ with reference-profile expression $m_{g,k}$, and the $\pi$ and $\rho$ are fit jointly from variant allele frequencies and expression data. The $\varepsilon$ terms are the scientific object: they say whether a silent-looking enhancer mutation becomes a driver only in the presence of a coding hit in the same clone — the sequence-level analogue of the pre-leukemic priming documented for single lesions Lee et al., Song.
How This Differs From the Closest Existing Work
Four systems come near this intersection, and each stops at a different edge. FiTree infers epistatic fitness effects from single-cell mutation trees, but its epistasis matrix covers only coding SNVs and is expressed as fitness parameters, not regulatory mechanism; we predict the molecular consequence of each clone's haplotype, including non-coding variants, and test additivity at the sequence-to-expression interface. SubNetDL is the strongest bulk-side competitor: it clusters mutations into subclones from bulk variant allele frequencies, propagates them over a protein-interaction network, and predicts drug response with graph attention networks — but it deliberately prunes to 253 curated protein-coding drivers, so non-coding variants are discarded before modeling begins, and it never asks whether two mutations in the same clone interact. AlphaGenome supplies our regulatory predictor but assumes variants act alone on a static reference background; it has never been asked a two-variant, clone-conditioned question. The purity-interaction eQTL model of Geeleher et al. separates tumor from microenvironment regulation, but treats the tumor compartment as one uniform genotype and scores each variant independently; our mixture model replaces the single tumor fraction with per-clone fractions, each carrying its own predicted regulatory program. scPGI-finder scPGI-finder detects genetic interactions from co-expression in single-cell RNA, which is downstream of DNA and confounded by cell state; we work at the DNA-haplotype level, where co-occurrence is physical rather than correlational.
Research Program
The build has four stages. First, subclone inference: from bulk whole-genome sequencing, cluster somatic variant allele frequencies into clones under purity and ploidy constraints, using the consistency equation of CNAqc as a validity check on the inferred state — this matters because CNAqc showed that 901 of 1,464 TCGA samples failed naive purity checks and that 88% could be rescued by a consistency-guided choice, so purity errors would otherwise poison every downstream clone assignment. Crucially, the clustering includes non-coding variants (promoter, enhancer, and silent regions), which requires whole-genome rather than targeted-panel data. Second, clone-conditioned regulatory prediction: for each clone, apply AlphaGenome to the reconstructed haplotype carrying that clone's combined coding and non-coding mutations, restricted to tracks active in the relevant cell type, chosen using single-cell accessibility references of the kind ChromFound is trained on — its genome-coordinate tokenization and in-silico enhancer knockout capability provide the cell-type conditioning that AlphaGenome's bulk reference tracks lack, while AlphaGenome's megabase context covers the distal enhancer contacts that ChromFound's ±200 kb attention window misses; we use the two as complementary oracles rather than choosing between them. Third, epistasis estimation: compute $\varepsilon_{A,B}$ for all within-clone coding–non-coding and non-coding–non-coding pairs via in-silico double mutagenesis, with uncertainty propagated from the sequence model's own variance across tracks. Fourth, joint mixture fitting: fit clone fractions $\pi_c$ and cell-type fractions $\rho_k$ to bulk RNA-seq, so that the residual expression not explained by composition is attributed to within-clone regulatory rewiring rather than to mixing — directly addressing the three-way ambiguity (mixture shift, cell-state shift, subclone-specific mutation) that no current deconvolution method resolves.
Proposed Experiments
Three evaluation tiers, all on data named in the corpus. First, ground-truth co-occurrence: the Morita acute myeloid leukemia single-cell mutation trees (123 trees, 31 patients, used by FiTree) give physical variant co-occurrence per cell, so we test whether our sequence-level $\varepsilon$ estimates for known interacting pairs such as DNMT3A–FLT3 agree in sign with FiTree's fitness-epistasis estimates; no reported result found on Lacuna for FiTree's numerical epistasis accuracy, so agreement is assessed by sign concordance and credible-interval overlap rather than a published scalar. Second, expression prediction on TCGA and PCAWG whole-genome cohorts: predict bulk expression of variant target genes (assigned by AlphaGenome's distal-contact tracks rather than nearest-gene, which is known to misassign Xu et al.), and compare the full epistatic mixture model against the additive-only ablation, reporting correlation of predicted with held-out observed expression, stratified by tumor purity — we expect degradation below roughly 20% purity, consistent with CNAqc's reported low-purity failure horizon. Third, cell-type conditioning and clinical transfer: condition track selection on the Salcher lung adenocarcinoma atlas of about 161,000 cells (used by scPGI-finder), and test whether adding our epistatic non-coding features to a SubNetDL-style response predictor improves on its reported median AUROC of 0.77 on the MSK_2020 and MSK_2018 anti-PD-1 cohorts.
Baselines
The naive reference point is nearest-gene annotation of non-coding variants plus aggregate mutation burden (tumor mutational burden). This is what most analyses effectively do, and it is already known to be weak: SubNetDL outperformed tumor mutational burden on both MSK immunotherapy cohorts ($P = 0.02$ on MSK_2020; $P = 6.0\times10^{-3}$ on MSK_2018) with a median AUROC of 0.77 SubNetDL, and proximity-based target assignment misassigns distal enhancer variants that contact promoters megabases away Xu et al.. We will run it on the identical TCGA and MSK splits with the same tuning budget; beating it means higher expression-prediction correlation and higher response AUROC than its numbers, which should be easy and is not the claim of interest.
The strongest competitor is additive single-variant scoring with AlphaGenome: predict each variant's effect in isolation against the reference genome and sum. This is the direct ablation of our key ingredient — it is our own pipeline with $\varepsilon$ forced to zero and clone structure collapsed to a single averaged tumor haplotype. It is strong because AlphaGenome reports eQTL sign prediction improved from 75% to 80% auROC on its evaluation suite AlphaGenome, so additivity already carries most of the signal we hope to exploit. To make the comparison fair, both variants of the pipeline use the same AlphaGenome checkpoint, the same tracks, the same data splits, and the same compute; the epistatic model gets no extra parameters at the sequence level, only the mixture decomposition and $\varepsilon$ estimation on top. We beat it only if clone-conditioned epistatic scoring raises eQTL-sign auROC above the 80% single-variant figure on matched tasks and explains held-out clone-specific expression residuals that the additive model leaves unexplained. A second substantive baseline is the purity-interaction eQTL regression of Geeleher et al. run on the same bulk data with the same purity estimates: it is the best existing method for separating composition confounding from tumor-cell regulation, and our model must attribute less of the expression signal to microenvironment than it does while fitting subclonal structure it cannot see.
Two ablations isolate the proposal's ingredients. The clone-shuffle ablation randomly reassigns variant co-occurrence across clones before computing $\varepsilon$; SubNetDL showed that shuffling subclonal assignments significantly degraded response prediction, and we expect the same collapse here if physical co-occurrence carries real signal. The composition-collapse ablation fits the model without the $\rho_k$ cell-type term, which should cause spurious $\varepsilon$ estimates in low-purity tumors — the falsifiable prediction that mixture confounding masquerades as regulatory epistasis unless explicitly modeled.
Why not just…?
Why not just score each variant independently with AlphaGenome and add the effects? This is the most serious alternative, and the outcome is genuinely uncertain. The additive model already achieves 80% auROC on eQTL sign prediction AlphaGenome, and epistatic terms could turn out to be small relative to main effects. The decisive experiment is the within-pipeline comparison described above: if, on the Morita trees and TCGA held-out expression, the additive model matches the epistatic one, the core claim is falsified and we will report that — it would itself be a useful negative result, since the entire oncogenic-synergy literature assumes interaction is the rule. What the additive model cannot do even in principle is distinguish a subclone-specific regulatory mutation from a composition shift, because it has no clone variable at all.
Why not just deconvolve the bulk sample into cell types first, then run standard eQTL per cell type? Reference-based deconvolution with tools like CIBERSORTx and MuSiC is the standard answer to composition confounding, but it addresses only the $\rho_k$ part of the problem: it recovers cell-type fractions, not subclonal genotypes, so any expression change confined to one malignant clone is attributed to whichever cell type it most resembles. The corpus also shows these linear tools are already being outperformed — Count Bridges beat CIBERSORTx and MuSiC on bulk deconvolution and STDeconvolve on spatial spots Count Bridges, and POPPY beat standard bulk encoders by up to 73% on silhouette-score integration POPPY. We will nevertheless run a CIBERSORTx-plus-per-cell-type-eQTL pipeline as a baseline at the decomposition stage; if it captures everything our joint mixture fit does, the clone component adds nothing and the proposal should shrink to a composition-correction method.
Why not just train a classifier on the flat binary mutation matrix? Random forests, support vector machines, and logistic regression on raw mutation calls need no subclone inference, no sequence model, and no mixture decomposition. SubNetDL already compared against these and against itself with subclonal features removed, with the structured model winning at a median AUROC of 0.74 across 507 TCGA patients spanning 10 cancer–drug combinations SubNetDL. Flat matrices also cannot contain our object of study at all: with non-coding variants one-hot encoded, there is no notion of within-clone haplotype or of interaction at the sequence level, and tree-model epistasis like FiTree's fitness parameters is lost entirely.
Decisive Checks
Three results would decide the proposal. The first is the sign-concordance test on the Morita AML trees: our sequence-level $\varepsilon$ estimates should recover the direction of FiTree's fitness epistasis for coding pairs, and, for the pairs where whole-genome data reveal non-coding co-mutations, predict effects the fitness-only model cannot represent. The second is the additive-versus-epistatic comparison on held-out TCGA and PCAWG expression: a meaningful gap in correlation, concentrated in low-purity and multi-subclone tumors, confirms that clone-conditioning and interaction terms carry signal beyond isolated-variant scoring; no gap falsifies the core claim. The third is the purity-stratified breakdown: if spurious $\varepsilon$ estimates cluster in the tumors where the composition-collapse ablation was run — that is, if removing the cell-type term manufactures epistasis — then the confounding decomposition is doing its job and any analysis that skips it is measuring mixture, not mechanism. Together these checks either establish non-coding regulatory epistasis at subclone resolution as a measurable phenomenon or cleanly bound how little it adds to single-variant interpretation — both outcomes worth publishing.
References
- CBL Is Frequently Altered in Lung Cancers: Its Relationship to Mutations in MET and EGFR Tyrosine Kinases
- microRNA-143/145 loss induces Ras signaling to promote aggressive Pten-deficient basal-like breast cancer
- AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model
- STAT5B leukemic mutations, altering SH2 tyrosine 665, have opposing impacts on immune gene programs
- Early lesions of follicular lymphoma: a genetic perspective
- Bayesian inference of fitness landscapes via tree-structured branching processes
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer
- Charting cell-type-specific positive genetic interaction at single-cell resolution for lung adenocarcinoma
- Computational validation of clonal and subclonal copy number alterations from bulk tumor sequencing using CNAqc
- ChromFound: Towards A Universal Foundation Model for Single-Cell Chromatin Accessibiltiy Data
- HiView: an integrative genome browser to leverage Hi-C results for the interpretation of GWAS variants
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level