Cell-Cycle-Resolved Allele-Specific Binding from Bulk ChIP-seq

Most genomic assays grind up a million cells and report one blended average. Two sources of blending are well known individually but almost never handled together. The first is cell type: a tumor biopsy mixes cancer, stromal, and immune cells, and bulk measurements hide rare subpopulations. The second is cell cycle: even a "pure" cell line contains cells scattered across division phases, and drugs that freeze cells at a checkpoint — the whole point of CDK4/6 inhibitors and similar therapies — shift which proteins sit on DNA. Meanwhile, the deep-learning models that predict how a DNA letter change outside a protein-coding gene affects regulation are trained on these averaged pictures. They can tell you a variant disrupts binding "on average," but not whether it disrupts binding only in cells arrested at a division checkpoint. The clinical case where this matters is precise: therapies that work by arresting cells, given to heterogeneous tumors, where the variant's effect may exist only in the arrested subpopulation that the drug creates.

Chromatin immunoprecipitation sequencing — ChIP-seq, which maps where a given protein physically binds DNA — can in principle see this, but two obstacles stand in the way. Bulk ChIP-seq needs millions of cells, so it averages over all cycle states at once; single-cell ChIP-seq recovers only hundreds to a few thousand DNA fragments per cell, leaving individual regulatory loci nearly empty. This proposal argues that both obstacles can be cleared computationally, and that doing so changes what non-coding variants we can interpret. The specific target is allele-specific binding: when a person's two chromosome copies differ at one DNA letter, and the binding protein prefers one copy over the other. That preference is the cleanest direct readout of a non-coding variant's effect, and today it is measured only in pooled, unsynchronized cultures.

Core Claim

The claim has a measurement half and a model half. First, a substantial fraction of allele-specific binding is cell-cycle-state-specific: it exists in G0/G1-arrested or G2/M-arrested cells and is diluted below statistical detection when reads from all states are pooled — a concrete instance of the masking that bulk averaging imposes on stochastic and subpopulation-level biology (Genome 2011). Second, a generative count model that treats bulk ChIP-seq reads as a mixture over cycle states, with binding strengths conditioned on DNA sequence, can recover these hidden effects from unsynchronized data alone. The model, which we call CycleBridge, is built by porting a discrete-count deconvolution machinery from transcriptomics to ChIP-seq and grafting on sequence-aware, state-specific binding affinities.

The generative story is one equation. Let $y_{s,a}$ be the ChIP-seq read count at regulatory locus $s$ on allele $a$ (reference or alternate), and let $k$ index cycle states — cycling, G0/G1-arrested, G2/M-arrested:

$$y_{s,a} \sim \mathrm{Poisson}\!\left(\sum_{k} \pi_k \, \exp\big(w_k^{\top} \phi_{s,a}\big)\right)$$

Here $\pi_k$ is the fraction of cells in state $k$, $\phi_{s,a}$ is a vector of sequence features for the allele (transcription-factor binding affinities of the kind GRAM computes from 515 DeepBind models), and $w_k$ is a state-specific activity weight — the ingredient that says a factor is active in G0/G1 arrest but silent in G2/M. Inference is an Expectation-Maximization procedure in which per-state counts are latent variables, using the integer-valued stochastic bridge dynamics that Count Bridges introduced for RNA counts; allele-specific binding per state is then tested with the beta-binomial model that AlleleDB established, applied within rather than across states.

What Exists and What Is Missing

Three lines of work come close and each stops one step short. MixChIP (MixChIP: a probabilistic method for cell type specific protein-DNA binding analysis) deconvolves bulk ChIP-seq into cell types with a Bayesian Poisson model, but treats the components as a few static, discrete lineages — validated almost entirely on two-cell mixtures — and makes no variant calls. AlleleDB (A uniform survey of allele-specific binding and expression over 1000-Genomes-Project individuals) calls allele-specific binding at scale by mapping reads to personalized diploid genomes, but on pooled unsynchronized cultures, so a binding difference active only in one cycle phase is diluted away; its companion study (Variation in Transcription Factor Binding Among Humans) found two-thirds of between-individual binding variation unexplained by local sequence variants, which unmodeled cycle heterogeneity could plausibly account for. The sequence predictors — Enformer (Effective gene expression prediction from sequence by integrating long-range interactions), AlphaGenome (AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model), and GRAM (GRAM: A GeneRAlized Model to predict the molecular effect of a non-coding variant in a cell-type specific manner) — predict variant effects from DNA, but their training tracks are bulk and static: GRAM's cell specificity comes solely from average bulk expression of transcription factors, with no notion that a factor's activity changes when a cell is arrested.

On the method side, two papers supply parts we reuse. Count Bridges (Count Bridges enable Modeling and Deconvolving Transcriptomic Data) shows that modeling exact integer counts with Poisson birth-death dynamics beats continuous deconvolution on RNA data — but never touches ChIP-seq or variants. SIMPA (Interpretable machine learning models for single-cell ChIP-seq imputation) imputes sparse single-cell ChIP-seq from bulk ENCODE references — but its reliance on bulk references guarantees that an arrested state absent from ENCODE cannot be imputed, which is exactly the state we care about. No work on record connects arrest states, ChIP-seq deconvolution, and variant interpretation; that triangle is the proposal.

Research Program

The program has three parts. Part one builds the reference data that does not yet exist: an arrest-state ChIP-seq atlas in diploid cell lines. We will treat 1000 Genomes lymphoblastoid cell lines — chosen so we inherit the genotypes, the p65/NF-κB and Polymerase II factors, and the diploid-mapping pipeline from the existing allele-specific binding studies — under three conditions: untreated cycling, G0/G1 arrest via a CDK4/6 inhibitor (palbociclib, whose clinical activity and arrest mechanism are established; A Decade After Approval of the First CDK4/6 Inhibitor), and G2/M arrest via CDC25B degradation (Inhibition of CDC25B With WG-391D Impedes the Tumorigenesis of Ovarian Cancer) or PLK1 destabilization through the USP7 axis (USP7 is a novel Deubiquitinase sustaining PLK1 protein stability and regulating chromosome alignment in mitosis). ChIP-seq for p65, Polymerase II, and one histone mark, with matched input controls following MixChIP's practice. Reads are mapped to personalized diploid genomes per AlleleDB, and allele-specific binding is called within each condition. A subset of cultures will also be sorted by DNA content (Hoechst FACS) into phase fractions and profiled separately, giving orthogonal ground truth. The headline deliverable of part one is the first catalog of arrest-state-specific allele-specific binding, and a direct measurement of how many events pooled bulk analysis misses.

Part two builds CycleBridge. The sequence component reuses GRAM's design of universal DeepBind affinities, but replaces GRAM's static weighting by average transcription-factor expression with learned state activity weights $w_k$ fit to the part-one atlas. The deconvolution component ports the integer count-bridge machinery from Count Bridges to ChIP-seq read counts, with ENCODE bulk profiles serving as the cycling-state reference and the atlas supplying the arrested-state references — closing the reference gap that SIMPA's ENCODE-only design cannot. The model outputs, for any bulk sample, state proportions $\pi_k$ and per-state, per-allele binding intensities, from which state-specific allele-specific binding calls and variant effect scores follow.

Part three evaluates against the benchmarks above and tests the model where its competitors structurally fail: reanalyzing the primary breast cancer estrogen-receptor ChIP-seq biopsies from the MixChIP study, asking whether part of what MixChIP attributes to cell-type proportions is actually cycle-state composition; and applying CycleBridge to the Grosselin single-cell ChIP-seq data to impute arrested-state profiles that a reference set built from ENCODE alone cannot represent.

Baselines

The do-nothing reference point is pooled allele-specific binding exactly as practiced today: the beta-binomial test on unsynchronized bulk reads with personalized diploid mapping and bias filtering, i.e. the AlleleDB procedure (A uniform survey of allele-specific binding and expression over 1000-Genomes-Project individuals), together with nearest-gene assignment for any variant-to-gene interpretation. Its reference number is the 6,136 allele-specific binding SNVs that procedure recovered across 382 bulk lymphoblastoid lines. Beating it means recalling, against the sorted-fraction ground truth of part one, state-specific binding events that the pooled test at the same false-discovery rate does not detect.

The strongest existing competitor is MixChIP (MixChIP: a probabilistic method for cell type specific protein-DNA binding analysis), a Bayesian Poisson deconvolution with matched input controls that jointly estimates cell-type proportions and type-specific binding. On simulated ENCODE mixtures of two cell lines at 10:90 to 90:10 ratios, it reports AUC values of 0.95 for HepG2 and 0.98 for K562 using only three mixed samples. We will run MixChIP on our simulated arrest-state mixtures (constructed by mixing our atlas profiles at the same ratios and site counts, with matched input controls) under the same compute and tuning budget as CycleBridge, and also feed it the sorted-fraction profiles as candidate "types." Two results would count as beating it: higher AUC when the mixture components are overlapping cycle states rather than distinct cell lines, and calibrated per-state allele-specific binding calls, which MixChIP does not produce at all.

The strongest static variant scorers are the sequence models: Enformer, which achieved an AUC of 0.764 on gene-centric evaluation of fine-mapped GTEx causal variants (Gene-centric evaluation of causal variant prediction for DNA models), and AlphaGenome, which reports eQTL sign-prediction AUROC of 80% (AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model). These will be scored on held-out state-specific allele-specific binding events from part one; beating them means higher AUROC on which variants show arrest-specific effects, a task their bulk training tracks never represented. Finally, one ablation isolates the proposal's key ingredient: the same model with the count-bridge replaced by continuous Gaussian deconvolution in the CIBERSORTx style, and a second ablation dropping the sequence conditioning (state weights only, no variant awareness). If the continuous variant matches the count model on ChIP-seq data, the count machinery is not earning its complexity and we will say so.

Why not just…?

Why not just synchronize the cells experimentally? For cell lines, this genuinely works, and we use it ourselves in part one to create ground truth — chemical blocks and FACS sorting give us the reference atlas. The limitation is scope, not validity: primary tumors such as the breast cancer biopsies in the MixChIP study are fixed or frozen on excision and cannot be cultured or synchronized; chemical blocks trigger non-physiological stress responses; and Hoechst sorting loses enough cells to starve a technique that needs $10^6$–$10^7$ inputs. The computational model targets exactly the samples synchronization cannot reach, and the decisive check is agreement between model-inferred and FACS-sorted profiles on samples where both exist — if they agree, the model earns its claim on the samples where only it can run.

Why not just run standard deconvolution like CIBERSORTx or MuSiC? Three pieces of evidence say this falls short. Count Bridges (Count Bridges enable Modeling and Deconvolving Transcriptomic Data) already outperformed both on integer count data, though the file records no numerical value for that margin; standard methods assume continuous readouts and miss the discrete Poisson structure of sequencing counts. POPPY (Cross-modal transfer learning for mapping bulk transcriptomes at cellular level) showed that rigid predefined signatures degrade across cohorts and conditions, improving patient-composition correlation by up to 714% over such baselines. Most fundamentally, none of these methods produce allele-specific output — they can estimate proportions but cannot say anything about a variant's effect, which is the point of the proposal. The continuous-deconvolution ablation settles empirically whether the count machinery specifically matters for ChIP-seq.

Why not just scan position-weight-matrix motifs and filter by conservation? This is the cheapest alternative and a real threat, so we run it as a baseline rather than argue around it. The evidence against it as a complete answer: GRAM (GRAM: A GeneRAlized Model...) found that conservation scores "contributed almost nothing" to predicting immediate molecular effects, while binding affinities did; AlleleDB found that 90% of motif-matching allele-specific binding followed the expected motif-disruption trend, which means 10% did not — and motif matches occur millions of times genome-wide without in vivo binding. Kasowski (Variation in Transcription Factor Binding Among Humans) documented cooperative binding (a Stat1 mutation disrupting adjacent NF-κB binding) that local motif scans cannot represent. Our experiment adjudicates directly: if PWM delta scores plus a conservation filter match CycleBridge's AUROC on state-specific allele-specific binding, the state-conditioning machinery is not needed and the simple path wins.

Decisive Experiments and Checks

Four checks carry the weight. First, the masking measurement: from part one's matched cycling-versus-arrested profiles in genotyped lines, what fraction of state-specific allele-specific binding events falls below detection when reads are pooled — and how does that fraction compare to the 6,136 events the bulk survey could see? A dilution effect concentrated in binding sites of cycle-regulated factors would confirm the core claim at its source. Second, the deconvolution test: AUC of CycleBridge versus MixChIP's reported 0.95/0.98 on two-component mixtures, re-run by us on arrest-state mixtures where its discrete-type assumption is violated. Third, the prediction test: AUROC of CycleBridge versus Enformer (0.764 gene-centric), AlphaGenome (80% eQTL sign AUROC), GRAM (0.72 on the GM12878 MPRA benchmark), PWM delta scores, and the 1.5 kb CNN baseline (0.695 variant-centric, from the gene-centric evaluation study) on held-out state-specific allele-specific binding. Fourth, the reanalysis: on the Kasowski-style data and the MixChIP breast cancer biopsies, does allowing cycle-state components explain part of the two-thirds of binding variation that local sequence variants leave unexplained? A positive answer on checks one and two, with at least a margin over the static sequence models on check three, would establish that cell-cycle arrest is not a nuisance confounder of non-coding variant interpretation but a determinant of it — and that bulk ChIP-seq, read through the right count model, can see it.