Inside a cell nucleus, proteins do their work in physical machines: they dock onto each other, assemble, and disassemble on timescales of minutes to hours. A growing family of algorithms predicts when these machines form by looking at messenger RNA measurements over time — if two genes' RNA levels rise and fall together, the algorithms declare their proteins interaction partners at that moment. The uncomfortable truth is that nobody has ever checked these predicted "dynamic interaction networks" against physical measurements of protein binding taken at many time points in a mammalian system. We propose to do exactly that, in a setting where the timing carries real biological meaning: the drug JQ1 pulling BRD4 — a reader protein that uses a pocket called a bromodomain to bind acetyl chemical tags on histones, the spools around which DNA is wound — off the chromatin of human neurons.
The proposal has two deliverables that reinforce each other. First, a benchmark: the first mammalian dataset in which a physically measured, time-resolved protein interaction network (captured by co-immunoprecipitation — "molecular fishing," where an antibody to a chosen bait protein reels out whatever is physically stuck to it — followed by mass spectrometry) is sampled at the same dense time points as matched RNA sequencing, during an acute epigenetic perturbation. Second, a method built on it: a longitudinal gene-set enrichment procedure whose gene-set membership is structured by the measured, time-varying physical complexes rather than by static database pathways or RNA co-expression guesses.
Core Claim
Our claim is falsifiable in both halves. First, dynamic protein interactions inferred computationally from RNA co-expression — the recipe used to build the yeast DPPIN benchmark, where an edge exists at time $t$ only if both genes exceed a three-sigma activity threshold and their sliding-window Pearson correlation is high — will diverge substantially from physically measured interactions during BRD4 displacement. Second, when the physically measured, time-varying complex memberships are used to structure longitudinal enrichment analysis, the analysis detects transient functional programs and orders them correctly relative to complex disassembly, where both unstructured longitudinal enrichment and static-network enrichment fail or blur.
The motivation comes from three gaps that no single paper has closed. DPPIN: A Biological Dataset of Dynamic Protein-Protein Interaction Networks built twelve dynamic yeast networks from 36 timestamps of expression data but validated zero of its time-dependent edges biochemically, and evaluated graph-mining metrics rather than biological function. Chromatin profiling in human neurons reveals aberrant roles for histone acetylation and BET family proteins in schizophrenia established the biology we perturb — BRD4 reads a hyperacetylated form of the histone variant H2A.Z in schizophrenia neurons, and the BET inhibitor JQ1 partially rescues transcriptional deficits — but treated JQ1 as a static endpoint, with no co-immunoprecipitation and no time course. And Gene set proximity analysis: expanding gene set enrichment analysis through learned geometric embeddings showed that embedding protein interaction networks improves enrichment sensitivity, but its networks are frozen in time and drawn from databases, not measurements. We combine the perturbation system of the second, the dynamic-network question of the first, and the enrichment machinery of the third, adding the one ingredient none of them has: physical, time-resolved interaction data.
Research Program
The experimental system is human induced pluripotent stem cell-derived neurons, both schizophrenia and control lines, following Farrelly et al., who showed the hyperacetylated H2A.Z state and BRD4 reading of it appear in mature schizophrenia neurons but not in earlier progenitor stages — so the mature neurons are where the reader's behavior is most at stake. We treat cultures with JQ1 or vehicle and sample at seven time points spanning the acute-to-adaptation window (0, 1, 2, 4, 8, 12, and 24 hours; the 12-versus-24-hour contrast is chosen because Enrichment Map showed pathway programs emerging between those two points that a single endpoint would miss entirely). At each time point we collect three matched layers: RNA sequencing with biological replicates; co-immunoprecipitation of BRD4 using antibody crosslinked to magnetic beads, following the biochemical protocol of Sun et al., with all time points multiplexed into single mass spectrometry runs so that time becomes a within-run label rather than a batch effect; and, at two selected time points, proximity ligation assays in intact cells, following the orthogonal-validation practice of Foster et al., to guard against the post-lysis re-association artifacts that co-immunoprecipitation is known to suffer. Because JQ1 acts pan-specifically across BET family members — a limitation Farrelly et al. themselves flag — we interpret results as BET-family displacement and include parallel pull-downs of BRD2 and BRD3 to attribute effects.
The computational deliverable is a kinetics-aware enrichment method. From the mass spectrometry quantification we build a graph of BRD4-associated proteins whose edges carry measured intensities at each time point; each time point gets its own learned geometric embedding of that graph (node2vec, the technique GSPA uses to place proteins in a space where physical partners sit close together); gene sets are expanded at each time point to include proteins within a radius of the set's members, exactly generalizing GSPA, which recovers standard GSEA when the radius is zero; and the resulting time-dependent gene sets are fed into the linear mixed-effects trajectory modeling of Time-Course Gene Set Analysis for Longitudinal Gene Expression Data, which fits spline trajectories per gene set and tests them with likelihood ratios while properly handling replicate correlation and missing time points. Finally, Enrichment Map clustering condenses the redundant term lists into a small number of functional themes whose rise and fall can be read against the physical disassembly timeline. The key conceptual move is that gene-set membership itself becomes time-dependent and physically grounded, rather than fixed by a static database.
A secondary system tests generality: mouse embryonic stem cells exiting pluripotency, where Gas41 links histone acetylation to H2A.Z deposition and maintenance of embryonic stem cell identity showed that the YEATS-domain reader Gas41 physically recruits the SRCAP remodeling complex to bivalent promoters, and that a W93A point mutant kills the acetyllysine binding. Comparing wild-type and W93A Gas41 across differentiation gives us a genetic, non-pharmacological perturbation of reader-complex coupling, with the mutant serving as a built-in negative control.
Baselines
The naive reference point is a static endpoint contrast: pool the 0-hour and 24-hour samples, run a standard differential expression test, and interpret the resulting gene list with Fisher's exact overrepresentation analysis against Gene Ontology and KEGG, the classic recipe described in the functional enrichment literature. This is what most wet-lab papers actually do, and it is the right null to beat: if dense time sampling and physical networks add nothing, this two-point design should recover the same functional story. It will be run on the identical RNA-seq data.
The strongest existing method for the enrichment question is TcGSA, from Time-Course Gene Set Analysis for Longitudinal Gene Expression Data. It is the strongest competitor because it already solves the longitudinal half of the problem — in the DALIA-1 vaccine trial (19 patients, 14 time points) it identified 69 significant gene sets while standard univariate gene-by-gene analysis and standard time-series GSEA both failed to detect any significant pattern change over time — and it does so without needing any of our new physical measurements. We will run TcGSA on the same RNA-seq, with the same false-discovery-rate control and the same spline flexibility choices, and give it the same compute; our method beats it only if the physically structured, time-varying gene sets detect significant trajectory gene sets that TcGSA's fixed database sets miss, or recover the JQ1-rescued neurological disease programs that Farrelly et al. identified, with earlier or more precise timing. Note that Farrelly et al. report only that JQ1 "partially rescued" the deficits; no reported result found on Lacuna for the rescue effect size, so our 0/24-hour arms will establish that number ourselves.
The strongest network-aware competitor is GSPA, from Gene set proximity analysis, which outperformed standard GSEA in area under the precision-recall curve on 75.8% of the GEO2KEGG datasets it was validated on. We run GSPA with its STRING-derived static embeddings on our data, with the same radius tuning budget our method receives, to isolate whether the gain comes from having any network structure or specifically from a physically measured, time-varying one. Two ablations complete the ladder: our full pipeline with the Co-IP edges collapsed to a single static graph (isolating the value of temporal edge dynamics), and our full pipeline with the Co-IP edges replaced by static STRING edges (isolating the value of physical measurement over database curation).
For the edge-prediction question, the competitor is the DPPIN inference recipe itself: the three-sigma activity filter plus sliding-window Pearson correlation from Fu and He, applied to our RNA-seq to predict dynamic BRD4-complex edges, scored against our measured Co-IP edges time point by time point. On its home benchmark DPPIN reports 95.58% dynamic node classification accuracy for a Temporal Graph Attention Network versus 94.35% for a static Graph Attention Network on the DPPIN-Ito network; we will retrain and evaluate both on our mammalian data under matched compute and tuning, so the comparison is about the data regime, not the model.
Why not just…?
Why not just compare untreated versus treated endpoints? Because the literature shows endpoints actively mislead in exactly this kind of transition: Liu et al. found myocardial infarction responses shift from an inflammatory phase to a metabolically distinct recovery phase, and Yan et al. found toxic-insult footprints move from early metabolic disruption to late cell-cycle arrest — a late snapshot captures only the end state and an early one only generic stress. Our own lead-lag analysis makes this concrete and testable: if physical complex disassembly precedes transcriptional change for direct BRD4-dependent programs, endpoint designs cannot see the ordering at all, and our seven-point sampling measures it directly.
Why not just infer the dynamic interactions from RNA co-expression and skip the co-IP? This is the real threat, and we do not argue around it — we test it. Two lines of evidence say inference will fall short: PRING: Rethinking Protein-Protein Interaction Prediction from Pairs to Graphs shows that even 650-million-parameter protein language models score below 0.5 in graph similarity when reconstructing human interaction networks and achieve high precision but low recall on protein complexes, and the co-IP literature — for example Li et al., who used co-IP to prove IQGAP3 drives signaling through direct binding to GLI1 rather than through an intermediate — exists precisely because correlation cannot distinguish a physical machine from a multi-step cascade. But if, on our benchmark, the DPPIN-style inferred edges match the measured ones with high precision and recall at each time point, we will report that the expensive physical arm was confirmatory, and the benchmark's value then lies in being the first mammalian ground truth that lets anyone make that statement quantitatively. Either outcome is a result; we pre-register the comparison so it cannot be quietly dropped.
Why not just run standard differential expression plus GSEA? The DALIA-1 result is decisive on the longitudinal half: standard time-series GSEA detected zero significant pattern changes where TcGSA found 69 gene sets. But the stronger honest answer is that this objection is partially right — TcGSA alone may capture most of the transcriptomic signal, which is why it is our strongest baseline rather than a dismissed one. What standard workflows cannot do, by construction, is connect a detected transcriptional program to the physical machine that drives it: gene sets treated as bags of names carry no information about whether the encoded proteins actually co-assemble, and the known discordance between RNA dynamics and protein complex assembly means the mapping must be measured. Our static-Co-IP and static-STRING ablations are designed precisely so that, if the physical-temporal structure adds nothing beyond TcGSA, the ablation table will say so plainly.
Decisive Checks
Three findings would make the case. First, the edge benchmark: DPPIN-style inference and static database edges are scored against measured Co-IP edges per time point, with the hypothesis (from PRING's sub-0.5 reconstruction scores) that inference shows the same high-precision, low-recall failure on transient complex subunits. Second, the enrichment comparison: our method must detect FDR-significant trajectory gene sets that TcGSA and GSPA miss, and must recover the JQ1-rescued neurological disease programs of Farrelly et al. with correct temporal placement relative to measured BRD4-complex dissociation. Third, prediction rather than description: edges measured at time $t$ should predict enrichment changes at the next time point better than an RNA-only model, tested on held-out time points and on the Gas41 W93A mutant, where loss of acetyllysine binding — shown by Hsu et al. to abrogate SRCAP recruitment — must break the physical-transcript coupling our method claims to capture. If all three hold, dynamic network biology gains its first physically validated mammalian footing; if the inference comparison goes the other way, the field learns that equally important fact, and far more cheaply than anyone currently assumes.
References
- DPPIN: A Biological Dataset of Dynamic Protein-Protein Interaction Networks
- Chromatin profiling in human neurons reveals aberrant roles for histone acetylation and BET family proteins in schizophrenia
- Gene set proximity analysis: expanding gene set enrichment analysis through learned geometric embeddings
- Enrichment Map: A Network-Based Method for Gene-Set Enrichment Visualization and Interpretation
- Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9)
- HPF-MPI: A Programming System Supporting Task and Data Parallelism
- Time-Course Gene Set Analysis for Longitudinal Gene Expression Data
- Gas41 links histone acetylation to H2A.Z deposition and maintenance of embryonic stem cell identity
- Potential biomarkers of acute myocardial infarction based on weighted gene co-expression network analysis
- New Insights Into the Persistent Effects of Acute Exposure to AFB1 on Rat Liver
- PRING: Rethinking Protein-Protein Interaction Prediction from Pairs to Graphs
- IQGAP3 activates Hedgehog signaling to confer stemness and metastasis via up-regulating GLI1 in lung cancer