Kinetic Reporter Labels for Inferred Gene-Regulation Edges

Modern biology has a strange asymmetry. High-throughput experiments can list thousands of genes whose activity rises and falls together over time, in response to a drug or during disease, but they cannot say which gene is directly turning which other gene on — and which pairs merely move together because they sit downstream of a shared cause. Meanwhile, two low-throughput laboratory assays answer exactly this question, but only one gene pair at a time: the reporter assay, in which a candidate gene's regulatory switch is glued to a light-emitting enzyme so that its firing rate becomes measurable photon flux, and co-immunoprecipitation (Co-IP), in which an antibody drags one protein out of a cell lysate to see which other proteins are physically stuck to it. No study in the corpus we reviewed has ever used these two assays, at scale and over time, as a labeled ground-truth set for the edges that computational network models infer. That is the proposal: build a biochemically labeled truth set of a few hundred inferred regulatory edges, each classified by kinetic reporter assays and Co-IP as direct promoter regulation, indirect cascade, post-transcriptional effect, or physical-complex-mediated cooperation, and then use that truth set to test, calibrate, and repair the computational pipelines — and to show that time-resolved features make biomarker panels better than the static abundance measures used today.

The motivation is not abstract. The benchmarking literature in this corpus contains two documented crises that such a truth set would directly attack. First, the "interventional paradox": on over 400,000 interventional single-cell samples, causal discovery algorithms failed to consistently beat simple correlation baselines at reconstructing gene networks, and the benchmark authors trace the failure partly to ground-truth databases whose labels are unreliable CausalBench. Second, effect-size inflation: deep "virtual cell" models predict perturbation effects roughly 23-fold larger than reality, and no deep model beat a sparse linear comparator on causal recovery Virtual Cell Models Inflate Perturbation Effect Sizes and Undermine Causal Gene Regulatory Network Recovery. Both crises are ground-truth problems. A truth set built from the two assays that measure the actual physical event — promoter firing and protein complex assembly — is the missing instrument.

Core Claim

The claim is that a modest, systematically chosen panel of biochemical assays can produce, for the first time, a labeled set of temporal regulatory edges good enough to answer three questions that no existing study can: (1) when a perturbation-data method infers an edge, how often is that edge a direct promoter interaction rather than an indirect ripple — the exact failure mode the BetterBoost authors concede, noting their method "might incorrectly infer direct edges between genes that are indirectly connected in a causal chain" BetterBoost - Inference of Gene Regulatory Networks with Perturbation Data; (2) by how much do inferred effect sizes deviate from measured promoter dose-response curves, extending the inflation finding from deep models to perturbation-data methods; and (3) do kinetic features of a biomarker — time of activation onset and rate of change, rather than abundance at a single clinic visit — improve early detection in a staged clinical funnel, where today's best practice, a static five-miRNA serum panel for breast cancer, reports an AUC of 0.948 with 100% specificity and 79% sensitivity Diagnostic value of 5 miRNAs combined detection for breast cancer.

Research Program

The program has three linked components. The first is computational: run simple, benchmark-vetted edge-inference methods — Pearson correlation and GRNBoost2, which the living benchmark geneRNIB found consistently outperforming complex multiomic pipelines and foundation models geneRNIB: a living benchmark for gene regulatory network inference, plus BetterBoost on the interventional data BetterBoost — on time-series perturbation data. The primary discovery system is the A549 lung-cancer 72-hour drug-perturbation time course used by FLeCS, a system where a learnable differential-equation model showed that only a small subset of nearly a million putative links actually drives state transitions A scalable gene network model of regulatory dynamics in single cells. We deliberately do not build a new deep model; the corpus evidence rules that out, and the interesting uncertainty is in the labels, not the learner.

The second component is the truth set itself, and it is the heart of the proposal. From the inferred edges we select roughly 150–200 targets, stratified by inference score (high-confidence, mid, and deliberately borderline), and for each we run a kinetic dual-luciferase assay: the target gene's own promoter driving a destabilized, short-lived luciferase, co-transfected with a constitutive control reporter whose ratiometric quotient cancels transfection noise — the standard design U4 at the 3′ UTR of PB1 Segment of H5N1 Influenza Virus Promotes RNA Polymerase Activity and Contributes to Viral Pathogenicity — but measured at multiple time points after the regulator is perturbed (CRISPRi or siRNA), not at a single endpoint. This is the crucial design choice. The classification logic is temporal: if the reporter's firing rate changes promptly after regulator perturbation while the endogenous transcript follows with a lag, the edge is direct promoter regulation; if the endogenous mRNA shifts while the reporter stays flat, the edge is post-transcriptional — exactly the regime where the field's most instructive contradiction lives. In the PIK3CA study, the reporter assay showed the protective T allele creating a miR-520a binding site that repressed reporter output, yet patients carrying the risk CC genotype had higher endogenous PIK3CA mRNA, a result the authors called seemingly counterintuitive A Functional Variant at miR-520a Binding Site in PIK3CA Alters Susceptibility to Colorectal Cancer in a Chinese Han Population. Our truth set is designed to make that class of contradiction measurable rather than anecdotal. For edges where two inferred regulators are predicted to cooperate, we add Co-IP, plus living-cell proximity assays to guard against lysis artifacts, to test whether the regulators physically co-assemble — because physical binding and regulatory output are orthogonal facts, as the IQGAP3–GLI1 work showed by proving direct contact that transcript correlation alone could not establish IQGAP3 activates Hedgehog signaling to confer stemness and metastasis via up-regulating GLI1 in lung cancer.

The third component closes the loop with a staged clinical funnel, replicating the three-stage design of the breast-cancer miRNA study (public-data screening, small testing cohort, large independent validation) but replacing static abundance with the kinetic features the truth set endorses. The model system is cisplatin-induced acute kidney injury, where the corpus documents a mechanism (NAT10 binding DDX17 in tubular epithelial senescence) validated by Co-IP but never assessed as a longitudinal clinical signal NAT10 Promotes Tubular Epithelial Cell Senescence in Cisplatin-Induced Acute Kidney Injury by Regulating DDX17; no reported result found on Lacuna for any diagnostic AUC in that setting, which is precisely the gap. Serial plasma samples during cisplatin therapy are clinically routine, making onset time and rate of change of candidate markers measurable in a testing cohort and then an independent validation cohort.

Baselines

The naive reference point is the standard static funnel: differential expression between cases and controls at one time point, RT-qPCR verification, and independent-cohort validation, exactly as executed for the five-miRNA breast-cancer panel, which reports an AUC of 0.948 with 100% specificity and 79% sensitivity Diagnostic value of 5 miRNAs combined detection for breast cancer. It is the strongest competitor because it is the field's working gold standard and its numbers define the bar for any staged panel. We will run it on the same cisplatin-cohort samples, at the same sequencing budget, using abundance at the standard clinical sampling point. Beating it means higher AUC, or equal AUC with detection earlier than the clinical gold standard of serum creatinine rise — because an early-warning claim is the only honest justification for the added cost of serial sampling.

The strongest existing method on the edge-classification question is computational-only inference: trust the top-ranked edges from perturbation-data methods as direct regulations. The leading instance is BetterBoost, which on the Replogle RPE1 Perturb-seq data at a 100% fraction of interventional data reported an average Wasserstein distance of 0.636, against 0.106 for GRNBoost returning 1000 edges and 0.130 for DCDI, figures quoted as reported in that paper BetterBoost. We will run these methods on our identical time-course data with the same compute and tuning budget, rank edges by their scores, and evaluate the scores against our biochemical labels. This baseline cannot be "beaten" by us in the usual sense — our contribution is the evaluation itself — so the decisive result is a measured accuracy number: the area under the receiver operating curve of computational score against directness label. If that number is very high, the proposal's practical value collapses to the clinical stage; if it is mediocre, the truth set becomes the necessary instrument. We state this openly because the experiment settles it either way.

The key ablation isolates the temporal ingredient: identical reporter constructs, identical cell system, but read out at a single endpoint time point, exactly as the PIK3CA and H5N1 studies did PIK3CA, H5N1 PB1. If endpoint-only readouts classify edges as accurately as kinetic readouts, the temporal claim of this proposal is refuted and the simpler assay suffices; we would then narrow the proposal to the labeling and calibration claims, which do not depend on it. A second ablation drops Co-IP and classifies edges by reporter evidence alone, to measure how many physical-complex-mediated cooperations the reporter layer mislabels.

Why not just…?

Why not just look up protein interactions in STRING or infer them from co-expression? Because the corpus shows the computational and database routes fail on exactly this question. The CausalBench evaluation found that ground-truth interaction databases are often not robustly above chance for context-specific networks CausalBench, and the IQGAP3–GLI1 study is a concrete case where transcriptomic correlation could not distinguish direct physical activation from a multi-step enzymatic cascade until Co-IP settled it IQGAP3. The truth set needs the physical assay, not the lookup.

Why not just measure endogenous mRNA by RT-qPCR or RNA-seq? Because steady-state mRNA conflates transcription initiation, splicing, export, and degradation, and the field's own data shows the conflation is not hypothetical: the PIK3CA study found reporter repression and endogenous elevation pointing in opposite directions for the same allele PIK3CA, and the H5N1 work needed a reporter to isolate polymerase initiation from replication feedback, showing the C4 variant at only about 37% of U4 activity at 37°C H5N1 PB1. We measure endogenous transcript alongside the reporter precisely so the truth set quantifies how often the two disagree.

Why not just use a deep virtual cell model to predict the dynamics? Because two independent benchmark efforts in this corpus reject that route: no deep model beat a sparse linear comparator on causal recovery, with roughly 23-fold effect-size inflation PerturbCausal, and simple correlation methods consistently beat foundation models in the living benchmark geneRNIB. We follow that evidence, use simple inference, and spend the budget on biochemical labels instead of parameters. Our calibration experiment does give deep models a second chance — if reporter-measured dose-response curves, used as training signal, fix their inflation, that would be a genuinely new result.

Why not just compare before and after (two time points)? Because collapsing the timeline averages mutually antagonistic regulatory programs: STOIC showed transcription-factor motifs operating in distinct temporal waves during neuronal differentiation A genome-wide, machine learning-guided exploration of the cis-regulatory code involved in neuronal differentiation, and myocardial infarction profiling shows acute inflammation giving way to metabolic recovery programs that a snapshot merges indistinguishably Zhihua Liu et al.. This is a real threat rather than a straw man, which is why the endpoint-only ablation is built in: if two-point readouts classify edges as well as kinetic ones, the temporal ingredient adds nothing and we will say so.

Proposed Experiments and Decisive Checks

Experiment one builds the truth set: infer edges on the A549 72-hour drug time course with GRNBoost2, BetterBoost, and Pearson baselines; select ~180 stratified target edges; run kinetic dual-luciferase with destabilized reporters at six post-perturbation time points; add Co-IP and proximity assays for the ~40 edges with predicted regulator cooperativity. The decisive outputs are the directness accuracy of each computational method (the number nobody currently has) and the distribution of ratio between inferred and reporter-measured effect sizes, to test whether the 23-fold inflation documented for virtual cell models PerturbCausal also afflicts perturbation-data inference. A sanity anchor comes from FLeCS, whose mechanistic interpretation matched known biology in 11 of 12 K562 knockdowns against differential expression analysis FLeCS; edges where FLeCS and our labels agree or conflict become case studies.

Experiment two resolves the PIK3CA-type contradiction mechanistically: for post-transcriptional-classified edges, Co-IP of Argonaute-containing complexes tests whether the discrepancy between reporter repression and endogenous abundance traces to ribonucleoprotein assembly, in the spirit of the TaBZR2.1–AGO4 physical bridge between a transcription factor and the silencing machinery Identification of the Brassinazole-Resistant (BZR) Gene Family in Wheat.

Experiment three runs the staged clinical funnel in the cisplatin-cohort setting: temporal-feature panel (onset time and slope of the top mechanistically validated candidates) versus the static abundance funnel, in a testing cohort and then an independent validation cohort, reporting AUC, sensitivity, specificity, and time-to-detection relative to serum creatinine. Success is beating the static funnel's numbers on the same samples, judged against the 0.948 AUC / 100% specificity / 79% sensitivity benchmark that the best static funnel in this corpus achieved in breast cancer Diagnostic value of 5 miRNAs combined detection for breast cancer. Failure modes are pre-registered: if kinetic features add nothing over endpoints, or if computational directness scores already exceed roughly 0.9 area under the curve, the respective claims are abandoned and the truth set is released as a public benchmark for the field — which, given that CausalBench's authors traced the interventional paradox to unreliable ground truth, would itself be a contribution no existing study provides.