Most genetic variants that raise risk for common immune diseases do not change any protein's amino-acid sequence. They sit outside coding regions and act like dimmers on gene expression — a view supported both by fine-mapping showing most causal GWAS variants are non-coding Discovery and fine-mapping of adiposity loci using high density imputation of genome-wide association studies in individuals of African ancestry: African Ancestry Anthropometry Genetics Consortium and by the finding that non-coding and synonymous variants carry disease-association effect sizes comparable to coding ones Non-Synonymous and Synonymous Coding SNPs Show Similar Likelihood and Effect Size of Human Disease Association. Time-course RNA measurements have recently shown something sharper: for 23% of genes with expression QTLs during T-cell activation, the dimmer only turns during a specific window of the response, and the effect is masked or attenuated if you look at static time points Division state reveals hidden genetic regulation during T cell activation and identifies immune disease–linked gene programmes. But that result, like nearly all genetic mapping of regulation, stops at mRNA.
Here is the problem. The molecules that actually coordinate an immune response are secreted proteins — cytokines such as IL-2, TNF, and IFN-γ — and mRNA levels are an unreliable proxy for them: translation, secretion, and extracellular degradation each add their own delays and buffering. Meanwhile, the one large study that mapped genetics to a circulating protein with an immunoassay Genome-wide studies reveal factors associated with circulating uromodulin and its relationships to complex diseases used a single static blood draw per person. Nobody has asked whether non-coding variants shift the timing of protein output — when a cytokine peaks, and how long it lasts — because nobody has run serial protein measurements across a genotyped, time-resolved cell-activation cohort. We propose to do exactly that, using the enzyme-linked immunosorbent assay (ELISA — a method that traps a target protein between two antibodies recognizing different parts of it, then uses an enzyme reaction to amplify the binding event into a measurable color signal) as the readout, because its two-antibody requirement guards against the epitope confounds that corrupt single-binder platforms.
Core Claim
A meaningful fraction of non-coding regulatory variants act on immune effector proteins through kinetics — shifting time-to-peak or duration of cytokine production — rather than through average abundance, and these variants are invisible to both static protein QTL designs and to current sequence-based variant-effect models. Serial ELISA sampling across a genotyped activation time course can map these "kinetic protein QTLs" (kinetic pQTLs) and, by pairing them with matched single-cell RNA trajectories, can measure how much of the dynamic mRNA regulation documented by Division state reveals hidden genetic regulation during T cell activation and identifies immune disease–linked gene programmes actually survives translation and secretion to reach the protein output the immune system uses.
Research Program
The study has three linked components. First, the cohort and assay design: we will recruit roughly 150 genotyped donors (deliberately exceeding the 119 individuals of Gozzard et al., whose own power analysis flagged small-effect interactions as undetectable at that scale), isolate naive CD4+ T cells, stimulate them under a uniform protocol, and collect culture supernatants at six to eight time points spanning 4 to 96 hours. Each supernatant will be quantified by sandwich ELISA for a focused panel — IL-2, TNF, IFN-γ, IL-10 — chosen because these are the effectors whose transcriptional dynamics the T-cell activation literature already implicates. In parallel, single-cell RNA-seq will be run at three to four matched time points, with cell division state inferred from transcriptomes in the manner of Gozzard et al.'s CellDivider model (a small neural network that predicts how many times a cell has divided from its RNA profile, reported at Pearson r = 0.881 on held-out data).
Second, the kinetic model and mapping. For each donor and cytokine we fit a two-parameter pulse to the serial ELISA curve,
$$p(t) = A \cdot \frac{t}{\tau}\, e^{1 - t/\tau},$$
where $A$ is the peak concentration and $\tau$ is the time at which the peak occurs, so that amplitude and timing are separated by construction. Each candidate non-coding variant $g$ is then tested against $\log A$ and $\log \tau$ separately. A variant with a significant effect on $\tau$ but not $A$ is a pure timing variant — precisely the class that a single-time-point design cannot see, because its effect depends entirely on when you look. Candidate variants come from three sources: lead SNPs and cis-eQTLs from Gozzard et al. (their own scan was restricted to these for computational reasons), decay-regulating variants (the NMD-QTLs mapped across 48 tissues by Mapping genetic variants for nonsense-mediated mRNA decay regulation across human tissues, which were 2.54 times more likely to be disease-associated than ordinary eQTLs and which their authors explicitly never applied to time courses), and 3'-UTR microRNA-site variants of the kind shown to tune transcript levels post-transcriptionally in lymphocyte lines miRNA-Mediated Relationships between Cis-SNP Genotypes and Transcript Intensities in Lymphocyte Cell Lines. The decay-QTL class carries a specific prediction: decay variants should shift protein amplitude (less mRNA means less translated protein over the window) whereas promoter or enhancer timing variants should shift $\tau$, giving a mechanistic signature readable straight off the ELISA curves.
Third, the model benchmark. We will take the lead kinetic-pQTL variants and ask whether current sequence-to-function models predict their effects. AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model reaches 80% auROC on predicting the direction of static eQTL effects (improved from 75% in prior models) but its authors state that condition-specific variant effects remain hard; Identifying non-coding variant effects at scale via machine learning models of cis-regulatory reporter assays reaches Pearson r ≈ 0.88–0.89 on episomal reporter activity but only r ≈ 0.69–0.76 on allelic skew, and its authors concede the episomal format cannot represent UTRs, repressors, or native chromatin. Our measured kinetic effects become a new evaluation set: do these models' predicted effect directions track the measured $\Delta A$ and $\Delta\tau$? We expect a sharp drop, which itself is a publishable, quantified negative result about where sequence models stand.
Baselines
The naive reference point is the static protein QTL design: a single ELISA measurement per donor at one late time point, analyzed exactly as Genome-wide studies reveal factors associated with circulating uromodulin and its relationships to complex diseases analyzed its cross-sectional serum draws. This is the field's default, so it must be run on our own data (same donors, same assays, same variant set, and the same multiple-testing correction) rather than cited from elsewhere; it will be beaten if it recovers significantly fewer loci than the kinetic design, or recovers the same loci with biased effect estimates. The strongest existing competitor is the dynamic mRNA-only approach of Gozzard et al. — single-cell RNA time courses with division state as an interaction term — used as a proxy for protein output: we will take their pipeline's dynamic eQTL effects on cytokine genes as a predictor of our measured ELISA curves and ask how much protein-level variance it explains. Their CellDivider inference layer is well validated (Pearson 0.881, MAE 0.428 on 25,694 held-out cells), but the mRNA-to-protein step is exactly what is untested; beating this baseline means showing measured protein kinetics deviate from the RNA-implied kinetics for a non-trivial set of loci in a direction predictable from decay-QTL annotations. The strongest in-silico competitor is AlphaGenome, which we will run on all lead variants under its released configuration, scored on effect-direction accuracy against our measured effects; beating its reported 80% auROC on static eQTL sign prediction is not the target — matching that number on kinetic-window variants would be — so the comparison is degradation, not superiority. Finally, an ablation isolates the proposal's key ingredient: the same cohort, the same ELISA data, but with the time dimension collapsed — pooling supernatant measurements into a single per-donor average before mapping. If the pooled analysis finds everything the kinetic analysis finds, the timing parameterization is unnecessary; the ablation exists to fail.
Why not just…?
Why not just use existing static eQTL and pQTL databases? Because the masking is already quantified: Gozzard et al. found 23% of eQTL genes lost their regulatory effects when analyzed at static activation time points rather than along the trajectory, and the RPS26 autoimmune locus acted only in early division states. Static protein designs inherit the same blindness, plus the RNA-to-protein gap. Our own ablation (pooled single-time-point mapping on the same data) directly measures how much a static design would miss here.
Why not just infer protein levels from the RNA-seq? This is the most serious alternative, and we treat it as such rather than arguing around it. Evidence says mRNA is an incomplete proxy: decay-regulating variants decouple transcript abundance from functional output Mapping genetic variants for nonsense-mediated mRNA decay regulation across human tissues, 3'-UTR variants repress translation without proportional mRNA change miRNA-Mediated Relationships between Cis-SNP Genotypes and Transcript Intensities in Lymphocyte Cell Lines, and in the sepsis setting tissue transcriptomes failed to explain roughly half of observed cellular changes driven by circulating signals A pairwise cytokine code explains the organism-wide response to sepsis. But for some cytokines RNA may indeed track protein well. The proposal settles this empirically rather than by assumption: the per-donor mRNA-trajectory versus ELISA-curve comparison is a headline result either way, and a finding that RNA is a good proxy for IL-2 but a poor one for IFN-γ would itself be useful.
Why not just run MPRAs or trust AlphaGenome in silico? Both models' own papers bound their reach: MPAC's episomal format cannot represent UTRs, repressors, or chromatin context, and its allelic-skew accuracy of r ≈ 0.69–0.76 leaves most of the variance in variant impact unexplained Identifying non-coding variant effects at scale via machine learning models of cis-regulatory reporter assays, while AlphaGenome's authors state that condition-specific variant effects remain challenging and its 80% eQTL sign auROC leaves one direction call in five wrong AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model. Neither has ever been evaluated against time-resolved protein measurements, which is precisely the evaluation we supply. They are baselines here, not substitutes.
Why not just use a multiplex aptamer platform instead of one-analyte-per-well ELISA? Because the aptamer format is known to confound post-translational state with abundance: in the uromodulin study the antibody assay found 3 loci including the true cis-regulatory switch, while the aptamer assay found 10 loci dominated by glycan-modifying enzymes that alter binder recognition rather than protein amount Genome-wide studies reveal factors associated with circulating uromodulin and its relationships to complex diseases. Cytokines undergo proteolytic processing and multimerization during activation, exactly the conditions under which single-binder platforms mislead. The honest cost is throughput and sample volume; we mitigate it by keeping the panel small and by validating a subset of donor-supernatant measurements with an independent second antibody pair per cytokine, which also serves as our control for the epitope-masking confound.
Decisive Checks
Three results decide the proposal. First, replication: lead kinetic pQTLs identified in a random half of donors must hold in the held-out half, with the split-half analysis pre-registered per cytokine and per kinetic parameter. Second, the static-versus-kinetic gap: the pooled single-time-point ablation must miss a quantifiable fraction of replicated loci — if the fraction is near zero, the core claim fails and the honest conclusion is that timing variants are rarer than the mRNA dynamics of Gozzard et al. suggested. Third, the proxy question: the correlation between RNA-implied and ELISA-measured trajectories, reported per cytokine per locus, determines whether the field can keep leaning on transcriptomics for secreted effectors; the decay-QTL-annotated variants are the pre-specified group where we expect the RNA proxy to break first. Sample-volume permitting, an exploratory arm extends the design to a second stimulus condition, testing whether the same variants show different kinetic signatures under different activation contexts — the condition-specificity that AlphaGenome's authors name as an open problem. No reported result found on Lacuna for any existing serial-ELISA kinetic pQTL study in activated human T cells, which is the gap this proposal occupies; the closest prior work, the developmental ELISA time series of Changes in Gonadal Sex Differentiation, Digestive Enzymes, and Growth-Related Hormone Contents in the Larval and Juvenile Black Scraper, Thamnaconus modestus, measured hormones across 91 days in a fish but with no genetic variation component, and no reported numerical summary statistic for its hormone trajectories is available in the file.
References
- Discovery and fine-mapping of adiposity loci using high density imputation of genome-wide association studies in individuals of African ancestry: African Ancestry Anthropometry Genetics Consortium
- Non-Synonymous and Synonymous Coding SNPs Show Similar Likelihood and Effect Size of Human Disease Association
- Division state reveals hidden genetic regulation during T cell activation and identifies immune disease–linked gene programmes
- Genome-wide studies reveal factors associated with circulating uromodulin and its relationships to complex diseases
- Mapping genetic variants for nonsense-mediated mRNA decay regulation across human tissues
- miRNA-Mediated Relationships between Cis-SNP Genotypes and Transcript Intensities in Lymphocyte Cell Lines
- AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model
- Identifying non-coding variant effects at scale via machine learning models of cis-regulatory reporter assays
- A pairwise cytokine code explains the organism-wide response to sepsis
- Changes in Gonadal Sex Differentiation, Digestive Enzymes, and Growth-Related Hormone Contents in the Larval and Juvenile Black Scraper, Thamnaconus modestus