Long noncoding RNAs (lncRNAs) are RNA molecules longer than 200 nucleotides that do not code for protein. Many are thought to regulate other genes, either by soaking up microRNAs or by holding proteins together. Almost everything known about their function comes from two kinds of evidence. One is a single snapshot comparing diseased with healthy tissue. The other is a cell-culture experiment in which the lncRNA is knocked down or overexpressed far outside normal ranges. Neither shows how an lncRNA behaves inside a person when that person's physiology is deliberately changed and then allowed to return. A randomized crossover trial does exactly that: each participant receives both conditions in random order, separated by a "washout" gap, and so serves as their own control. Crossover trials treat anything that persists through the washout as a nuisance called carryover. We propose to treat that persistence as the measurement of interest instead. A regulator that acts through chromatin should leave a longer tail after the stimulus stops than one that acts on RNA or protein in the cytoplasm.
Concretely, we would add densely timed blood RNA sequencing to a two-period dietary crossover trial, sampling during the intervention periods and through the washout itself. We would then fit a functional mixed-effects model that splits each transcript's trajectory into four parts: a stable personal setpoint, a randomized response curve, a decay curve after the stimulus stops, and noise. This yields three quantities per lncRNA that no cross-sectional or cell-culture study can separate. The first is its personal baseline. The second is its causal response to the intervention. The third is how long that response persists. Within-person coupling between an lncRNA and its proposed targets then provides an in vivo test of whether a published mechanism operates at physiological levels.
Core Claim
The claim is that three separable, randomized quantities per transcript can be estimated with usable precision in a cohort of about 37 people, and that they carry mechanistic information. The first is a setpoint: a random intercept, meaning a per-person offset that holds across both periods. The second is an on-response curve. The third is an off-response, or persistence, curve. The persistence ratio, defined below, should differ systematically between lncRNAs whose proposed mechanism is cytoplasmic and those whose mechanism involves chromatin or super-enhancers. Coupling between lncRNA and target curves within the same person should follow the mechanism class: shared mRNA dynamics for microRNA sponges, and protein-without-mRNA dynamics for protein scaffolds. If persistence ratios turn out to be indistinguishable across classes, or if carryover is negligible everywhere, the proposal fails on its own terms, and that failure is itself a useful result for anyone designing omics crossover trials.
Why this gap is real
The lncRNA mechanism literature in the file is uniformly static. SNHG12 in renal cell carcinoma established a sponge axis (SNHG12 soaks up miR-199a-5p, protecting HIF1α mRNA) using 20 surgical tissue pairs and shRNA knockdown. That knockdown raised apoptosis from about 11% to about 72%, an effect far from physiological. GAS5 and YBX1 in stomach cancer showed a scaffold mechanism: GAS5 exon 12 binds YBX1 protein and shields it from degradation without changing YBX1 transcription, and stabilized YBX1 then induces p21. The clinical support, from 55 resections, is a Spearman correlation of 0.3279 between GAS5 and p21. FOXP4-AS1 in colorectal cancer likewise bridges USP7 to ZEB1 using specimens and cell lines. NEAT1 in mesangial cells is the most relevant to a metabolic intervention: high glucose induces NEAT1, which sponges miR-222-3p and regulates CDKN1B. All four papers motivate this work because they supply candidate axes with explicit, testable predictions. None observes those axes changing within a person over time.
The crossover literature constrains the design. The resistant starch trial ran 37 participants through 8-week periods with a 4-week washout, collecting metagenomics, metabolomics, MRI and clamp measurements. It reported a 2.8 kg mean weight loss and noted that physiological parameters drifted during the washout. It did not measure the host transcriptome, and it analyzed endpoints with conventional repeated-measures comparisons. The ice-bath time-of-day trial and the traffic-pollution nasal microbiome trial also used crossovers for hormones, lipids and microbes, but not for noncoding RNA. We reuse the resistant-starch design as our template because it is a crossover trial with an intervention strong enough to shift glycemia and adiposity, which is the regime where NEAT1-type glucose-responsive lncRNAs should move.
The longitudinal literature supplies the method. Berk et al.'s mixed-effects smoothing splines (MESS) models each gene's trajectory as a smooth population curve plus smooth subject-level deviations. On BCG vaccination data it moved the transient-peak gene TNF from rank 303 to rank 15 relative to EDGE, a model that assumes a constant shift over time. That paper studied a single challenge series of protein-coding microarray probes, with no crossover and no washout. LMM-VAE embeds a linear mixed model inside the latent space of a variational autoencoder and scales to high dimensions, but it has no design matrix for periods, sequences or carryover. Venkatesh et al. motivates keeping setpoint and change apart: they showed that the genetic loci governing BMI change are largely distinct from those governing baseline BMI. We ask whether the same separation holds for lncRNA levels, which would mean that some "dysregulated" lncRNAs in snapshot studies are really personal setpoints.
Research Program
Trial and sampling. The trial follows the template of a 2×2 randomized, double-blind crossover of resistant starch versus calorie-matched control starch. It has two 8-week periods separated by a 4-week washout, with a target of 37 participants. Peripheral blood mononuclear cells (PBMCs) are sampled at days 0, 3, 7, 14, 28 and 56 of each period, and at days 3, 7, 14 and 28 of the washout. This gives about 16 samples per person and roughly 600 total-RNA libraries. We also sample after the second period, so that carryover-free decay is observed in both sequences. Glucose and insulin measurements already collected in the template design serve as time-varying physiological covariates.
Quantification. Transcripts are quantified with Salmon against a comprehensive GENCODE annotation. The lncRNA quantification benchmark found that pseudoalignment tools such as Salmon produced 2 discordant expressed-lncRNA calls on simulated TCGA data, compared with 187 for HTSeq and 143 for featureCounts. They also detected about 16% of annotated lncRNAs versus 10–13% for alignment-based methods, and alignment missed or underestimated HOTAIR among others.
Model. For transcript $g$, person $i$ and calendar time $s$ since enrolment, we fit
$$y_{gi}(s) = b_{gi} + \mu_g(s) + \sum_{k \in \{A,B\}} \Big[ f_{gk}\big(s - s^{\text{on}}_{ik}\big)\,\mathbb{1}_{\text{on}} + h_{gk}\big(s - s^{\text{off}}_{ik}\big)\,\mathbb{1}_{\text{off}} \Big] + v_{gi}(s) + \epsilon_{gi}(s).$$
The terms are defined as follows. $b_{gi}$ is the person's random setpoint. $\mu_g(s)$ is a smooth period or secular trend shared by everyone. $f_{gk}$ is the population on-response curve to condition $k$, evaluated from that person's start date $s^{\text{on}}_{ik}$. $h_{gk}$ is the off-response curve after condition $k$ stops at $s^{\text{off}}_{ik}$. The indicator $\mathbb{1}_{\text{on}}$ equals one while condition $k$ is active, and $\mathbb{1}_{\text{off}}$ equals one after it has stopped. $v_{gi}(s)$ is a smooth person-level deviation, and $\epsilon_{gi}(s)$ is noise. Curves are penalized cubic splines, as in MESS.
Because the off-curve $h_{gk}$ extends into the next period, carryover is part of the model rather than an assumption. The model is identified because the two sequences place the off-curve of A over the on-period of B in one group and the reverse in the other, and because sampling during the washout observes $h$ directly. From the fitted curves we define a persistence ratio, $r_g = h_{gA}(28\text{ d}) / \max_t f_{gA}(t)$: the fraction of the peak diet-induced change still present at the end of the washout. We test differential response with an $L_2$ distance between the fitted A and B response curves, as MESS does, and calibrate it by permuting sequence labels.
Mechanism tests. For the NEAT1 → miR-222-3p → CDKN1B axis, a functioning sponge predicts that the within-person CDKN1B mRNA response curve lags and follows the NEAT1 curve. It further predicts that this coupling is specific: it should be weaker for mRNAs lacking miR-222-3p binding sites. We test this by comparing the coupling against a matched set of expression-matched mRNAs without the site. For GAS5 → YBX1 → p21, a scaffold predicts that GAS5 changes are followed by YBX1 protein changes, measured by targeted proteomics on PBMC lysates at six timepoints, with no YBX1 mRNA change, and then by p21 mRNA induction. The cytoplasmic versus nuclear location of candidates is checked by fractionation followed by qRT-PCR on a subset of samples, the protocol used in the file's sponge papers. The persistence hypothesis is tested by asking whether $r_g$ differs by fractionation class and by overlap with chromatin-regulatory lncRNAs. The latter are defined by transferring the class labels of lncRNAs with published chromatin or $m^6$A roles in the file, such as USP30-AS1 and BDNF-AS, to PBMC-expressed lncRNAs with similar chromatin association. We state openly that this labeling is coarse.
Baselines
The naive reference point is the textbook crossover analysis. For each transcript, it fits treatment, period and sequence effects to period-end values only (day 56), followed by standard false-discovery control. It uses the same Salmon counts and the same cohort. It is the analysis most trials would run, and it matches the conventional repeated-measures endpoint comparisons of the resistant starch trial. No reported result for lncRNA crossover data was found on Lacuna. Beating it means detecting more transcripts at matched empirical false discovery rate on simulated data, and producing a less biased treatment estimate when carryover is simulated.
The strongest existing longitudinal method is MESS, applied separately to each period and treated as a two-condition time course, which is how it was designed to be used. On its synthetic 100,000-gene benchmark, MESS reports 85.1% power at 90% specificity versus 70.4% for EDGE (Berk et al.). We rebuild that simulator, extend it with a crossover schedule, washout sampling and carryover tails, and reproduce the MESS-versus-EDGE gap as a sanity check before adding our model. Both models get the same spline basis, the same smoothing-parameter selection procedure and the same compute. Beating MESS means higher power at 90% specificity specifically in the scenarios with non-zero carryover. We expect a tie when carryover is zero, and a tie there is an honest outcome, not a failure.
For high-dimensional joint modeling we include LMM-VAE with treatment and period as covariates, fit to the whole transcript matrix. Its reported numbers are on Health MNIST (imputation MSE 0.002) and Rotating MNIST (0.014); no reported result was found on Lacuna for RNA-seq or crossover data. We compare it on held-out-timepoint imputation of lncRNA expression, with equal hyperparameter-search budgets.
Three ablations isolate the key ingredients. The first removes the off-curve $h$, forcing the usual no-carryover assumption. The second keeps $h$ but drops the four washout samples, which tests whether dense washout sampling is what makes persistence estimable. The third replaces Salmon with featureCounts, which measures how much of the lncRNA signal the pipeline choice alone destroys.
Why not just…?
Why not just run a longer washout and use the standard crossover analysis? Without measuring decay, there is no way to know what length is long enough. The resistant starch trial found drift during its 4-week washout. For chromatin-linked transcripts, persistence may exceed any practical washout, and the persistence is precisely the quantity we want to observe. That said, if the no-carryover ablation fits held-out data as well as the full model and the estimated $h$ curves are indistinguishable from zero, this alternative wins. We would then report that simple crossover analysis suffices for PBMC lncRNAs.
Why not just run a parallel-group longitudinal trial? Between-person variance dominates for lncRNAs, so a parallel design needs many more participants for the same power, as the PRIZE sub-study design argues in the clinical setting. More fundamentally, a parallel design cannot estimate a person's setpoint separately from their response, because each person sees only one condition. We quantify the cost directly by re-analyzing period 1 alone as a parallel trial and comparing standard errors.
Why not just validate mechanisms in cell lines, as the SNHG12 and GAS5 papers did? Cell-line knockdown shows what an lncRNA can do at extreme doses. It cannot show whether the same coupling appears at the modest, physiologically driven changes seen in blood. We do not claim to replace wet-lab validation. The in vivo coupling test is a filter that says which published axes are worth pursuing in human tissue. A null coupling for a sizable within-person lncRNA response is informative, and the matched no-site control mRNAs guard against mistaking shared co-regulation for sponging. This last point is a real threat: co-regulation by the diet itself could produce coupling. The time-lag and binding-site specificity tests are what distinguish the two explanations.
Why not just use an endpoint-only mixed model with a random slope? This is the most serious competitor. If diet-induced lncRNA changes are sustained shifts rather than transient spikes, a random-slope model on a few timepoints will perform as well at much lower sequencing cost. The simulation comparison and a real-data comparison on the held-out-timepoint task settle it. If the spline model does not beat random slopes on real data, we would recommend the sparse sampling schedule for future trials.
Proposed Experiments
The first experiment is the simulation benchmark described above. It covers transient spikes, sustained shifts, and carryover tails ranging from zero to half the peak response, and reports power at 90% specificity plus bias in the treatment curve for all baselines and ablations. The second is an internal replication check on the real trial data. The on-response curve estimated only from the AB sequence's first period should match the curve estimated from the BA sequence's second period once $h$ is included, and should mismatch when $h$ is removed. This comparison is the most direct evidence that carryover is real and correctly modeled. The third tests setpoint stability: the correlation of each person's fitted $b_{gi}$ across periods, and its correlation with response magnitude, testing whether setpoint and response are distinct as Venkatesh et al. found for BMI. The fourth is the pair of mechanism coupling tests for NEAT1/CDKN1B and GAS5/YBX1/p21, together with the persistence-by-mechanism-class comparison of $r_g$.
The decisive result is the combination of three findings. The sequence-replication check must pass only with the off-curve included. Persistence ratios must separate by localization class. At least one published axis must show its predicted coupling signature in vivo. If all three hold, carryover becomes a measurable property of lncRNAs rather than an obstacle to crossover analysis.
References
- Overexpression of SNHG12 regulates the viability and invasion of renal cell carcinoma cells through modulation of HIF1α
- lncRNA GAS5 enhances G1 cell cycle arrest via binding to YBX1 to regulate p21 expression in stomach cancer
- LncRNA FOXP4-AS1 facilitates colorectal cancer invasion and migration by enhancing USP7 interaction with ZEB1
- LncRNA NEAT1 Promotes High Glucose-Induced Mesangial Cell Hypertrophy by Targeting miR-222-3p/CDKN1B Axis
- Resistant starch intake facilitates weight loss in humans by reshaping the gut microbiota
- Effects of time-of-day on the noradrenaline, adrenaline, cortisol and blood lipidome response to an ice bath
- Effects of Short-Term Traffic-Related Air Pollution Exposure on Nasal Microbiome in Young Healthy Adults: A Randomized Crossover Controlled Trial
- Longitudinal analysis of gene expression profiles using functional mixed-effects models
- Latent mixed-effect models for high-dimensional longitudinal data
- Characterising the genetic architecture of changes in adiposity during adulthood using electronic health records
- Benchmark of lncRNA Quantification for RNA-Seq of Cancer Samples
- Nuclear and cytoplasmic USP30-AS1 coordinately regulate breast cancer progression through HnRNPF/p21 and EZH2/c-Myc/p21 axes
- Apelin-13-Mediated Upregulation of METTL3 Ameliorates Alzheimer’s Disease via Inhibiting Neuroinflammation Through m6A-Dependent Regulation of lncRNA BDNF-AS
- 3 Rationale and design of the Medical Research Council Precision medicine with Zibotentan in microvascular angina (PRIZE) trial MRI sub-study