{"paper_id":"922d18a1-8e49-44a9-a7bd-ee192f72bfb5","body_text":"Evaluating the Identifiability of Temporal Delay and\nResponse Attenuation in Endometrial Decidualization\nElena Krikun, Wael Fahmy, Abedalrhman Alkhateeb\nLakehead University\nThunder Bay, Ontario, Canada\nAbstract\nEndometriosis is associated with impaired endometrial decidualization, yet conventional differential expres-\nsion analyses cannot readily distinguish delayed transcriptional activation from reduced response amplitude.\nThese mechanisms may produce similar expression differences at individual time points while implying\ndistinct biological interpretations. Establishing when they can be distinguished is therefore a computational\nidentifiability problem.\nWe implement an identifiability-aware analysis of temporal displacement and amplitude attenuation\nin longitudinal transcriptomic responses. The approach combines reference trajectories, constrained\namplitude-only and amplitude-plus-delay models, analytical examples of observational equivalence, and\nsimulation-based assessments of mechanism distinguishability. Held-out time-point evaluation and culture-\nlevel sensitivity analyses examine the stability of model-dependent conclusions.\nIn GSE246591, a candidate-delay-dependent dynamic-range filter excluded 30.9–38.3% of initially dynamic\nfeatures under a hypothetical 24-hour shift across four endometrial stromal cultures. A two-component\ndiagnostic separating retained dynamic range from affine shape similarity revealed that small projection\nresiduals can arise from either loss of observable dynamics or similarity of response shapes. These findings\ndemonstrate why a single separation statistic may be insufficient for interpreting temporal identifiability.\nComplementary independent in vivo analyses provided biological context. In GSE51981, an endometrial\nprogesterone-response-associated transcriptional score was reduced in endometriosis samples (phase-model\nestimate 0.155, 95% CI 0.222 to 0.088). GSE6364 showed a directionally consistent but statistically\ninconclusive difference (0.033, 95% CI 0.068 to 0.002). Neither observation independently establishes a\ntemporal delay.\nTogether, these results support selection-aware evaluation of temporal transcriptomic responses, distin-\nguishing structural ambiguity and loss of dynamic signal from evidence of a biological timing change.\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted November 4, 2026. ; https://doi.org/10.64898/2026.11.03.764027doi: bioRxiv preprint \n\n1 Introduction\nDecidualization is the hormone-associated differentiation of human endometrial stromal cells that prepares\nthe uterine lining for implantation [1]. An attenuated decidualization-associated transcriptional signal in\nendometriosis has been reported in multiple endometrial studies [2, 3]. Yet a lower signal at one sampling time\ncan arise from a smaller response, a later response, altered trajectory shape, menstrual-cycle composition, or\nan unmeasured cellular mixture. Distinguishing those explanations requires information beyond cross-sectional\ndifferential expression. Menstrual-cycle-associated confounding further complicates endometrial comparisons\n[4].\nTemporal gene-expression methods answer related but distinct questions.DEtime estimates when two\nperturbation-specific time courses diverge using Gaussian processes [5];ImpulseDE2 tests and models time-\ndependent expression using an impulse family [6]. Neither method, without a suitable matched design and\nspecified nuisance parameters, establishes that a lower single-time signal is specifically a delayed version of\nanother response. Identifiability analysis asks whether different parameter settings can generate the same\nobservable data, a question with a long history in dynamical modeling [7, 8].\nWe operationalize the distinction in two steps. First, we identify when a shifted curve can be represented\nexactly by changing only its amplitude and baseline (observational aliasing). Second, we quantify the\nremaining separation on a verified endometrial RNA-sequencing schedule, with explicit checks of expression\nassayability, reference-curve assumptions, and uncertainty limits. The resulting audit separates algebraic\nnonidentifiability from what measured expression values can establish. A supplementary software example\ndocuments a related computational workflow, but does not serve as biological validation.\n2 What does it mean to identify a temporal delay?\n2.1 A minimal model and an exact criterion\nFor sampling timesT ={t1,...,tm}and a specified reference expression profilef (t), compare an amplitude-\nonly response with an amplitude-plus-shift response:\nM0 :y(t i) =b 0 +a 0f(ti) +εi,(1)\nMτ: y(ti) = bτ+aτf (ti−τ) +εi. (2)\nHereb is the baseline expression level,a is the response amplitude, andτis a displacement relative to a chosen\nreference trajectory,not an experimentally observed molecular-onset time. Theεi represent measurement\nerror or departures from the assumed trajectory. We first consider the noise-free case with unconstrained\nbaseline and nonzero signed amplitudes.\nProposition 1(Observational equivalence on a finite time grid).Let f = (f (ti))m\ni=1 and fτ= (f (ti−τ))m\ni=1\n.\nThe two response families can generate the same noiseless vector with nonzero amplitudes if and only if\nfτ=cf+d1(3)\nfor an admissiblec̸= 0and offsetd. If amplitudes are restricted,c must also respect those restrictions.\nProof. Equality bτ1 +aτfτ= b01 +a0f yields c = a0/aτand d = (b0−bτ)/aτ. Conversely, substituting\nEq. (3) and settinga0 =aτc, b0 =bτ+aτd makes both vectors identical. Constant or zero-amplitude edge\ncases must be treated separately.\nFor example,f (t) = ekt has f (t−τ) = e−kτf (t), whereasf (t) = αt+βhas f (t−τ) = f (t)−ατ. For a\nfixed shift τ0, f (t) = ektq(t) with q(t−τ0) = q(t) also aliases. These are familiar algebraic counterexamples,\nnot new solution families. Such exact identities assume the shape extends over the shifted domain: a separate\nrule that clampsf (t) before stimulation may break that identity.\n1\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted November 4, 2026. ; https://doi.org/10.64898/2026.11.03.764027doi: bioRxiv preprint \n\n2.2 A measurable, but conditional, diagnostic\nSet X = [1, f ], letPX =XX + be the least-squares projector (withX + the Moore–Penrose pseudoinverse),\nand define\nδ(τ;T,f ) =∥(I−PX )fτ∥2, δ RMS =δ/√m.(4)\nProposition 2 (Signal–shape decomposition of a projection residual).Let g = fτbe the modeled shifted\ncurve and supposerange(f )> 0. DefineD =δRMS/range (f ), R = range(g)/ range(f ), and, for nonconstant\ng,\nS =∥(I−PX )g∥2\n∥g−g1∥2\n, D=R sd(g)\nrange(g)S≤R\n2,(5)\nwhere sd(g) uses divisorm. If g is constant,D = 0, R = 0, andS is undefined. For two varying curves, the\nunconstrained signed-amplitude projection givesS2 = 1−r(f, g)2, wherer is their Pearson correlation.\nProof. Because 1∈col(X), the projection residual ofg is no larger than the centered norm ofg. With its\nrange B, sd(g)≤B/2 by the range–variance inequality. Substitution yields Eq.(5). The correlation identity\nfollows from ordinary least-squares regression of one centered vector on the other.\nA smallD can therefore arise from weak retained signal (Rsmall) or similarity after affine reparameterization\n(S small). These are different limits on interpretation. Neither quantity is a confidence interval or a calibrated\nonset test. The two components do not alone determineD; its exact product also involvessd(g)/ range(g).\nWhich quantities remain identifiable? For the illustrative exponential referencef (t;k) = ekt, a delayed\nresponse y(t) = b +aek(t−τ)can be written asb +Aekt with theidentifiable combinationA =ae−kτ. No\nexpression-only observation separatesa from τin this unrestricted family. Yet, in noise-free observations at\nthree equally spaced timest,t +h,t + 2h, a nonzero response withk̸= 0obeys\nk =h−1 log\n(︃y(t + 2h)−y(t +h)\ny(t+h)−y(t)\n)︃\n.(6)\nThusk may remain identifiable even whenτdoes not. This is a model-specific illustrative reparameterization,\nnot a general estimator of a biological rate; near-flat curves or noisy observations can makek practically\nunrecoverable [9].\nIdentifiability-aware feature eligibility.If Eq. (3) holds on the sampled grid, the identity\nrange(fτ) =|c|range(f ) (7)\nis exact. Therefore requiringboth observed and modelled-shifted ranges to exceed a fixed threshold can exclude\nsome exactly aliased trajectories in a shift-dependent way. This conclusion concerns a candidate-dependent\nshifted-range gate; it does not establish that conventional filtering on observed expression alone invalidates\ntemporal analyses. The direction of the selection effect depends on the trajectory, shift and pre-onset boundary\nrule [10]. We report eligibility separately from projection residuals.\nThis is the residual after allowing the unshifted curve to change its baseline andsigned amplitude. Zero\nresidual denotes exact equivalence on the sampled grid; nonzero residual indicates geometric separation\nconditional on the assumed reference and boundary rule. It is neither a p-value nor a confidence interval and\ncannot determine an onset time by itself. With only two sampled times and unrestricted offset/amplitude,\neven distinct shifts can always have zero residual whenX spans the observation space. Sampling more times\ncan break some ambiguities but cannot resolve a globally exponential or affine identity by itself.\n3 Datasets and computational protocol\n3.1 Observed endometrial time course\nThe Gene Expression Omnibus (GEO) datasetGSE246591 contains RNA-sequencing (RNA-seq) profiles of\nfour primary human endometrial stromal-cell cultures, S169, S170, S506, and S508, each sampled at0, 3, 6,\n2\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted November 4, 2026. ; https://doi.org/10.64898/2026.11.03.764027doi: bioRxiv preprint \n\n0 1 2 3 4 5 6 7 8\nIllustrative time (arbitrary units)\n1.0\n1.5\n2.0\n2.5Reference response (arbitrary units)\nHypothetically shifted reference\nUnshifted reference, rescaled\nIllustrative sample times\nVerify\nGEO + 40 IDs\nNormalize\nlog2(1+CPM)\nReference\nLinear + pre-0\nCompare\nShift vs scale\nReport\nResidual + limits\nShifted values are modeled from the same ten measured times; no observed biological-lag truth.\nA Exact equivalence for an exponential reference\nB Reproducible GSE246591 analysis\nFigure 1: Exact ambiguity and analytical pipeline.(A) The two profiles overlap algebraically for\nan illustrative exponential reference; dots arenot measured RNA-seq. (B) The mandatory audit sequence\nchecks sample identities, fixes the reference and nuisance parameters, compares amplitude with displacement,\nassesses remaining geometric separation, and reports limits. The programmatic implementation records\nsource files and checksums.\n9, 12, 18, 24, 36, 48, and 96 hoursafter decidualization induction [11]. We verified all 40 culture-by-hour\ncolumns in the original deposited counts file (SHA-256 in Supplementary Methods). There are 60,676 source\nfeature rows. These are source identifiers, not necessarily 60,676 uniquely mapped genes. No independent\nwithin-culture, same-hour replicate series is available for estimating measurement variance or ground-truth\ndelay.\nFor each feature and culture, raw counts were divided by the total reads assigned to the sample, scaled\nto one million (counts per million; CPM), and transformed aslog2(1 + CPM). We considered hypothetical\ndisplacements τ∈{3, 6, 12, 24}hours. The measured ten-point curve was connected bypiecewise-linear\ninterpolation; for shifted arguments below time zero we imposedf (t < 0) = f (0). No expression at an\nunmeasured shifted time was treated as a new observation. For each candidate displacement we computed\nEq. (4) and dividedδRMS by the measured ten-point expression range. Features entered each descriptive\nsummary only if both their measured and modeled shifted expression ranges were at least 0.5 log2(1+CPM).\nConsequently, the set of included features changes with shift and culture; our summaries are not paired\ngene-level contrasts.\n3.2 Evidence roles and safeguards\nSample identities, input checksums, eligibility rules and model-specific statistics are recorded in the Supplement.\nWe present no estimate of power or classification accuracy for molecular onset. The threshold of normalized\n3\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted November 4, 2026. ; https://doi.org/10.64898/2026.11.03.764027doi: bioRxiv preprint \n\nTable1: Distinctanalyticalroles. Thereporteddatasetsare not pooled; samplesdonotrepresentexchangeable\nunits across studies.\nGEO series Design Role / boundary\nGSE246591 Four cultures;\nten sampling\ntimes each\nReference-shift geometry; no known biological lag or within-time replicate\nvariance.\nGSE94036 PGR/FOXO1\nperturbation\nin stromal\ncells\nSource of the PGR-oriented score, not a timed-lag experiment.\nGSE51981 Tissue\nmicroarrays;\nMSE subset\n28 vs 8\nFixed-score clinical contrast, not measured delay.\nGSE6364 Tissue\nmicroarrays;\nMSE subset 9\nvs 8\nExternal directional score check, not an hourly time course.\nprojection residual below 0.05 isdescriptive only: it is not derived from the experimental error distribution. We\nseparately describe a retrospective shape-sensitivity exercise on 48 features but do not call it a head-to-head\nbenchmark with DEtime or ImpulseDE2.\n4 Results\n4.1 Exact counterexamples: why timing alone may be unidentifiable\nAll three algebraic examples in Fig. 2 satisfy Eq.(3) for the specified shifts and allowed nuisance parameters.\nThus zero-noise expression values alone need not distinguish a temporal shift from rescaling or changing\nbackground expression. These mathematical witnesses do not establish that the corresponding shape is\nbiologically correct. A separate historical implementation verified the three identities to numerical tolerance;\nwe treat the identities themselves, rather than an inferred biological lag, as the result.\n4.2 Real measured trajectories show conditional, heterogeneous separation\nThe deposited GSE246591 counts matrix passed the 40/40 sample-identity and integer-count checks. Across\nthe four cultures and four shifts we obtained 16 descriptive summaries (Supplementary Table S1). Forτ= 3\nhours, medianδRMS/range(f ) was 0.220 for S169 and 0.153–0.165 for the remaining cultures. At 6 hours\nthe four medians were 0.206–0.214; at 12 hours they were 0.226–0.231; at 24 hours they were 0.209–0.219\n(Fig. 3). Thus a general assertion thatmost measured response curves are nearly exactly compatible with a\nshifted-and-rescaled response is not supported by this diagnostic.\nAt the descriptive cutoffδRMS/range(f )< 0.05, the fraction of included features ranged from approximately\n0.02% to 3.20% across the 16 culture-by-shift settings. This is neither an estimate of the proportion of\ngenuinely delayed genes nor a noise-calibrated test: the source includes only one profile per culture-hour\nand the hypothetical shifted values are computed from the same measured series. Larger shifts also change\nthe eligible feature universe; for example S169 included 14,019 rows at three hours but 9,116 at 24 hours.\nWe therefore do not interpret across-shift median differences as paired gene-level effects. The reported\nproportions also condition onboth dynamic-range gates. In particular, exact aliased profiles with shrinking\nmodeled response ranges need not remain eligible (Eq.(7)). A selection-aware audit instead reports the\ndenominator passing the observed-range gate, the additional exclusions caused by the candidate shift, and\n4\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted November 4, 2026. ; https://doi.org/10.64898/2026.11.03.764027doi: bioRxiv preprint \n\n0.0 2.5 5.0 7.5 10.0\nTime (illustrative)\n0.2\n0.4\n0.6\n0.8\n1.0Normalized responseshift = scale\nA   Exponential\nShifted\nAffine-reexpressed\n0.0 2.5 5.0 7.5 10.0\nTime (illustrative)\n0.2\n0.4\n0.6\n0.8\n1.0\nshift = baseline change\nB   Affine\n0.0 2.5 5.0 7.5 10.0\nTime (illustrative)\n0.2\n0.4\n0.6\n0.8\n1.0\nat shift = 6\nC   Exponential × periodic\nFigure 2: Exact analytic counterexamples, not measured data.An exponential, an affine function\nand an exponential multiplied by a periodic factor can each alias a specified shift through scale and/or offset.\nThe periodic construction establishes equality for its stated shift, not all possible shifts. The curves here are\nunclamped; the empirical analysis uses an explicit flat pre-zero boundary.\nthe distribution of the same geometric residual in retained and excluded features (Supplementary Methods).\nThis is a retrospective diagnostic, not a population estimate of lag identifiability.\n4.3 Separating signal loss from affine shape similarity\nThe two-component audit of the same GSE246591 matrix identifies distinct explanations for a smallD\n(Fig. 4). Among features passing the initial observed-curve range screen, 982 culture–feature records at\na three-hour candidate shift hadD < 0.05; 977 (99.5%) also hadS < 0.20, whereas none hadR <0.20.\nWithin each culture the near-zero group’s medianR was 0.968–0.975 and medianS was 0.131–0.141. Thus\nshort-shift small residuals commonly arose from affine shape similarity despite retained dynamic range, under\nthe specified reference model.\nAt a 24-hour shift, 730 culture–feature records were both excluded by the shifted-range screen and below\nD <0.05. Of these, 557 (76.3%) hadR< 0.20, while only 20 (2.7%) hadS <0.20; one shifted curve was\nexactly flat and had undefinedS. MedianR among these low-D excluded features was 0.136–0.186 across\ncultures, whereas median definedS was 0.866–0.989. Their small residuals more often reflected loss of the\nshifted signal than similarity of two varying shapes. The 122 near-zero records that remained shift-eligible at\n24 hours showed comparatively higher medianR (0.578–0.616) and lowerS (0.225–0.236). The illustrative\n0.20cutoffs do not define statistical significance.\nThe distinction also explains why the secondary dynamic-range gate changes the denominator: of 56,988\ninitially dynamic culture–feature records at 24 hours, 19,676 (34.5%) were excluded by the shifted-range\ncondition. Among all initially eligible records, 852 (1.50%) hadD< 0.05, compared with 122/37,312 (0.33%)\nof the retained set. We donot interpret this difference as proof that filtering preferentially removes structurally\naliased trajectories: Eq.(5) makes smallD mechanically possible whenR is small. A fixed intersection of\neligible features across all four candidate shifts (7,858–9,960 rows per culture) provides a stable denominator\nfor supplementary descriptive comparisons. Additional culture- and shift-specific summaries are reported in\nthe Supplement.\n4.4 Sensitivity to the assumed reference-curve shape\nA separate, retrospectively selectedfixed 48-feature subset was evaluated using piecewise-linear interpolation,\na Gaussian-process (GP) approximation with fixed parameters, and a two-sigmoid impulse-like approximation.\nIn leave-one-time-point-out reconstruction, linear interpolation had the lowest median feature-level root-mean-\nsquare error (RMSE) in each culture (0.595–0.737 log2CPM). The GP and impulse-like curves often produced\nsmaller conditional residuals for a three-hour hypothetical shift, illustrating the sensitivity of Eq.(4) to the\n5\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted November 4, 2026. ; https://doi.org/10.64898/2026.11.03.764027doi: bioRxiv preprint \n\n3 6 12 24\nHypothetical displacement τ (hours)\n0.12\n0.14\n0.16\n0.18\n0.20\n0.22\n0.24\n0.26Median projected residual / observed range\nS169 S170 S506 S508\nFigure 3: Measured-source, model-conditional GSE246591 geometry.Medians over included feature\nrows of the root-mean-square residual after projecting a hypothetical displaced curve onto baseline and\nsigned amplitude, divided by each feature’s measured expression range. The values are computed from actual\nGEO counts, while displaced coordinates rely on piecewise-linear interpolation and the flat pre-zero rule.\nFeature eligibility differs byτ(Supplementary Fig. S3); connecting lines is for readability, not evidence of a\ncontinuously estimated lag.\nreference representation. The pilot emitted numerical warnings from matrix operations despite finite exported\ntables. We therefore classify this exercise asexploratory and not numerical method validation; its complete\ndescriptive readouts and caveats are in Supplementary Table S3. These approximations are not the published\nimplementations of DEtime or ImpulseDE2; they serve only to test sensitivity to curve representation.\n4.5 Independent biological context does not establish temporal delay\nThe PGR (progesterone receptor)-dependency score was derived from a separate stromal-cell perturbation\nexperiment (GSE94036) and evaluated in endometrial cohorts [2, 3, 12–14]. In the 28 endometriosis versus\neight strict-control mid-secretory endometrium (MSE) samples from GSE51981, the estimated score difference\nwas−0.154922. The phase-interaction modelusing all eligible menstrual phases (n = 109) yielded a 95%\nconfidence interval (CI) based on HC3 heteroskedasticity-robust standard errors[−0.222227,−0.087617] for\nthe MSE contrast andp = 1.386×10−5. A separate mid-secretory-only model with HC3 standard errors\nproduced [−0.223889,−0.085955] (p = 6.247×10−5); adjusting that MSE-only model for three composition-\nproxy components gave−0.076561 (95% CI [−0.133443,−0.019678], p = 0.009975). These are different\nmodels and must not be assigned one another’s p-values.\nIn the independent GSE6364 mid-secretory subset (nine endometriosis cases and eight controls), the\nfrozen-score difference was−0.032765. Its unequal-varianceWelch 95% CI was[−0.067656, +0.002126], with\np = 0.063752. Its direction was consistent but its interval included zero. The GSE51981 values from the phase\nmodel and the GSE6364 Welch estimate explain the rounding in the registered abstract. Neither endometrial\ncomparison repeatedly samples thesame women over hours after a controlled perturbation; no delayed\nmolecular onset can be inferred from them. We do not pool effect sizes across these distinct estimands.\n6\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted November 4, 2026. ; https://doi.org/10.64898/2026.11.03.764027doi: bioRxiv preprint \n\n0.0 0.2 0.4 0.6 0.8 1.0\nDynamic-range retention, R\n0.0\n0.2\n0.4\n0.6\n0.8\n1.0Affine shape separation, S\nFour cultures per stratum\nA  Near-zero residual strata (culture medians)\n3 h: initially dynamic\n24 h: shift-excluded\n24 h: shift-retained\n3 h: initial dynamic\n(n=982)\n24 h: shift-excluded\n(n=730)\n0\n20\n40\n60\n80\n100Within near-zero residuals (%)0.0%\n76.3%\n99.5%\n2.7%\nB  Descriptive component thresholds\nR < 0.20\nS < 0.20\nGSE246591 · Median points and cutoff counts from author-verified summary only; not per-gene scatter or calibrated rates.\nFigure 4: Two explanations for a small geometric projection residual. (A) Each point is a\nculture-level median among near-zero(D <0.05) records, not a per-gene joint scatter; the three symbols\ndistinguish the three-hour initially dynamic stratum from 24-hour shifted-range-excluded and retained strata.\n(B) Descriptive percentages belowR = 0.20 and S = 0.20 within the corresponding near-zero groups. The\ngroups differ in size and selection rule; thresholds are illustrative, not calibrated tests. Source: GSE246591\nand the archived author-run dual-component audit.\n5 Discussion\nThe central distinction is betweenstructural ambiguity(different mechanisms giving exactly the same noiseless\nobservations) andempirical information (the separation suggested by an observed curve under specified\nassumptions). Algebra shows that some shape families can alias delay exactly. With an empirical curve,\nprojection residuals quantify the distance between hypothetical shifted and unshifted profiles only after\nreference interpolation, boundary behavior, background and amplitude freedom have been specified. Even\na nonzero residual cannot imply practical distinguishability without replicate-level noise and appropriate\ndependence modeling [\n7, 8]. The actual GSE246591 grid showed heterogeneous, mostly non-negligible\nnormalized distances among the eligible features rather than pervasive near-zero residuals. That descriptive\nfinding refines the motivating question; it does not recover the onset of a biological event.\nA constructive positive implication is that when a shift is absorbed by an amplitude–baseline transformation,\na parameter describingshape may remain measurable. For an exponential family, the ratek is an example\n(Eq. (6)); time to a specified observed plateau is not automatically an identifiable substitute and would\nrequire its own model-specific analysis. The appropriate estimand therefore depends on which parameters\nsurvive the nuisance reparameterization.\nThe signal–shape decomposition additionally shows why a low range-normalized residual is not, by itself,\nevidence of affine equivalence: reduced shifted dynamic range can lower the residual without reducing the\nrelative shape-separation ratio. We recommend reporting signal retention, affine shape separation, non-\nevaluable flat profiles and eligibility counts together, rather than using a single near-zero residual fraction.\nThis is an interpretive auditing protocol built from standard regression geometry, not a novel statistical test\nor a proven method for recovering onset.\nThe main limits are a single measured expression vector per culture-hour, no known imposed lag, no\nmatched untreated curve at every time, uncertain interpolation and pre-zero behavior, source feature identifiers\nnot mapped to unique genes, and retrospective eligibility/filter choices. Thousands of feature rows should not\nbe counted as thousands of independent biological replicates. In addition, the amplitude projection is signed;\na constrained nonnegative-amplitude model would yield different distances when fitted slopes are negative.\nThe Supplement distinguishes that issue and the separate fixed-panel historical analyses. The curve-shape\npilot is not a validated comparison with published state-of-the-art methods and emitted numerical warnings;\nits values must not be used to rank algorithms.\nRelated preprints by the present authors discuss assayability and cohort transfer [15], and the distinction\nbetween computational stability and biological interpretation [16]. They motivate separate reporting of\n7\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted November 4, 2026. ; https://doi.org/10.64898/2026.11.03.764027doi: bioRxiv preprint \n\nmeasurement compatibility, statistical reproducibility and biological meaning, but do not validate this\nendometrial timing hypothesis. A decisive follow-up would preregister amplitude and curve-shape constraints,\ncollect matched untreated time courses from independent donors with within-time biological replicates, impose\na known perturbation delay, and assess candidate estimators on the same prediction target. Such a study\nwould also permit a fair comparison with DEtime, ImpulseDE2 and other established time-course approaches.\nWhere those conditions are absent, the correct result is a limit on interpretability, not a claimed delayed\nresponse.\n6 Conclusions and software availability\nDelay and attenuation can be observationally equivalent even with perfect measurements, but how closely they\nagree on a real sampling grid is reference- and assumption-dependent. Our endometrial RNA-seq audit provides\nreproducible descriptive evidence for that distinction without claiming biological-onset recovery or a new\nidentification theorem. Code and reproducibility instructions: https://github.com/Helen2813/paper7.\nThe separate OnsetAudit application, its genuine screenshot and its mouse-cell elapsed-exposure-time\nbenchmark are documented in the Supplement; they are not evidence for human endometrial lag.\n8\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted November 4, 2026. ; https://doi.org/10.64898/2026.11.03.764027doi: bioRxiv preprint \n\nReferences\n[1] Birgit Gellersen and Jan J. Brosens. Cyclic decidualization of the human endometrium in reproductive\nhealth and failure.Endocrine Reviews, 35(6):851–905, 2014. doi: 10.1210/er.2014-1045.\n[2] Richard O. Burney, Said Talbi, Amy E. Hamilton, Kim Chi Vo, Mette Nyegaard, Camran R. Nezhat,\nBruce A. Lessey, and Linda C. Giudice. Gene expression analysis of endometrium reveals progesterone\nresistance and candidate susceptibility genes in women with endometriosis.Endocrinology, 148(8):\n3814–3826, 2007. doi: 10.1210/en.2006-1692.\n[3] John S. Tamaresis, Juan C. Irwin, Gabriel A. Goldfien, Joseph T. Rabban, Richard O. Burney, Camran\nNezhat, Louis V. DePaolo, and Linda C. Giudice. Molecular classification of endometriosis and disease\nstage using high-dimensional genomic data.Endocrinology, 155(12):4986–4999, 2014. doi: 10.1210/en.\n2014-1490.\n[4] Almudena Devesa-Peiro, Patricia Sebastian-Leon, Antonio Pellicer, and Patricia Diaz-Gimeno. Guidelines\nfor biomarker discovery in endometrium: Correcting for menstrual cycle bias reveals new genes associated\nwith uterine disorders. Molecular Human Reproduction, 27(4):gaab011, 2021. doi: 10.1093/molehr/\ngaab011.\n[5] JingYang, ChristopherA.Penfold, MurrayR.Grant, andMagnusRattray. Inferringtheperturbationtime\nfrom biological time course data.Bioinformatics, 32(19):2956–2964, 2016. doi: 10.1093/bioinformatics/\nbtw329.\n[6] David S. Fischer, Fabian J. Theis, and Nir Yosef. Impulse model-based differential expression analysis of\ntime course sequencing data.Nucleic Acids Research, 46(20):e119, 2018. doi: 10.1093/nar/gky675.\n[7] Andreas Raue, Clemens Kreutz, Tim Maiwald, Julie Bachmann, Marcel Schilling, Ursula Klingmueller,\nand Jens Timmer. Structural and practical identifiability analysis of partially observed dynamical models\nby exploiting the profile likelihood.Bioinformatics, 25(15):1923–1929, 2009. doi: 10.1093/bioinformatics/\nbtp358.\n[8] Franz-Georg Wieland, Adrian L. Hauber, Marcus Rosenblatt, Christian Tönsing, and Jens Timmer.\nOn structural and practical identifiability.Current Opinion in Systems Biology, 25:60–69, 2021. doi:\n10.1016/j.coisb.2021.03.005.\n[9] Nicolette Meshkat, Marisa Eisenberg, and Joseph J. DiStefano. An algorithm for finding globally\nidentifiable parameter combinations of nonlinear ODE models using Gr\"obner bases.Mathematical\nBiosciences, 222(2):61–72, 2009. doi: 10.1016/j.mbs.2009.08.010.\n[10] Richard Bourgon, Robert Gentleman, and Wolfgang Huber. Independent filtering increases detection\npower for high-throughput experiments. Proceedings of the National Academy of Sciences, 107(21):\n9546–9551, 2010. doi: 10.1073/pnas.0914005107.\n[11] NCBI Gene Expression Omnibus. GSE246591: Transcriptomic exploration of decidualizing human\nendometrial stromal cell dynamics. Gene Expression Omnibus dataset, 2024. URLhttps://www.ncbi.\nnlm.nih.gov/geo/query/acc.cgi?acc=GSE246591.\n[12] NCBI Gene Expression Omnibus. GSE94036: Progesterone receptor- and FOXO1-dependent transcrip-\ntomes of decidualized human endometrial stromal cells. Gene Expression Omnibus dataset, 2017. URL\nhttps://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE94036.\n[13] NCBI Gene Expression Omnibus. GSE51981: Molecular classification of endometriosis and disease stage.\nGene Expression Omnibus dataset, 2014. URL\nhttps://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?\nacc=GSE51981.\n[14] NCBI Gene Expression Omnibus. GSE6364: Gene profiling of endometrium reveals progesterone\nresistance in endometriosis. Gene Expression Omnibus dataset, 2007. URL\nhttps://www.ncbi.nlm.\nnih.gov/geo/query/acc.cgi?acc=GSE6364.\n[15] Elena Krikun and Abedalrhman Alkhateeb. External evaluation of multi-omics prognostic models across\nTCGA-BRCA and METABRIC: Transportability, stability, and incremental performance. bioRxiv, 2026.\nURL https://doi.org/10.64898/2026.09.18.752709. Preprint; not peer reviewed.\n[16] Elena Krikun and Abedalrhman Alkhateeb. Latent-factor stability is not sufficient: Evaluating SABLE-\nMO in simulated and cancer multi-omics data. bioRxiv, 2026. URLhttps://doi.org/10.64898/2026.\n10.12.758631. Preprint; not peer reviewed.\n9\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted November 4, 2026. ; https://doi.org/10.64898/2026.11.03.764027doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}