When a developing brain is damaged, when the damage happens matters as much as how much. Identical chemical or genetic insults produce severe, irreversible outcomes if they land inside a narrow developmental window and almost nothing if they land outside it — the principle biologists call a window of susceptibility. Two measurement traditions have mapped this timing problem from opposite ends without ever meeting. Molecular biologists track how gene expression unfolds over developmental time and use it to predict when disorders like autism become likely (Identifying Windows of Susceptibility by Temporal Gene Analysis). Electrophysiologists measure how the brain's rhythmic electrical activity — decomposed into frequency bands from slow delta waves to fast gamma oscillations — matures over childhood, charting for example how coupling between slow waves and high-frequency bursts grows in visual cortex with a slope of +0.047/√year (Developmental atlas of phase-amplitude coupling between physiologic high-frequency oscillations and slow waves). Nobody has connected the two: we do not know whether the gene-expression programs that open a susceptibility window are the same programs that shape which brain rhythms mature when, and no study in the literature couples spectral frequency decomposition with dynamical transcriptomic models across developmental windows.
This proposal builds that coupling as a predictive, mechanistic model, and then tests it with a targeted experiment. The core move is to take the machinery of FLeCS — a model that learns ordinary differential equations over gene regulatory networks to simulate how single-cell gene expression evolves over time (A scalable gene network model of regulatory dynamics in single cells) — and attach a spectral readout, so that the same regulatory dynamics that push a cell through maturation also drive predicted changes in frequency-band power and cross-frequency coupling. The model is trained on two existing public resources: the CORTECON dataset of gene expression across nine stages of human cortical differentiation in a dish, and the intracranial EEG atlas of phase-amplitude coupling across ages 1.0 to 41.5 years. Its decisive test is a perturbation experiment: knock down autism-risk genes in silico at different developmental times and ask whether the model predicts band-specific rhythm deficits — and whether those predictions match what actually happens in mice exposed to the drug valproate at different embryonic days, a manipulation known to hit interneurons hardest at one specific day (Prenatal Valproic Acid Exposure Impairs Offspring Cognition Through Disturbing Interneuron Development).
Core Claim
The claim is that the developmental trajectory of cross-frequency electrophysiology is not an independent maturation curve but a readout of gene regulatory dynamics, and that this coupling is testable in three escalating ways. First, a gene-dynamical model fit only on longitudinal expression data should predict region- and age-specific spectral trajectories better than a purely chronological model does. Second, perturbing the gene dynamics at a transcriptomically-defined susceptibility window should produce band-specific spectral deficits — gamma and high-frequency effects from interneuron-related gene modules, not uniform broadband changes. Third, those in silico predictions should match the spectral phenotype of real animals insulted during the corresponding window, which no existing study has measured.
Three gaps in the literature make this claim open rather than incremental. The windows-of-susceptibility clock (Identifying Windows of Susceptibility by Temporal Gene Analysis) operates purely on RNA counts and can say nothing about rhythms. The developmental coupling atlas (Developmental atlas of phase-amplitude coupling between physiologic high-frequency oscillations and slow waves) charts the spectral side with quantitative precision — occipital coupling rising at +0.047/√year, frontal high-frequency oscillation rates falling at −0.41/√year, consistent with synaptic pruning — but records no transcriptomic data and so cannot say what drives those curves. And the valproate study (Prenatal Valproic Acid Exposure Impairs Offspring Cognition Through Disturbing Interneuron Development) profiled gene expression and behavior but never recorded electrophysiology, so whether the uniquely vulnerable E12.5 exposure leaves a specific spectral fingerprint is completely untested. Our proposal is designed to close all three at once, using each existing result as a constraint rather than a competitor.
Research Program
The model couples two dynamical layers. The gene layer reuses FLeCS's formulation: gene expression states $z$ evolve under learned production and decay functions computed by message passing over a gene regulatory network with adjacency $A$,
$$\dot{z} = P_\theta(z, A) \;-\; D_\theta(z)\odot z, \qquad \hat{y}(t) = h_\phi\big(z(t)\big),$$
where $P_\theta$ and $D_\theta$ are the learned production and decay rates and $\odot$ is elementwise multiplication. The new ingredient is the readout $h_\phi$, which maps the regulatory state at developmental time $t$ onto spectral features $\hat{y}$: log power in the canonical delta-through-gamma bands, high-frequency oscillation rates, and the phase-amplitude coupling Modulation Index as computed by the winPACT pipeline used in the atlas study. The readout is deliberately structured so that biologically annotated gene groups — ion-channel genes, synaptic scaffolds, interneuron-lineage markers — feed band-specific outputs rather than a single shared trunk, making the band specificity of any perturbation a property of the learned wiring rather than a post hoc interpretation. We acknowledge one known limitation we inherit: FLeCS's production rates combine regulators linearly and are deterministic, which oversimplifies combinatorial logic and transcriptional bursting; we treat this as a first-generation constraint and include an ablation with a nonlinear production head.
The hardest sub-problem is temporal alignment, and we treat it as a first-class research question rather than a preprocessing step. CORTECON gives nine stages of in vitro cortical differentiation (days in culture); the EEG atlas gives chronological years in human patients; the mouse valproate model gives embryonic days E10.5–E14.5. These three clocks must be registered to each other. Our approach is to anchor the registration on shared molecular events with known timing on both axes — for example the expression peaks of interneuron-lineage and synaptogenesis markers — and to report the registration's uncertainty explicitly. A skeptical reader should expect this alignment to be the proposal's largest source of error, and the experimental program below is structured so that the headline claims do not collapse if the alignment is only approximately right: band specificity and window specificity of perturbation effects are robust to monotone time-warping, even if exact ages are not.
The perturbation screen is the scientific payload. The susceptibility-window study found autism-associated genes enriched at the Neural Differentiation, Cortical Specification, and Upper Layers stages of CORTECON. We will re-run the FLeCS-style intervention protocol it implies — knock down each enriched gene module at each of the nine stages — but now read out the predicted spectral consequence, producing a gene-by-window-by-band map of susceptibility. The central mechanistic prediction is asymmetric: perturbations hitting interneuron-related modules during the E12.5-equivalent window should selectively depress gamma-band power and high-frequency oscillation rates (fast-spiking inhibitory networks are the recognized generators of gamma rhythms, and their maturation is precisely what the valproate study showed to be disrupted), while earlier neuroectoderm-stage perturbations should shift broadband trajectories without gamma specificity.
Baselines
The "do nothing" reference point is the atlas's own statistical model: a mixed-effects regression of Modulation Index and high-frequency oscillation rate on chronological age (plus the square root of age), fit per brain region. This is the correct null because it already encodes everything chronological regression can know; the atlas paper reports its occipital coupling slope as +0.047/√year with p < 0.001 and its frontal high-frequency oscillation rate slope as −0.41/√year (Developmental atlas of phase-amplitude coupling between physiologic high-frequency oscillations and slow waves). Beating it means predicting held-out subjects' spectral trajectories with higher out-of-sample R² and correctly predicting trajectory changes under perturbation, which an age-only model cannot do even in principle. No reported result found on Lacuna for age-regression or trajectory-prediction accuracy on this dataset, so we will establish the age-only number ourselves under identical folds.
The strongest existing competitor is a static gene-to-spectrum regression: take the CORTECON eigengenes from the susceptibility-window pipeline (Identifying Windows of Susceptibility by Temporal Gene Analysis) at each stage, regress them onto the atlas spectral features at the registered age, with no differential equation in between. This is the fair version of "just correlate the two datasets" and it isolates exactly what the ODE adds. We will also run a spectral-only deep baseline that never sees gene data: a graph neural network over electrode contacts built with the GNN4EEG toolkit (GNN4EEG: A Benchmark and Toolkit for Electroencephalography Classification with Graph Neural Network), trained to predict age and regional coupling trajectories from band features — the toolkit's own benchmark shows cross-subject generalization is the hard part there (best cross-subject model, RGNN, reaches 31.3% accuracy on the 9-class FACED task, versus 94.5% for HetEmotionNet intra-subject), which calibrates how much cross-subject variance any EEG-only model can absorb. All baselines get the same data splits, the same held-out subjects and regions, and the same hyperparameter tuning budget under nested cross-validation, following the GNN4EEG protocol; compute is matched by fixing total parameter count and training FLOPs across contenders.
Two ablations isolate the proposal's key ingredients. Replacing the ODE layer with the static regression above isolates the value of dynamics; if the static map matches the coupled model on trajectory prediction and perturbation direction, the differential equation is decoration and the claim reduces to correlation. Replacing the band-decomposed readout with a single broadband output isolates the value of frequency structure; prior work makes a clear prediction here — motor-cortex signals show simultaneous low-frequency decreases and high-frequency increases that cancel in broadband (Spectral Changes in Cortical Surface Potentials during Motor Movement), and Brant's own ablation showed that removing its explicit frequency-band encoding caused immediate drops across seizure detection, forecasting, and imputation tasks (Brant: Foundation Model for Intracranial Neural Signal) — so the broadband readout should specifically lose the gamma-versus-slow-wave dissociation that our claim depends on.
Why not just…?
Why not just train a deep model on raw broadband EEG and skip band decomposition? The evidence against this is consistent across the corpus: opposing directional shifts in low and high frequencies cancel in broadband (Spectral Changes in Cortical Surface Potentials during Motor Movement), sub-band connectivity captures diagnostic signal that broadband averages destroy (Resting-State Multi-Spectrum Functional Connectivity Networks for Identification of MCI Patients), different bands carry non-redundant information about response strength versus timing (Different LFP frequency bands convey complementary information about the BOLD signal), and both Brant and CodeBrain — which reports a +9.11% Cohen's Kappa gain over its strongest baseline on BCIC 2020-T3 — needed explicit frequency tokenization to work at all (CodeBrain: Bridging Decoupled Tokenizer and Multi-Scale Architecture for EEG Foundation Model). More decisively for us: our claim is band-specific perturbation effects, which a broadband model cannot even express. The broadband ablation settles this empirically.
Why not just do endpoint RNA-seq after the insult, as the valproate study did? Endpoint snapshots confound direct damage with secondary compensation. FLeCS aligned with known biology in 11 out of 12 K562 knockdowns where standard differential expression analysis did not, precisely because the dynamical model separates direct regulatory effects from downstream ripples (A scalable gene network model of regulatory dynamics in single cells); the susceptibility-window study needed all nine CORTECON time points to localize when disease genes act (Identifying Windows of Susceptibility by Temporal Gene Analysis); and maturation status was only interpretable by projecting cells onto a longitudinal axis spanning stem cell to adult tissue (Transcriptome-Guided Functional Analyses Reveal Novel Biological Properties and Regulatory Hierarchy of Human Embryonic Stem Cell-Derived Ventricular Cardiomyocytes Crucial for Maturation). The static-regression ablation is the honest test of whether our dynamics earn their cost on top of these arguments.
Why not just run the mouse experiment and skip the model? This is a real threat and we do not argue around it: the mouse experiment is the ground truth, and we include it as the decisive check rather than claiming the model substitutes for it. But the experiment alone cannot do three things the model can. It cannot screen hundreds of genes and windows — only a handful of exposure days are ethically and practically testable, while the in silico screen covers the full gene-by-window-by-band grid. It cannot bridge to human data, since mouse embryonic days never directly meet human years without a registered molecular clock. And it cannot generate counterfactuals. The division of labor is explicit: the model chooses which window and which band to measure, and the experiment adjudicates.
Why not just correlate the SWOT-clock windows with the atlas milestones post hoc? A raw correlation across these datasets would be confounded at every joint: the transcriptomic data come from cells in a dish at nine coarse stages, the EEG data come from epilepsy surgery patients (the atlas authors used nonepileptic electrode sites, but the cohort is still not healthy), and the two time axes are incommensurable. Correlation would also be unfalsifiable in the perturbation sense — it can note that an autism-enriched stage and a coupling-growth period overlap after registration, but it cannot predict the sign or band of an intervention. The coupled model converts the observation into a testable prediction, which is exactly what the ablation against the static regression is designed to demonstrate.
Proposed Experiments
Experiment 1 tests trajectory prediction. Train the coupled model on CORTECON expression dynamics and atlas spectral features (8,251 nonepileptic sites, 114 patients, ages 1.0–41.5, via OpenNeuro ds004551 and winPACT), with regions and patients held out. Metric: out-of-sample R² and RMSE for per-site Modulation Index and high-frequency oscillation rate as functions of age, per lobe, against the age-only mixed-effects baseline and the static gene-to-spectrum regression. Success means beating the age-only model's R² especially in early childhood, where the atlas shows the steepest occipital coupling growth (+0.047/√year, p < 0.001) and where a gene-driven mechanism should concentrate its explanatory power (Developmental atlas of phase-amplitude coupling between physiologic high-frequency oscillations and slow waves).
Experiment 2 is the in silico perturbation screen. Knock down each autism-risk gene module enriched at the Neural Differentiation, Cortical Specification, and Upper Layers stages (Identifying Windows of Susceptibility by Temporal Gene Analysis) at each of the nine CORTECON stages, and record the predicted change in each spectral band. The preregistered prediction is an interaction: interneuron-module perturbations at the E12.5-equivalent stage depress gamma power and high-frequency oscillation rates while leaving slow-wave power comparatively intact, whereas neuroectoderm-stage perturbations shift trajectories without band specificity. The decisive statistic is the gene-module × window × band interaction term, tested against the static-regression and broadband ablations, both of which should fail to reproduce the interaction if the mechanism claim is right.
Experiment 3 is the wet-lab adjudication and directly fills the untested gap in the valproate study. Pregnant mice receive valproate at E10.5, E12.5, or E14.5 exactly as in the original protocol (Prenatal Valproic Acid Exposure Impairs Offspring Cognition Through Disturbing Interneuron Development); offspring get chronic cortical electrocorticography around P28 with spectral decomposition and winPACT-style coupling analysis, plus a short longitudinal RNA-seq series from embryonic through juvenile stages so the mouse arm carries its own expression clock. The model is scored on whether it predicts, ahead of measurement, (a) which exposure day yields a spectral phenotype, (b) which band is affected, and (c) the direction of the effect. The developmental-timing literature predicts E12.5 uniquely matters for the interneuron route, and the timing-behavior work in autism models shows early- versus late-onset circuit alterations cascade into qualitatively different adult outcomes (The developmental timing of spinal touch processing alterations predicts behavioral changes in genetic mouse models of autism spectrum disorders), so a spectral fingerprint restricted to the E12.5 cohort would corroborate the chronogeneity logic on a new measurement axis.
The decisive checks, in order of importance, are these. If the static gene-to-spectrum regression matches the coupled ODE model on Experiment 2's interaction term, the mechanistic claim fails and the work reduces to a correlation study. If the mouse experiment shows a spectral phenotype at all three exposure days, or at none, the window-specificity claim fails regardless of model fit. And if the broadband readout ablation reproduces the band-specific interactions, our reliance on frequency decomposition was unnecessary — though the converging negative evidence from Spectral Changes in Cortical Surface Potentials during Motor Movement and Brant's frequency-encoding ablation makes this the least likely failure mode. Each check is designed so that its failure is informative rather than fatal: a failed ODE layer still leaves the first gene-to-spectrum bridge and the first spectral phenotyping of the valproate window design, both of which the literature currently lacks entirely.
References
- Identifying Windows of Susceptibility by Temporal Gene Analysis
- Developmental atlas of phase-amplitude coupling between physiologic high-frequency oscillations and slow waves
- A scalable gene network model of regulatory dynamics in single cells
- Prenatal Valproic Acid Exposure Impairs Offspring Cognition Through Disturbing Interneuron Development
- GNN4EEG: A Benchmark and Toolkit for Electroencephalography Classification with Graph Neural Network
- Spectral Changes in Cortical Surface Potentials during Motor Movement
- Brant: Foundation Model for Intracranial Neural Signal
- Resting-State Multi-Spectrum Functional Connectivity Networks for Identification of MCI Patients
- Different LFP frequency bands convey complementary information about the BOLD signal
- CodeBrain: Bridging Decoupled Tokenizer and Multi-Scale Architecture for EEG Foundation Model
- Transcriptome-Guided Functional Analyses Reveal Novel Biological Properties and Regulatory Hierarchy of Human Embryonic Stem Cell-Derived Ventricular Cardiomyocytes Crucial for Maturation
- The developmental timing of spinal touch processing alterations predicts behavioral changes in genetic mouse models of autism spectrum disorders