Atlas-based electrical source imaging for mouse EEG: Application to an Fmr1 knockout model

Abstract

Electrical source imaging (ESI) reconstructs where in the brain a scalp EEG signal arises. It is routine in humans but unestablished in the mouse, whose brain is some 3000 times smaller where it is not obvious that a sparse array can separate one region from another. We built an atlas-based ESI pipeline for 30-channel mouse EEG and tested it against simulated dipoles, across 18 combinations of source space, head model and inverse operator, with conductivity and noise sweeps. The best, a region-of-interest source space with an ellipsoid head model and sLORETA, recovered superficial sources to a mean error of 0.93 mm, rising to 5.08 mm at depth. We then applied it to resting EEG from male Fmr1 knockout (n = 18) and wild-type (n = 17) mice. Gamma power was elevated in the knockouts (Low Gamma q = 0.004, standardized mean difference 1.02; High Gamma q = 0.004, 0.67) and localized to six parcels, three of them in both bands. Relative power in Delta, Theta and High Gamma, and both aperiodic parameters, varied by region (q < 0.001 to 0.035). The electrode analysis recovered neither the Delta pattern nor any named structure. The gamma findings survived adjustment for acquisition cohort (p = 0.001 and 0.002). Simulation bounds what the pipeline recovers and which region it assigns activity to, but cannot show that a labeled region is the true source in a live animal. Replication across recording systems, sex, age and injection-naive animals is needed.

1. Introduction

Electrical source imaging (ESI) estimates the brain-region origins of scalp-recorded EEG by combining a head-conduction forward model, typically boundary element method (BEM) head geometry, with an inverse solution (Grech et al., 2008; Michel & Brunet, 2019). In human neuroimaging, ESI is a mature methodology that converts sensor-space recordings into anatomically resolved maps of neural activity. Mouse EEG has become a central translational platform for neurodevelopmental disorders, including Fragile X syndrome (FXS) (Jonak et al., 2024; Pedapati et al., 2022), where both patients and Fmr1 knockout (KO) mice exhibit elevated resting gamma power linked to excitatory–inhibitory imbalance (Ethridge et al., 2017; Lovelace et al., 2018; Wang et al., 2017). Yet mouse EEG is almost universally analyzed at the electrode level, leaving the spatial origin of its signals unresolved. Extending ESI to the mouse is therefore an open methodological problem rather than a routine application.

Volume conduction limits scalp EEG: activity in one region spreads through tissue and is detected at multiple electrodes (Nunez & Srinivasan, 2006), inflating inter-electrode correlations (Haufe et al., 2013; Schoffelen & Gross, 2009) and preventing attribution of spectral power to specific regions. In human EEG, ESI addresses this by estimating activity at brain regions of interest (ROIs) rather than at electrodes.

The central obstacle is scale. The mouse brain is orders of magnitude smaller than the human brain (Herculano-Houzel, 2009), so it is not obvious that a sparse surface array can resolve adjacent regions, or that any inverse solution recovers true source geometry rather than a smoothed artifact. Prior rodent work demonstrated proof-of-concept with subject-specific models (Bae et al., 2015; Jiricek et al., 2021), but did not quantify localization accuracy across the configuration space, nor provide a standardized, atlas-based framework applicable across laboratories. Whether mouse ESI is well-posed has two parts that are easily conflated: whether the inverse problem is numerically tractable at this scale, and whether recovered activity can be attributed to anatomy in tissue, which simulation cannot establish. This study characterizes the first and bounds the second.

We address this gap by developing and benchmarking an atlas-based ESI pipeline for 30-channel mouse EEG. The pipeline pairs the Antwerp Mouse Brain Atlas (Praet et al., 2018) (BEM geometry, electrode registration) with an Allen-32 parcellation derived from the Allen Common Coordinate Framework v3 (CCFv3) (Wang et al., 2020) (32 ROIs, 16 bilaterally symmetrical pairs). We systematically evaluate 18 configurations (2 BEM × 3 source spaces × 3 inverse methods) through forward-inverse dipole simulations with conductivity and SNR robustness testing, then apply the selected configuration to male Fmr1 KO (n = 18) and WT (n = 17) mice (76–85 days at recording, all with adolescent ICV vehicle exposure from a parent gene therapy study), asking whether source-level band power and aperiodic parameters differ by genotype, and whether spatially heterogeneous effects emerge that the electrode tier cannot resolve, together with what such a pipeline should report (§4.9). KO vs. WT contrast, six-band absolute and relative power, and aperiodic exponent/offset were prespecified; methodological refinements were fixed before examining biological results.

2. Methods

2.1. Pipeline overview

We developed an electrical source imaging (ESI) pipeline implemented in Python using MNE-Python 1.11.0 (Gramfort et al., 2013) with custom extensions for mouse-scale analysis. The pipeline proceeds from electrode registration through BEM head modeling, source space creation, forward and inverse solutions, ROI extraction, and spectral analysis. All steps are configured through YAML preset files, enabling reproducible and systematic comparison of configurations. Fig. 1 shows the sequence and the file each step writes.

Fig. 1.

The source-imaging pipeline. A schematic of the six steps applied to each recording, in the order given in §2.1, with the file each step writes: electrode registration, BEM head modeling, source-space creation, the forward solution, the sLORETA inverse and ROI extraction. One command runs the whole pipeline; every step is set by one YAML preset.

2.2. Anatomical atlas

The Antwerp Mouse Brain Atlas (Praet et al., 2018) provides the coordinate space, brain mask, and BEM mesh-fitting volume; all ROI assignments use a parcellation derived from the Allen CCFv3 (Wang et al., 2020), registered to the Antwerp coordinate space via Advanced Normalization Tools (ANTs) nonlinear (SyN) registration. We refer to this parcellation as Allen-32: 32 ROIs (16 bilaterally symmetrical pairs) organized into 10 functional regions (Frontal-Anterior, Motor, Somatosensory, Auditory, Visual/Parietal, Lateral Cortex, Retrosplenial, Olfactory, Hippocampal, Deep Subcortical; Supplementary Table S1). Regions are named without a hemisphere and individual parcels with one, for example Somatosensory_R. Coordinate-frame accuracy was validated by the Bregma–Lambda landmark check (<1% error; §2.3 below). Fig. 2 shows the Allen-32 ROIs, 30-channel surface MEA, and Antwerp brain template in the shared coordinate space.

Fig. 2.

Atlas co-registration and localization accuracy. (A–C) Allen CCFv3 ROIs, the 30-channel array (red) and the brain template in one coordinate space; sagittal X = 0.2 mm, coronal Y = 1.4 mm, axial Z = 0.9 mm. (D) Expected misplacement of each parcel’s reconstructed peak against its distance to the nearest electrode, colored by region as in A–C. Configuration throughout: ROI-based source space, ellipsoid BEM, sLORETA.

2.3. Electrode registration

The 30-channel NeuroNexus surface microelectrode array (MEA) was registered to the atlas coordinate frame: manufacturer-specified electrode positions were projected onto the skull surface (intensity-based surface detection from the atlas). Registration accuracy was validated by Bregma–Lambda distance (measured 4.16 mm vs. published ~4.20 mm (Paxinos & Franklin, 2012); <1% error).

2.4. Boundary element method head model

Three-layer (brain/skull/scalp) BEM head models were constructed in two geometric variants: a sphere BEM (three concentric spheres fitted to the atlas brain volume; brain radius = 95th percentile of voxel distances from centroid ~6.4 mm, skull/scalp scaled by 1.06 and 1.15) and an ellipsoid BEM (three nested ellipsoids fitted by principal component analysis [PCA]; semi-axes 5.73, 8.04, 3.98 mm, capturing the elongated rostro-caudal mouse brain). Each surface was discretized as a triangulated mesh (642 vertices, 1280 triangles). Tissue conductivities: brain = scalp = 0.33 S/m, skull = 0.0042 S/m (brain-to-skull ratio 78.6:1) (McCann et al., 2019; Oostenveld & Oostendorp, 2002; Vorwerk et al., 2014).

2.5. Source space construction

Three volumetric source-space approaches were evaluated (Table 2). Cartesian source spaces placed dipoles on a regular lattice through the brain volume, spaced automatically to a budget of seven sources per channel so that the source budget is matched to the ROI-based space rather than fixed by a grid pitch (141 and 82 retained sources under ellipsoid and spherical BEM; Table 2) and Shell source spaces used concentric shells matched to the BEM geometry; both assigned each source post-hoc to an Allen-32 ROI by direct nearest-atlas-voxel lookup. ROI-based source spaces placed sources directly within ROI boundaries at creation time, unambiguously assigning each source to a single Allen-32 ROI and eliminating boundary ambiguities inherent in post-hoc assignment. Sources were allocated across ROIs in proportion to ROI volume under a fixed total budget of 200 sources, subject to a floor of one source per ROI. For the Allen-32 parcellation this yielded 199 allocated sources (mean 6.2 per ROI, range 1–18), increased to 204 after enforcing left–right mirror symmetry of source placement. The total source budget and the volume-proportional allocation rule were fixed before biological analysis; sensitivity to the total budget was not tested. Four sources fall outside the spherical BEM and are dropped there, so the sphere variant carries 200; the depth distribution of the retained sources is in Supplementary Fig. S1.

Table 2.

Depth-stratified localization error (mm) for the six sLORETA configurations, sorted by 1–4 mm mean error within source-space type. The bracketed interval is the 95% CI for the 1–4 mm mean, clustered on test position; the trial-naive interval, which is roughly half as wide, is in Supplementary Table S7. Depth bins indicate distance from test dipole to nearest electrode. Test dipoles were placed on a uniform 1 mm grid (381 positions, 25 trials at SNR = 10 dB), independent of the source space, ensuring true spatial estimation rather than grid recovery. Bold row = configuration used for the biological application.

BEM	Source Space	Sources	1–4 mm	1–2 mm	2–3 mm	3–4 mm	4–5 mm	5+ mm	Runtime (min)
Ellipsoid	Cartesian	141	1.28 [1.18–1.39]	1.02	1.10	1.62	3.24	4.21	2.1
Sphere	Cartesian	82	1.73 [1.57–1.90]	1.74	1.51	1.92	1.83	2.33	1.8
Ellipsoid	ROI-based	204	1.67 [1.50–1.85]	0.93	1.49	2.34	3.91	5.08	2.3
Sphere	ROI-based	200	2.14 [1.96–2.33]	1.77	2.15	2.39	3.08	3.75	2.3
Ellipsoid	Shell	139	1.82 [1.62–2.03]	1.04	1.47	2.67	4.07	4.33	2.0
Sphere	Shell	91	2.17 [2.01–2.35]	1.85	1.95	2.58	3.58	4.17	1.7

2.6. Inverse solution

Three linear inverse methods were evaluated: minimum norm estimate (MNE; L2-regularized (Hämäläinen & Ilmoniemi, 1994)); dynamic statistical parametric mapping (dSPM; MNE normalized by estimated noise sensitivity (Dale et al., 2000)); and standardized low-resolution electromagnetic tomography (sLORETA; MNE normalized by estimated variance, zero localization error for single dipoles under idealized conditions (Pascual-Marqui, 2002)). sLORETA was the prospective primary based on this property and prior benchmarking (Grech et al., 2008); all three were nonetheless validated. The zero-error property concerns where the peak is placed, not how concentrated the reconstruction is; point-spread dispersion and deep peak contrast are reported in §3.1.2. The regularization parameter λ² (lambda-squared) was set to 1/9 (corresponding to SNR = 3).

2.7. Validation: dipole simulation

We benchmarked pipeline performance using forward-inverse consistency testing (Michel et al., 2004): simulated dipoles were placed at known brain locations, synthetic scalp EEG was generated via the forward model with added Gaussian sensor noise, the inverse solution was applied, and localization error (Euclidean distance to peak estimated source, mm) and atlas-label attribution were computed. Because a simulated dipole must be at an existing source-space location, each requested position was snapped to the nearest available source; error is measured from the requested position and so includes the discretization cost of the source space.

The validation comprised the test categories in Table 1. The baseline swept all 18 configurations over a uniform 1 mm grid of 381 positions, 25 trials each at SNR = 10 dB, giving 9,525 simulations per configuration. Atlas-label attribution was measured separately, from 20 in-parcel probes per parcel in the selected configuration (§3.1.5). The conductivity and SNR sweeps were restricted to the 6 sLORETA configurations, following sLORETA’s prospective selection.

Table 1.

Validation test design. Coverage is staged rather than uniform. Localization error and the end-to-end configuration sweep span all 18 configurations. Conductivity, noise and point-spread cover the sLORETA family, sLORETA having been prospectively selected and confirmed best at baseline. Attribution covers the ROI-based family at parcel centroids and the selected configuration at 20 probes per parcel, and the multiple-source, extended-source and fidelity tests compare the selected configuration against a Cartesian source space. The point-spread row covers eight configurations rather than six: the sLORETA family plus MNE and dSPM at the selected geometry, so the three inverse operators can be compared. Every test but the configuration sweep used forward-inverse dipole simulation with depth-stratified evaluation.

Category	Configurations	Simulations	Details	Reported in
Baseline	18	171,450	3 source spaces × 2 BEM × 3 inverse methods; uniform 1 mm grid, 381 positions × 25 trials at SNR = 10 dB	§3.1.1; Table 2; S7
Conductivity	6	400,050	All sLORETA configs; brain: skull ratios 10:1–640:1 against a true 78.6:1	§3.1.4; S9
SNR sensitivity	6	285,750	All sLORETA configs; 0 to 20 dB in 5 dB steps	§3.1.4; S9
Attribution, centroid probes	6	4,800	ROI-based family, 2 BEM × 3 inverse; one probe at each parcel centroid × 25 trials	§3.1.3; S7
Attribution, in-parcel probes	1	16,000	Selected configuration; 20 probes spanning each parcel × 25 trials	§3.1.3, §3.1.5; Table 3; S10
Point-spread	8	1,265	One noise-free on-grid probe per source; dispersion and peak contrast	§3.1.2; S8
Two co-active	2	4,800	48 parcel pairs, correlation 0 and 0.9, 25 trials	§3.1.6; S11
Three co-active	2	1,000	20 depth-stratified triplets, 25 trials	§3.1.6; S11
Extended sources	2	4,800	Patch radii 0.5, 1.0 and 1.5 mm on each parcel	§3.1.6; S11
Fidelity	2	3,200	Single and two-source, waveform 30–55 Hz	§3.1.6; S11
Configuration sweep	18	—	Not a simulation: the full pipeline re-run on all 35 recordings in each of the 18 configurations, for 5 pre-committed outcomes	§3.6; Fig. 7; S12
Total		893,115

The primary metric was localization error; the secondary metric was atlas-label attribution, reported only for ROI-based source spaces, where sources are placed directly within atlas ROI boundaries. All metrics were computed in 1 mm depth bins from the nearest electrode (Table 2).

Each simulated dipole had a fixed moment of 50 nAm and an isotropic random orientation. Seeds were set from the position and trial indices alone, so every configuration met the same requested positions and the same draws; the forward model and source-space discretization still differ. Per-configuration runtime is given in Table 2 and Supplementary Table S7.

2.8. Animals and EEG recording

This is an observational genotype comparison (Fmr1 KO vs. WT) within the vehicle-treated arm of a parent gene therapy study (Lacher et al., 2026); all animals received the same adolescent (6–7 weeks) intracerebroventricular (ICV) vehicle injection, were implanted with a 30-channel NeuroNexus surface MEA, and contributed resting EEG. Procedures complied with the NIH Guide for the Care and Use of Laboratory Animals and were IACUC-approved at Cincinnati Children’s Hospital Medical Center; reporting follows ARRIVE 2.0 (with STROBE for the observational comparison; checklist in Supplementary Table S6). Full husbandry, stereotaxic, and surgical detail is in Extended Methods §EM-1.

EEG was recorded from 36 male C57BL/6J mice (n = 18 KO, n = 18 WT pre-QC) at 76–85 days of age (mean 78.5 ± 3.5 days). Animals were sourced from The Jackson Laboratory (separate KO and WT orders, not litter-matched). Genotype was masked at all stages (recording, preprocessing, and statistical analysis) and unblinded only after the primary models had been finalized. At 6–7 weeks all mice received bilateral ICV vehicle (PBS with 0.001% pluronic F-68; 2 μL/ventricle at bregma AP −0.3, ML ±1.0, DV −2 mm) under isoflurane; the MEA was implanted 4 weeks later, and resting EEG was recorded 1 week post-implant (15-min acclimation, 30 min resting at 1000 Hz). Generalization to injection-naive animals is addressed in Limitations.

Recordings were processed offline in EEGLAB (MATLAB 2018a): downsampled to 500 Hz, high-pass filtered at 0.5 Hz, band-stop filtered at 57–65 Hz, manually screened for artifacts, with bad channels replaced by spherical spline interpolation (capped at 10% per subject) and an average reference applied.

Sample size.

The sample size was the full vehicle-treated cohort of the parent gene therapy study (n = 18 KO + 18 WT pre-QC); it was not derived from a power calculation. At this size, detectable subject-level Hedges’ g with 80% power and α = 0.05 is roughly |g| ≥ 0.97 (two-tailed Welch’s t); the headline gamma findings (g = 1.28 Low Gamma, 1.26 High Gamma; both 95% CI lower bounds > 0.5) sit comfortably above this floor, while null or modest-effect ROI findings are not evidence of absence. The sample is comparable to prior Fmr1 KO mouse EEG studies (Jonak et al., 2020; Lovelace et al., 2018).

Exclusions and missing data.

Of 36 attempted recordings (18 KO + 18 WT), one WT subject was removed by a pre-specified automated QC threshold (broadband power z = −2.89, ≈4 × below the group mean), yielding the analyzed sample of n = 18 KO + n = 17 WT (35 total). No other animals were excluded; no recording was lost to equipment failure or post-hoc judgment. All 35 retained subjects contributed complete data on every reported outcome (no per-outcome missingness). Retained-epoch counts after artifact removal did not differ between groups (Welch’s t(25.0) = 1.15, p = 0.26).

2.9. Electrical source imaging

The pipeline-selection sequence was: (i) the 18-configuration benchmark was run on synthetic dipoles before any biological data were loaded; (ii) the analysis configuration (ROI-based + ellipsoid BEM + sLORETA) was selected using a three-criterion rule (source-to-atlas correspondence as primary, depth-stratified classification accuracy, and whole-brain localization error; analytical-vs-accuracy tradeoff in §4.1); (iii) biological EEG was loaded and analyzed only after the configuration was locked.

2.10. Epoch subsampling

To reduce between-subject variance from unequal epoch counts (KO: 118–627; WT: 96–662), we used bootstrap subsampling: 80 non-overlapping 2-s epochs per subject × 500 draws (fixed seed = 42), with subject-level estimates averaged across draws. We preferred subsample-and-average over modeling epoch count as a precision covariate (which would conflate measurement-precision with the biological contrast) or inverse-variance weighting (the weights would themselves be epoch-count estimates, creating circularity); equalizing data length per subject removes the imbalance at source rather than adjusting downstream. Subsample-size and draw-count justification are in Extended Methods §EM-2.

2.11. Frequency bands and aperiodic parameterization

Power spectral density was computed via Welch’s method (2-s Hann windows, 50% overlap) and integrated within six frequency bands: Delta [1–4 Hz], Theta [4–10 Hz], Alpha [10–13 Hz], Beta [13–30 Hz], Low Gamma [30–55 Hz], and High Gamma [65–80 Hz]. The Delta through Low Gamma definitions follow Jonak et al. (2020); the 55–65 Hz gap brackets the 57–65 Hz preprocessing notch around 60 Hz line noise, and the 80 Hz High Gamma upper bound was selected as a conservative cap below the frequencies at which myogenic artifact typically dominates in freely behaving rodents.

Aperiodic spectral parameters were estimated from the same Welch power spectra using the specparam algorithm (the spectral-parameterization successor to FOOOF; Donoghue et al., 2020). Spectra were parameterized over 2–50 Hz in the fixed aperiodic mode (no spectral knee), with peak-width limits of 1–12 Hz, a maximum of six periodic peaks, and a minimum peak height of 0.1. The window was prespecified and lies within the analyzed range below the 55–65 Hz notch. The aperiodic exponent (slope of the 1/f component) and offset (broadband power) were retained as outcomes; periodic peak parameters were not analyzed here.

2.12. Statistical analysis

Model specification.

Source-level band power was analyzed using linear mixed-effects models (LMMs) at two prespecified spatial tiers: a 10-region model, power ~ group × region + (1 | subject), and a 32-ROI model, power ~ group × ROI + (1 | subject). Both used Type III ANOVA with Satterthwaite degrees of freedom. Mixed models were fitted in R 4.6.1 with lme4 2.0.1, lmerTest 3.2.1 and emmeans 2.0.3, and spectral parameterization used specparam 2.0.0rc6 under Python 3.12 with MNE-Python 1.11.0. The Python environment is pinned in the releases cited in Data and software availability. The random-effects term is (1 | subject) only because each subject × spatial-unit cell contains a single mean per band, making a nested (1 | subject:ROI) unidentifiable. The exception is the six-region electrode aperiodic model, (1 | subject) + (1 | subject:region), where each subject × region cell aggregates ~5 channels; the six-region band-power model carries (1 | subject) only (Extended Methods EM-3). Two standardized effect sizes are reported. Model-based standardized mean difference (SMD) is the emmeans group contrast divided by the LMM residual standard deviation and is used for all mixed-model effects. Subject-level Hedges’ g, small-sample corrected (Hedges & Olkin, 1985), is used where subject means are compared directly, and the two are not interchangeable.

Multiplicity control.

Benjamini–Hochberg false-discovery-rate (BH-FDR) correction was applied within two prespecified families: oscillatory (12 tests = 6 bands × 2 power types) and aperiodic (2 tests = exponent, offset). FDR was applied independently for each contrast, spatial tier (10-region or 32-ROI), and effect type (group main effect vs. Group × spatial-unit interaction). Supplementary Table S2 lists FDR-significant and borderline rows; the full per-test inventory is deposited with the source data. Post-hoc per-ROI contrasts were gated by a significant omnibus group main effect or a significant Group × spatial-unit interaction at the matching tier, with Holm step-down correction within each (band × power-type) cell across the spatial units in that tier (10 or 32). Adjusted values are reported as q_adj. Source and electrode analyses differ in spatial granularity (32-ROI / 10-region vs. six-region / 30-channel), in multiplicity burden, since BH-FDR family sizes and gating differ, and in random-effects structure as noted above; differences in detection can reflect any of these and do not isolate localization sensitivity.

Sensitivity.

As a sensitivity to per-family BH-FDR, we re-applied BH-FDR pooled across the entire prespecified omnibus list (126 tests), the most conservative defensible single-family correction. Outcomes are reported in Supplementary Table S2.

2.13. Bias and confounding

The disease contrast (KO vs. WT) is observational, so we audited measured cohort variables to rule out batch- or time-related confounding. Recordings spanned five acquisition cohorts (FS8, FS9, FX9, sbpro, allego) over a 14-month window. Cohort was balanced between groups (χ² = 2.04, p = 0.73; permutation p = 0.84), recording dates overlapped (Mann-Whitney U = 141, p = 0.70; medians within 5 days), and prebootstrap epoch counts did not differ (Mann-Whitney p = 0.33). Per-cohort counts in Supplementary Table S3.

To confirm the headline findings did not depend on this balance, each significant disease-effect outcome was refit with the cohort tag as a fixed-effect covariate (subject-level OLS regression on the 32-ROI mean). Adjusted estimates closely tracked unadjusted and no finding changed status (Low Gamma p = 0.001, High Gamma p = 0.002 after adjustment); per-outcome estimates are in Supplementary Table S4.

Several potential confounders we could not measure or did not vary are: sex (all animals male); age (tightly controlled by design, 76–85 days, SD = 3.5 days); body weight at recording; circadian phase (recordings during the light cycle, not Zeitgeber-locked); and litter relatedness (separate KO and WT shipments, not littermate-matched). These limitations are addressed in §4.8.

3. Results

3.1. Simulation-based benchmarking

These results are simulation-based benchmarks of pipeline performance via forward-inverse consistency testing, not estimates of anatomical accuracy in biological tissue.

3.1.1. Localization error across configurations

All sLORETA configurations achieved approximately 1 mm localization error at superficial cortical depths (1–2 mm from electrodes), with error increasing systematically with depth (Table 2).

Within the ellipsoid BEM + sLORETA family, ROI-based and Cartesian were similar superficially (1–2 mm: 0.93 vs 1.02 mm), with Cartesian achieving lower mean error across 1–4 mm (1.28 vs 1.67 mm). Among volume source spaces, Cartesian sLORETA was the only configuration to achieve lower 1–4 mm error than the selected one, by 0.386 mm (95% CI 0.267–0.512; position-level paired bootstrap over the 222 shared test positions in that depth range). Cartesian and Shell source spaces require post-hoc atlas assignment; ROI-based assigns at construction. ROI-based was selected as primary on this basis (selection-timing protocol in §2.9; analytical-vs-accuracy tradeoff in §4.1).

3.1.2. Systematic effects on localization error

Inverse method was the dominant factor. Averaged over 1–4 mm, sLORETA gave the lowest error in five of six geometries. At the most superficial bin the separation narrows: sLORETA and MNE overlap (0.93–1.85 and 1.00–1.64 mm) while dSPM is clearly worse (1.66–3.02 mm). Ellipsoid BEM achieved lower superficial errors and better numerical conditioning (condition number 251 vs. 1148 for sphere) but steeper depth degradation. All configurations showed systematic depth dependence (Table 2). Spatial specificity is conventionally characterized by point-spread and cross-talk functions (Hauk & Stenroos, 2014); we report the point-spread function and, in place of the cross-talk function, the peak-label attribution of §3.1.5. Median point-spread dispersion is 4.65 mm for the selected configuration and 4.52–5.25 mm across the ellipsoid family, against a median parcel root mean square radius of 1.35 mm: the reconstruction of a single source is several times wider than the parcel it is assigned to. Peak contrast falls from 4.49 in the most superficial bin to 1.43 beyond 5 mm, so at depth the point-spread function is nearly flat. Its peak, and therefore the zero peak error of §2.6, then carries little information (Supplementary Table S8).

3.1.3. ROI attribution

ROI attribution was strongly depth-dependent. Recall was 69.6% at 1–2 mm, 37.3% at 3–4 mm and 1.7% beyond 5 mm, 37.5% overall across the 32 parcels (Table 3), and median precision fell alongside it from 68% to 10%, so deep parcels are not merely missed: the labels assigned to them are also unreliable. Recall and precision describe different failure modes, which a single accuracy figure cannot separate. Attribution was also checked across the ROI-based family at parcel centroids, where the selected configuration recalls 44.1% against 34.4% for the spherical BEM and 24.0–26.4% for MNE and dSPM: the inverse operator matters more to attribution than the head-model geometry does, and the prospectively selected operator is the best of the three (Supplementary Table S7).

Table 3.

Atlas-label attribution by source depth, for the selected configuration. Depth-bin aggregates of the per-parcel values in Supplementary Table S10, from 20 in-parcel probes per parcel rather than from the 381-position baseline grid. Recall is not monotonic with depth: the 2–3 mm bin holds three parcels and 1,500 trials, and its 34.3% sits below the seven-parcel 3–4 mm bin.

Depth (mm)	Parcels	Trials	Recall	Median precision
1–2	9	4,500	69.6%	68%
2–3	3	1,500	34.3%	40%
3–4	7	3,500	37.3%	27%
4–5	10	5,000	20.4%	27%
5+	3	1,500	1.7%	10%
All	32	16,000	37.5%	32%

3.1.4. Robustness

Robustness was tested across the 6 sLORETA configurations, sweeping conductivity and SNR (§2.7). Ellipsoid BEM was sensitive to conductivity in the expected direction (r = 0.76–0.93), its 1–4 mm minimum at the true ratio. Sphere BEM was largely insensitive. Four of six showed the expected negative SNR–error relationship, shell with ellipsoid BEM was flat (r = −0.11), and sphere with Cartesian rose slightly with SNR, indicating a floor set by model mismatch rather than by noise. Full results are in Supplementary Table S9.

3.1.5. ROI-label confusion

We tabulated the full 32 × 32 confusion matrix, reporting recall, precision and the displacement between the true parcel and the one receiving the reconstructed peak (Supplementary Table S10). Displacement measures how far the estimate moved (mean 2.45 mm, against 3.8–6.9 mm under random assignment) and sorts the parcels into four tiers (Supplementary Fig. S3). Precision exposes a separate failure: bilateral Lateral Cortex attracts more than 1100 trials each originating elsewhere, at precision 7–8%, and is treated as unresolved despite its cortical position. Misattribution is depth-dependent: cortical parcels sent errors predominantly to adjacent parcels (85.0%), deep structures to distant ones (hypothalamic 84.4%). Left–right confusion was 5.9% of errors pooled across trials, or 9.0% as the mean of per-parcel fractions; the olfactory bulbs are two thirds homotopic, so olfactory activity is well localized yet cannot be assigned to a hemisphere. The parcels carrying the genotype effects below have low expected displacement (Somatosensory_R 1.17 mm, bilateral Frontal-Anterior 1.65–1.70 mm), except Hippocampus_Post_R at 2.33 mm and Thalamus_R at 3.08 mm; precision differs among them, 77% for Somatosensory_R against 39–40% for Frontal-Anterior. These bounds govern the regional language used in §4.

3.1.6. Multiple sources, extended sources, and time-series fidelity

With two co-active parcels, both appeared among the two highest-power reconstructed parcels in 14.3% of trials at zero correlation and 13.0% at r = 0.9, rising to 31.5% for homotopic pairs; recovery was depth-dependent, from 45.6% at 1–2 mm to 8.2% at 3–4 mm. With three co-active parcels, all three were recovered in none of the 20 triplets tested, and 85.1% of reconstructed power fell outside the driven parcels (Supplementary Table S11).

Reconstructed extent was 3.79–3.80 mm irrespective of simulated extent, so extent is set by the point-spread function, not the source. The source space is sparse enough that requested radii of 0.5–1.5 mm produced patches of only 0.05–0.98 mm (Supplementary Table S11).

For a single source, simulated and recovered parcel courses correlated at |r| = 0.96–0.99 across depth bins; with a second source this fell from 0.97 superficially to 0.57 beyond 5 mm. Polarity is not identifiable, so correlations are absolute. The waveform was band-limited to 30–55 Hz, so this bounds Low Gamma recovery and does not extend to aperiodic estimation.

3.2. Source-level power spectral density

Grand-average power spectral density curves at representative Allen-32 ROIs (KO n = 18, WT n = 17) showed visible KO-vs-WT separation in the gamma range, particularly in Frontal-Anterior and Somatosensory cortex (Fig. 3).

Fig. 3.

Grand-average source-level PSD at six representative Allen-32 ROIs. KO (red, n = 18) and WT (blue, n = 17), mean ± SEM. Shading marks the frequency bands; the hatched region is the 55–65 Hz notch gap. ROI-based, ellipsoid, sLORETA.

3.2.1. Global group effects

Group main effects are reported here with unadjusted p values from the omnibus LMMs (see Statistical Analysis); BH-FDR applies to the omnibus tests and Holm to the per-ROI / per-region post-hoc contrasts reported below. Low Gamma [30–55 Hz] absolute power was significantly elevated in KO relative to WT (F = 15.10, p < 0.001, SMD = 1.02 [95% CI 0.48–1.55]). High Gamma [65–80 Hz] absolute power was also elevated (F = 14.45, p < 0.001, SMD = 0.67 [95% CI 0.31–1.03]). Both survive BH-FDR (q = 0.004) and reproduce at the 10-region tier (Low Gamma F = 13.47, High Gamma F = 12.94; both q = 0.006). Beta [13–30 Hz] absolute power was borderline at the unadjusted threshold (F = 3.85, p = 0.058, SMD = 0.66 [95% CI −0.02–1.35]). No other bands showed significant main group effects.

3.2.2. Group × ROI interactions

At the 32-ROI tier, three of six relative-power bands showed significant Group × ROI interactions after FDR (Supplementary Table S2): Delta (F = 3.38, q < 0.001), High Gamma (F = 1.92, q = 0.012), and Theta (F = 1.76, q = 0.026). No absolute-power interaction reached significance (all q ≥ 0.136), so the gamma group main effect reflects a spatially distributed shift rather than a localized one. At the 10-region tier, neither gamma absolute interaction reached significance (High Gamma F = 2.34, q = 0.087; Low Gamma F = 1.78, q = 0.172); Somatosensory and Frontal-Anterior carry the largest individual SMDs descriptively.

3.2.3. Post-hoc contrasts

At the 10-region tier, five regions survived Holm correction for Low Gamma absolute and four for High Gamma absolute, all KO > WT (largest: Somatosensory SMD = 2.07 / 1.72; Frontal-Anterior SMD = 1.98 / 1.53). At the 32-ROI tier, nine ROI-by-band results, across six distinct ROIs, survived Holm correction within their (band × power-type) cell, all in Low or High Gamma absolute and all KO > WT (largest: Thalamus_R Low Gamma SMD = 1.92, q_adj < 0.001; Somatosensory_R Low Gamma SMD = 1.79, q_adj = 0.001; bilateral Frontal-Anterior also surviving, together with posterior hippocampal and olfactory-bulb parcels whose hemispheric assignment is not resolved by this pipeline (§3.1.5); full list in Fig. 4).

Fig. 4.

Genotype differences in gamma power (KO – WT). Each ROI’s Hedges’ g is spread by a Gaussian whose standard deviation is that parcel’s own measured peak displacement (0.33–5.87 mm; Supplementary Table S10). The map shows where an effect can be attributed, not a new estimate; a value at any point is not that parcel’s tested effect size. Bold outlines mark Holm q < 0.05, and those parcels and their effect sizes are given in §3.2.3. ROI-based, ellipsoid, sLORETA.

Delta relative power produced one of the strongest Group × ROI interactions (F = 3.38, q < 0.001), with no significant global group difference in absolute Delta. The spatial pattern (Fig. 5) is bidirectional: ventral and lateral cortical ROIs show KO > WT, dorsal cortical and several deep subcortical ROIs KO < WT.

Fig. 5.

Genotype differences in relative Delta power (KO – WT). Color and scale as Fig. 4: each parcel’s Hedges’ g spread by its own measured peak displacement, showing where an effect can be attributed rather than a new estimate. No ROI reached Holm q < 0.05, so no outlines appear. ROI-based, ellipsoid, sLORETA.

3.3. Aperiodic spectral parameters

Neither exponent nor offset showed a significant main group effect (exponent: F = 1.26, p = 0.269; offset: F = 0.06, p = 0.811). Both showed significant Group × ROI interactions (exponent: F = 1.82, q = 0.008; offset: F = 1.52, q = 0.035; FDR-corrected within the 2-test aperiodic family). Under maximally-pooled global BH-FDR across the full prespecified omnibus list (Methods §2.12 sensitivity note), the exponent interaction remains significant whereas the borderline offset interaction does not (global q = 0.160).

No individual ROI survived Holm correction for either parameter. For exponent, all 32 ROIs showed negative effect sizes (KO < WT; Fig. 6). For offset, effects were mixed (16 of 32 ROIs positive). The strongest uncorrected exponent effects fell in Visual_Parietal_L, bilateral Frontal-Anterior and bilateral Hippocampus, parcels whose attribution ranges from the best in the study to among the weakest (Supplementary Table S10), so no anatomical pattern is read from them (§4.8).

Fig. 6.

Genotype differences in the aperiodic exponent (KO – WT). Color and scale as Fig. 4: each parcel’s Hedges’ g spread by its own measured peak displacement, showing where an effect can be attributed rather than a new estimate. No ROI reached Holm q < 0.05. The offset (F = 1.52, q = 0.035), 96% collinear with the exponent at subject level, is reported with it rather than mapped. ROI-based, ellipsoid, sLORETA.

3.4. Electrode-level analysis

At the 30-channel tier, electrode-level LMMs detected significant group main effects for Low Gamma absolute (F = 8.77, p = 0.006) and High Gamma absolute (F = 8.57, p = 0.006). At the six-region tier, post-hoc contrasts survived Holm correction for Low Gamma and High Gamma absolute in four of the six regions. No Group × Region interaction reached FDR significance at the six-region tier under the prespecified 12-test oscillatory family; Delta relative power was the strongest but did not survive correction (F = 2.48, raw p = 0.031, q = 0.367).

For aperiodic parameters, the six-region nested electrode LMM showed no group main effect for either exponent (F = 1.37, p = 0.250) or offset (F = 0.19, p = 0.667), and no Group × Region interaction (exponent F = 1.11, p = 0.351; offset F = 0.59, p = 0.705). At the 30-channel electrode tier (with subject as the only random effect), neither exponent (F = 1.42, p = 0.241) nor offset (F = 0.20, p = 0.655) showed a significant group main effect, but both showed significant Group × Channel interactions (exponent q = 0.026; offset q = 0.019; BH within the 2-test aperiodic family). Neither survives the pooled correction of Supplementary Table S2 (q = 0.136 and 0.067).

3.5. Electrode vs source comparison

Subject-level electrode and source power were strongly correlated across all 12 band × power-type combinations (r = 0.77–0.94, all p < 0.001; Table 4). KO and WT source–electrode discrepancies did not differ for any outcome (all p > 0.10, Welch’s t-test).

Table 4.

Source versus electrode at matched granularity. Subject-level correlation between source and electrode power, the omnibus Group × unit test at each level, and the number of spatial units surviving Holm correction, per band and power type. The electrode column is the 30-channel tier throughout, matched to the source tier in granularity and in correction breadth. Full per-band statistics, and the six-region tier, are in Supplementary Table S13.

(a) Absolute power (bands) and aperiodic parameters.
Band / parameter	r a	Group × ROI q (32 ROI)	Group × channel q (30 ch)	Post-hoc surviving Holm (source · electrode)b
Delta	0.93	0.136	0.099	—
Theta	0.92	0.212	0.035	— · 0 / 30
Alpha	0.87	0.136	0.035	— · 0 / 30
Beta	0.91	0.220	0.035	— · 0 / 30
Low Gamma	0.94	0.220	0.284	6 / 32 · 1 / 30
High Gamma	0.80	0.136	0.773	3 / 32 · 1 / 30
Exponentc	0.93	0.008	0.026	0 / 32 · 0 / 30
Offsetc	0.92	0.035	0.019	0 / 32 · 0 / 30
(b) Relative power (bands).
Band	r	Group × ROI q (32 ROI)	Group × channel q (30 ch)	Post-hoc surviving Holm (source · electrode)
Delta	0.94	< 0.001	0.099	0 / 32 · —
Theta	0.88	0.026	0.035	0 / 32 · 0 / 30
Alpha	0.77	0.673	0.502	—
Beta	0.85	0.298	0.485	—
Low Gamma	0.94	0.115	0.008	— · 1 / 30
High Gamma	0.87	0.012	0.042	0 / 32 · 0 / 30

^ar = Pearson correlation between subject-level source and electrode power (n = 35, all p < 0.001); subject-level effect sizes are Hedges’ g, small-sample corrected (Hedges & Olkin, 1985).

^bPost-hoc counts are ROIs (of 32) and channels (of 30) surviving Holm correction, source before the separator; a dash means post-hoc contrasts were not computed, because neither omnibus effect was significant.

^cFor the aperiodic parameters the six-region tier is reported in Supplementary Tables S2 and S13; the electrode column here is the 30-channel tier. Cross-level differences are descriptive: the units are not interchangeable and the correction families differ, so no formal cross-level test is conducted.

Absolute power.

Subject-level g_subj at source modestly exceeded the matched electrode value in five of six bands (largest Δg_subj 0.30 in Low Gamma; see §4.5 for interpretation). No gamma absolute Group × Region interaction reached significance at any tier (source 10-region: High Gamma q = 0.087, Low Gamma q = 0.172; matched electrode 30-channel: Low Gamma q = 0.284, High Gamma q = 0.773; unmatched electrode six-region: both q = 0.978; source 32-ROI: all q ≥ 0.136), which matches a spatially distributed effect captured comparably by both tiers in absolute units.

Relative power.

g_subj values were small at both levels (|g_subj| ≤ 0.66), with direction varying across bands, matching the fact that relative power is a distributional rather than total-magnitude measure. The omnibus diverges: three of six relative bands showed a significant Group × ROI interaction at source 32-ROI (Delta q < 0.001; High Gamma q = 0.012; Theta q = 0.026), and three of six at the matched electrode 30-channel tier, though not the same three (Theta q = 0.035; Low Gamma q = 0.008; High Gamma q = 0.042). At the unmatched six-region tier none reached significance (smallest q = 0.367, Delta relative).

Aperiodic.

Source 32-ROI Group × ROI interactions were significant for both exponent (q = 0.008) and offset (q = 0.035). The six-region electrode model showed no interaction for either parameter (exponent q = 0.701; offset q = 0.705); at the 30-channel tier both interactions were significant (exponent q = 0.026; offset q = 0.019).

3.6. Stability across pipeline configurations

To test dependence on the forward- and inverse-modeling choices, the pipeline was run across 18 configurations: three source spaces (ROI-based, Cartesian, shell) × two BEM geometries × three inverse operators, with subjects, bands and models held identical. Depth weighting is the one exception: it is 0.0 in the ROI-based and shell arms and 0.8 in the Cartesian arm, inherited from the pipeline presets, so the arms are not matched on a parameter that acts on depth. It is 0.0 throughout the simulation benchmark of §3.1. The full grid is in Supplementary Table S12 and Fig. 7.

Fig. 7.

Genotype effects across 18 pipeline configurations. Five pre-committed outcomes across three source spaces (ROI-based, Cartesian, shell) × two BEM geometries × three inverse operators, grouped by source space. Group main effects are the model-based SMD with 95% CI, the estimator of §2.12 and Supplementary Table S12; interactions are F, since a pooled effect size cannot represent spatial heterogeneity. Filled markers are FDR-significant; the dashed rule marks the selected configuration.

Low Gamma absolute power was elevated in KO in all 18 configurations and significant in 17 (SMD 0.74–1.25; the 1.02 of §3.2.1 falls mid-range), the exception (shell, spherical, dSPM; q = 0.063) being directionally consistent. High Gamma was elevated in all 18 and significant in 15, the exceptions all spherical and retaining their effect size. Across the outcome × ROI SMDs, configurations correlated with the primary at r = 0.74–0.92 (descriptive). The gridded spaces do not all resolve the full atlas (19 to 31 of the 32 parcels, spherical geometries worst), so each correlation is over the parcels its configuration resolves.

The interactions were less stable, systematically so. Delta relative reached significance in 15 of 18, and the aperiodic exponent and offset in 11 each, divided by head-model geometry rather than source space (ellipsoidal 8 of 9, spherical 3 of 9, for both). Exponent and offset are 96% collinear at subject level and are reported here as one aperiodic result, not two. The exponent contrast was negative in direction in all 18 (SMD −0.73 to −1.07, a summary of direction rather than the tested term, with intervals spanning zero) even where the interaction was null, though the group main effect is not significant in the primary configuration (§3.3); that distinction, global versus regional, is the one that separates two claims (§4.8).

4. Discussion

4.1. Pipeline selection: sLORETA, ellipsoid BEM, ROI-based source space

The benchmark answers the prior question of feasibility: region-resolved electrical source imaging is tractable from a 30-channel mouse array, with depth-graded error and atlas-label attribution bounded across an 18-configuration space.

The ROI-based source space was selected over Cartesian even though Cartesian achieved slightly lower whole-brain localization error (1.28 vs 1.67 mm at 1–4 mm depth). The optimization criterion was not localization error but the removal of a free parameter from ROI-level inference: ROI-based places dipoles inside ROI boundaries at construction, so every source carries an unambiguous single-ROI label, whereas gridded spaces must assign sources to parcels afterwards by a rule the ROI-level result then depends on (Hipp et al., 2012; Sanchez-Bornot et al., 2021; Kida et al., 2023). In this comparison both gridded spaces used nearest-atlas-voxel assignment. The cost of the choice is 0.386 mm of mean localization error over 1–4 mm (95% CI 0.267–0.512). That comparison is between source spaces matched on source budget rather than against a dense lattice, which is the better-controlled contrast but leaves the high-resolution case untested.

The biological conclusions are stable across that choice: the anchor findings replicate in both Cartesian and Shell, and the per-ROI effect-size vectors agree with the ROI-based pipeline at r = 0.85 and 0.84 (0.79 between the two gridded spaces). Which source space is used is therefore a question of how ROIs are assigned, not of which regions the data implicate.

One structural difference follows from geometry rather than from performance: a uniform lattice does not guarantee that every parcel receives a source, and under nearest-voxel assignment two parcels receive none, whereas ROI-based places sources within each parcel by construction. This favors ROI-based for atlas-anchored ROI-level analyses rather than indicting the alternatives, which are better suited to whole-brain mapping.

The configuration was locked from the simulation benchmark before biological data were loaded (Methods §2.9). sLORETA showed depth-invariant peak localization, which concerns the placement of the peak alone and not the concentration of the estimate (§3.1.2). Error-versus-depth correlations are not by themselves a ranking criterion, because a flatter profile can also be produced by a central bias that pulls estimates toward the middle of the head; we therefore base the preference on peak localization and on ROI-level attribution (§3.1.5) rather than on the slope of error against depth. Ellipsoid BEM additionally showed the expected U-shaped conductivity sensitivity, with its 1–4 mm minimum at the true 78.6:1 ratio and 1–4 mm error at or below 2.16 mm across an 8-fold band of assumed ratios. Whether these simulation rankings extend to biological recordings (where distributed, correlated sources are common) remains to be tested empirically.

4.2. Regional heterogeneity of elevated gamma-band power

Absolute gamma elevation was observed at both source-level omnibus tiers without a significant Group × ROI interaction (§3.2), indicating a bilaterally distributed elevation rather than one concentrated in a single ROI; nine ROI-by-band results across six distinct ROIs nonetheless survive Holm correction within their (band × power-type) cell. Of these, the largest dorsal-cortex effects (peak displacement 1.17–1.70 mm) fall in Somatosensory_R (Low Gamma SMD = 1.79) and bilateral Frontal-Anterior (1.40–1.58); the largest deep effects (Thalamus_R = 1.92, Hippocampus_Post_R, Olfactory_Bulb_L) sit deeper, where peak displacement reaches 2.33 mm and 3.08 mm and individual-ROI attribution carries higher uncertainty (Supplementary Table S10; §4.8). The frontal and olfactory findings are not independent of one another: a third of trials originating in bilateral Frontal-Anterior are attributed to the ipsilateral olfactory bulb (33.6% and 31.0%, each parcel’s largest sink) while the return leak is under 2%, so the olfactory effect may carry frontal activity rather than the reverse (Supplementary Table S10). Recall at the dorsal parcels is itself modest (25–63%), and modest recall alongside low displacement indicates a peak falling near but outside the parcel boundary rather than in a different structure, which is what a spatially dispersed elevation would produce and what the absent interaction independently implies. The High Gamma relative-power interaction reaches significance (q = 0.012; Low Gamma relative did not survive correction, q = 0.115), indicating that the spectral redistribution is itself spatially patterned even where the absolute elevation is uniform, a pattern that aligns with broadband aperiodic flattening rather than a single localized inhibitory deficit. Two constraints bound how these loci are read (§3.1.5, §3.1.6): co-active ROIs are resolved simultaneously in a minority of trials, though far more often at the depths these loci occupy than the pooled rate suggests (§3.1.6), so these are the strongest contributors to a spatially broad effect rather than independently resolved generators; and deep misattribution is predominantly non-local, so elevation at thalamic, hypothalamic, amygdalar and basal-ganglia labels is deep-structure activity of unresolved origin.

The Frontal-Anterior and Somatosensory loci, lying within dorsal cortex, parallel reported regional vulnerability of parvalbumin-expressing (PV+) interneurons in Fmr1 KO frontal cortex (Contractor et al., 2015; Selby et al., 2007) and disrupted inhibitory maturation in barrel cortex (Harlow et al., 2010); PV+ interneurons regulate gamma oscillations (Cardin et al., 2009) and reduced PV+ function disinhibits excitatory networks (Gibson et al., 2008; Patel et al., 2013). The deeper effects are anatomically less interpretable, and how far they should be weakened differs by structure (§3.1.5): the posterior hippocampal effect is moderately localized (2.33 mm, 45% below chance) and read as a neighborhood descriptor; the thalamic effect is weakly localized (3.08 mm, 19% below chance), better than random but not by a margin that supports naming the structure; and olfactory misattribution is two thirds contralateral, so no side-specific claim is made. The overall pattern is consistent with a PV+/E−I account combined with brain-wide spectral flattening; this remains a testable hypothesis rather than a demonstrated mechanism.

4.3. Aperiodic spectral parameters

The aperiodic Group × ROI interactions (§3.3), with KO showing negative exponent effect sizes in all 32 ROIs, indicate a brain-wide spectral flattening with regionally graded magnitude rather than a localized deficit. The strongest uncorrected effects fell in lateral cortical and deep ROIs, which are the least well attributed parcels in the study (Supplementary Table S10); no ROI survived Holm correction, so the interactions support a distributed shape change rather than specific-ROI claims. Attribution uncertainty bounds this interpretation. The parcels carrying the largest uncorrected effects span the full range of attribution quality (§3.1.5), which is a further reason to read the interaction as a distributed shape change rather than as parcel-specific localization. We can therefore claim that aperiodic shape differs by genotype, but not pinpoint where. The more borderline offset interaction accordingly does not survive maximally-pooled global BH-FDR (§2.12), while the stronger exponent interaction does. The aperiodic exponent has been proposed as an indirect E/I proxy (Gao et al., 2017), but this interpretation remains debated and the present design cannot adjudicate a mechanistic account.

4.4. Relative vs absolute power

Absolute gamma power was elevated near-uniformly across ROIs, whereas relative power showed Group × ROI interactions in three of six bands. Delta is the instructive case: total Delta power did not differ between groups, yet its spatial pattern was bidirectional: ventral and lateral cortex KO > WT, dorsal and deep structures KO < WT. That is a spatially heterogeneous redistribution invisible at the global level and unresolved at the six-region electrode tier, although no individual ROI survived Holm correction for Delta relative power and the ventral and lateral parcels carrying the positive direction are among the least reliably attributed (§3.1.5).

4.5. Electrode vs source comparison

Source and electrode estimates of subject-level power were strongly correlated across all twelve band × power-type combinations (r = 0.77–0.94). Comparing the levels requires matching spatial granularity, so we compare the 32-ROI source tier with the 30-channel electrode tier, at which correction breadth is also comparable. Post-hoc contrasts for absolute gamma survived Holm correction at six of 32 ROIs for Low Gamma and three of 32 for High Gamma, against one of 30 channels in each case; regional heterogeneity was resolved in three of six relative-power bands at each level, though not the same three: Delta, Theta and High Gamma at source, Theta, Low Gamma and High Gamma at the electrode tier. One explanation consistent with the pattern, which these data cannot test, is that aggregating sources within a parcel raises the signal-to-noise ratio of each unit while removing the volume-conduction mixing that makes neighboring channels covary, so detection rises while apparent heterogeneity does not. These comparisons are descriptive rather than a formal test of sensitivity (Table 4), but they are matched in granularity and in correction breadth, which the six-region comparison is not (Supplementary Table S13).

4.6. Atlas framework trade-offs

The atlas-based design enables standardization without subject-specific imaging, but it embeds the parcellation in the source space before inversion, so the inverse cannot discover a boundary the atlas does not contain: the resolution of any result here is the parcellation, and portability to another atlas is untested. It does not generate the findings: in Cartesian spaces, labeled only after inversion, both gamma anchors reproduce in all six configurations (§3.6). The design also assumes adult C57BL/6J geometry, so cross-strain, cross-age and disease-related variation cannot be represented.

4.7. Comparison to prior work

Multichannel mouse EEG established topographic sensor-level mapping of large-scale cortical activity (Mégevand et al., 2008) but did not localize sources. Prior rodent ESI has used subject-specific MRI-derived models (Bae et al., 2015; Jiricek et al., 2021), which require individual neuroimaging and provide no common coordinate framework for cross-study comparison. Our atlas-based framework benchmarks 18 configurations and compares source- and electrode-level analysis in a disease-model dataset, trading marginal individual accuracy for standardization across laboratories.

4.8. Limitations

The simulation benchmark evaluates mathematical recovery of single dipoles under known conditions (Michel et al., 2004; Michel & Brunet, 2019) and does not exercise per-subject electrode misregistration, in-vivo conductivity deviation, or head-model mismatch. Correlated and simultaneous sources were tested (§3.1.6) and are a stated limitation rather than an untested assumption: two co-active parcels are both recovered in 13–14% of trials and three never are. Empirical validation with simultaneous surface and depth recordings is still needed. The findings are also conditional on the imaging configuration: across the 18 configurations of §3.6 the gamma effects were stable, but the aperiodic Group × ROI interactions divided by head-model geometry (ellipsoidal 8 of 9, spherical 3 of 9), so regionally graded flattening is conditional on ellipsoidal BEM geometry in a way the gamma claims are not.

Biological findings come from a convenience sample of male, adolescent ICV-vehicle-treated mice. At n = 18 KO / n = 17 WT the 95% CI on subject-level Hedges’ g spans roughly ±0.7 (e.g., Low Gamma absolute 1.28 [0.55–2.01]); null or modest-effect ROI findings are not evidence of absence. Generalization to injection-naive, female, or differently-aged Fmr1 KO animals, and translation to human FXS, requires separate evidence; the cross-sectional design supports group associations rather than causal inference.

4.9. Toward validation criteria for mouse ESI

Rodent source imaging has no agreed validation standard, and in its absence a single accuracy number tends to be read as a general guarantee of validity. We suggest that a mouse-ESI pipeline report its measured status in each of the four domains: numerical recovery, whether the inverse operator returns a dipole the forward model placed (§3.1.1); atlas-label attribution, whether that activity reaches the correct parcel (§3.1.3, §3.1.5); inferential sensitivity, whether a group conclusion survives the pipeline’s analytic choices (§3.6); and empirical validation against intracranial ground truth, which this study does not provide. The criteria concern reporting rather than thresholds: the field does not yet know what localization error suffices in the mouse.

4.10. Future directions

ESI enables ROI-level functional connectivity (Schoffelen & Gross, 2009), whole-brain vertex mapping, and cross-modal comparison with Allen-aligned fMRI. Biomarker development from any region-specific measure would require, at minimum, test-retest reliability, effect-size stability across recording systems, replication in independent cohorts, and responsiveness to therapeutic perturbation. This is a high bar, given that clinically probed mechanisms of action have so far failed to normalize translational EEG phenotypes in Fmr1 knockout mice (Janz et al., 2025). In the interim, the framework is positioned as an analytic tool for preclinical pharmacological and gene therapy studies.

5. Conclusions

We report a systematically benchmarked atlas-based ESI framework for mouse EEG, pairing the Antwerp Mouse Brain Atlas (coordinate space, BEM geometry) with the Allen-32 parcellation (ROI assignments). In an 18-configuration simulation benchmark, sLORETA + ellipsoid BEM achieved the lowest depth-graded localization error among ROI-anchored configurations, and the selected configuration held 1–4 mm error at or below 2.16 mm across an 8-fold band of assumed conductivity; the ROI-based source space was selected under a prespecified priority of deterministic atlas labeling, accepting 0.386 mm of mean localization error over 1–4 mm (95% CI 0.267–0.512) relative to Cartesian sLORETA; gridded spaces remain preferable for discovery-oriented mapping, where the question is where activity is rather than how much of it a named parcel carries.

Applied to Fmr1 KO (n = 18) vs. WT (n = 17) mice, the framework recovered elevated gamma-band absolute power at both source-level mixed-model omnibus tiers (32-ROI q = 0.004, 10-region q = 0.006), bilaterally distributed. The loci with the lowest expected peak displacement sit in dorsal cortex (Somatosensory_R and bilateral Frontal-Anterior); additional deep loci carry larger effect sizes but substantially higher attribution uncertainty, graded by structure: hippocampal effects localize to a neighborhood and thalamic effects only weakly. Olfactory effects are well localized but cannot be assigned to a hemisphere (Supplementary Table S10). The framework also resolved spatially heterogeneous Delta relative-power effects that the electrode analysis did not detect, and localized the gamma elevations to six named parcels, though no individual parcel survived correction for Delta. Aperiodic interactions were detected at both the source (32-ROI) and electrode (30-channel) tiers, with the source analysis detecting the gamma effect at more spatial units than the matched electrode tier, and adding anatomical attribution that is reliable for superficial cortical parcels and not for deep ones (§3.1.5). Across 18 modeling configurations the gamma effects were stable and the regional aperiodic interactions were not. Replication across laboratories, recording systems, sex, age and injection-naive animals will be needed to establish generalizability.

Supplementary Material

Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.neuroimage.2026.122218.

Data and software availability

Software. Both pipelines are publicly available at https://github.com/alexedmon1/source-localization and https://github.com/alexedmon1/source-analytics. Source reconstruction and all validation analyses were produced with source-localization v0.4.1; all statistical analyses with source-analytics v0.4.0, which pins the 2–50 Hz aperiodic fitting window. Later releases resolve that window from configuration, so reproducing the aperiodic values requires the pinned tag. Each release records its commit hash and the environment it was built against.

Configurations and derived data. The pipeline presets and the analysis code are in those releases, with the study configuration. Every derived table behind the figures and statistics reported here is in the Supplementary Material. The per-subject tables that feed the mixed models and the trial-level simulation outputs are available from the corresponding author on request.

Raw EEG recordings. Raw EEG recordings are subject to the data-sharing terms of the parent gene therapy study and are available from the corresponding author on reasonable request, subject to those terms.