LEARNING FROM DATA

  • The elusive search for genes and variants that increase fluid intelligence

    Positive genetic associations offer clues, but the route from a higher-score estimate to a convincing biological mechanism is difficult. MYEF2 shows both the promise and the uncertainty.

    Conceptual illustration of integrating variant annotations
    Conceptual illustration of integrating constraint, missense pathogenicity and loss-of-function annotations. The ribbons are artistic motifs, not measured protein structures or statistical results.

    Can human genetic variation identify biological changes associated with higher fluid-intelligence scores? Rare coding variants offer a way to investigate the question: they alter proteins through specific amino-acid substitutions, splice changes and truncations. Their effects can point in different directions within the same gene. A positive estimate becomes more persuasive when it survives scrutiny of consequence class, constraint, allele frequency, uncertainty and replication.

    We analyzed Genebass summary statistics for fluid intelligence in 128,302 UK Biobank participants. The phenotype is the inverse-rank-normalized fluid-intelligence score, UK Biobank field 20016. We asked a focused question: within each gene, do variant effects differ systematically with HMM constraint, AlphaMissense pathogenicity, and high-confidence loss-of-function status?

    The frozen analysis contains 4,977,720 variant–gene records from 19,287 retrieved genes; 19,215 genes support a testable joint baseline model. The archived coverage table identifies every unavailable target. In the primary analysis, 2 genes cross the planned threshold of p < 2.58 × 10-6, corresponding to 0.05 divided by all 19,373 target genes. The genes crossing this threshold are ANKRD12, KDM5B.

    Joint annotation-model scan of fluid intelligence across 19,287 retrieved genes
    Primary per-gene annotation F-test versus an intercept-only common effect, AF≤0.05. The dashed line is 0.05/19,373; all planned genes remain in this denominator. Points are genes. FWER primary hits are purple diamonds; 5% BH-FDR secondary hits are lime outlined diamonds; nominal MYEF2 is an orange star. BH FDR is computed within the 19,215 testable baseline joint-model genes. BH FDR uses only the baseline joint-model testable-gene family. Available-family BH classifications are conditional on the available source coverage. The QQ plot does not account for unavailable LD or carrier covariance.

    A model for separating direction from annotation patterns

    The analysis adapts the unified meta-regression framework developed by Lauer and Rivas and applied to rare-variant studies by Aguilar, Rivas, and Rivas. The present specification uses three moderators and explicit annotation imputation; it is an adaptation rather than a literal replication of either preprint’s full model. Each observation is a single-variant association estimate, not an individual participant. Within gene g, we fit:

        \[\begin{aligned}\widehat\beta_{ig}&=\alpha_g+\gamma_{H,g}H_{ig}+\gamma_{A,g}A_{ig}+\gamma_{L,g}L_{ig}+\varepsilon_{ig},\\ w_{ig}&=\widehat s_{ig}^{-2},\qquad\operatorname{Var}(\varepsilon_{ig})=\sigma_g^2\widehat s_{ig}^{\,2}.\end{aligned}\]

    NotationMeaning
    i,\ gVariant index and gene index. The observations are variant association estimates.
    \widehat\beta_{ig},\ p_{ig},\ \widehat s_{ig}Genebass variant effect, reported two-sided association p-value, and reconstructed standard error for the inverse-rank-normalized score.
    H_{ig},\ A_{ig},\ L_{ig}Transformed HMM constraint and AlphaMissense scores, and observed HC-LoF indicator (or its explicitly imputed proportion).
    \alpha_g,\ \gamma_{H,g},\ \gamma_{A,g},\ \gamma_{L,g}Gene intercept and adjusted annotation slopes. The intercept refers to zero-valued predictors; it is not a missense-class coefficient.
    \varepsilon_{ig},\ \sigma_g^2,\ w_{ig}Residual error, gene-specific residual variance scale, and inverse-variance weight.
    n_g,\ r_gUsable variant count and rank of the full design matrix. Residual degrees of freedom are n_g minus r_g.
    \mathrm{RSS}_{0g},\ \mathrm{RSS}_{1g},\ F_gWeighted residual sums of squares for the intercept-only and annotation models, and their rank-aware F statistic.
    \Phi^{-1},\ \mathrm{AF}Standard-normal quantile function and source phenotype-specific allele frequency. AF cutoffs select nested sensitivity sets.

    HMM summarizes regional constraint; AlphaMissense supplies a predicted missense pathogenicity score; and the loss-of-function term distinguishes variants labeled high confidence by LOFTEE. We clip HMM and AlphaMissense probabilities to [0.01, 0.99] and transform them as −log(1 − score). The fit gives more weight to effects estimated with greater precision.

    The primary F-test compares this model with an intercept-only model. Its null is H_{0,g}:\ \gamma_{H,g}=\gamma_{A,g}=\gamma_{L,g}=0, for the annotation contrasts that can be estimated. A small p-value means that annotations help explain differences among variant effects. The fitted coefficients and observed effect signs supply the directional evidence.

        \[F_g=\frac{(\mathrm{RSS}_{0g}-\mathrm{RSS}_{1g})/(r_g-1)}{\mathrm{RSS}_{1g}/(n_g-r_g)}\ \overset{H_0}{\sim}\ F_{r_g-1,\,n_g-r_g}.\]

    We estimate the residual scale and use analytic, rank-aware F- and t-tests. When an annotation is constant within a gene, its separate coefficient cannot be identified; the joint test uses the remaining independent contrasts. The detailed results include the three single-annotation models and confidence intervals for estimable coefficients.

    The three-moderator meta-regression model and imputation procedure
    The actual three-moderator model used in this scan, adapted from Lauer and Rivas and Aguilar, Rivas and Rivas. SE recovery is an approximate normal inversion of beta and the reported two-sided p-value. Missing annotations are mean-imputed; source effects are never imputed. The rank-aware residual-scaled F-test compares the full annotation model to an intercept-only model, not a zero-effect burden null.

    The clearer evidence points toward lower scores

    ANKRD12, already implicated in adult cognitive function by Chen et al. (2023), has a joint-model result of F(3, 886) = 17.45, p = 5.14 × 10-11 across 890 baseline variants. It meets the primary planned-familywise criterion. The source Genebass HC LoF burden effect is negative, indicating association with lower scores (p = 5.28 × 10-14). Source-matched annotations identify 53 observed HC LoF variants, with 13 positive and 40 negative estimates; 0 LoF values remain imputed. The varying observed LoF indicator supplies information that constant whole-gene imputation cannot create. This is an annotation-based reanalysis of UK Biobank data, with cohort overlap rather than independent replication.

    KDM5B encodes an H3K4 histone demethylase, an enzyme that removes methyl marks from histone H3 (Yamane et al., 2007). Its relationship to adult cognition is already established: Chen et al. (2023) associated rare protein-truncating variants with lower cognitive performance, including verbal-numerical reasoning, in UK Biobank. The present work revisits UK Biobank summary statistics through an annotation model. It adds a view of differences among variant effects; it does not establish a new cognition gene or provide replication in an independent cohort.

    For KDM5B, the joint model gives F(3, 626) = 10.81, p = 6.24 × 10-7; the Bonferroni-adjusted value across all planned genes is 0.01208. This crosses the planned exome-wide primary threshold. Its source Genebass high-confidence loss-of-function burden effect is negative (p = 1.38 × 10-6), indicating an association with lower fluid-intelligence scores.

    The adjusted loss-of-function contrast in the KDM5B meta-regression is -0.538 (95% CI -0.752 to -0.323). This coefficient compares high-confidence loss-of-function variants with the model’s non-LoF reference at the same HMM and AlphaMissense values. It is a different quantity from the published burden effect, so the two magnitudes should not be interpreted interchangeably.

    MYEF2: following the positive missense clue

    MYEF2 illustrates why both direction and the statistical question matter. Its published Genebass loss-of-function burden association is negative (p = 0.001207), while its missense/low-confidence-LoF burden association is positive (p = 0.003002). Both source burden p-values are below 0.01. The annotation model, using 220 variants, gives F(3, 216) = 8.032 and p = 4.25 × 10-5. This does not cross the planned exome-wide primary threshold or the secondary BH-FDR criterion; we show it as a nominal association and an illustrative case study.

    To examine the positive missense clue directly, we fit a focused consequence-class meta-regression in MYEF2, keeping all 220 baseline variants and their QC flags. The 159 strict missense variants have a positive class coefficient of 0.0662 (95% CI 0.0162 to 0.116; nominal p = 0.00966). The observed HC-LoF class coefficient is -0.0957. Their missense-minus-LoF contrast is 0.162 (95% CI 0.0834 to 0.24; p = 6.67e-05).

    This is a secondary analysis prompted by the MYEF2 case study. Each class coefficient is a precision-weighted mean of single-variant effects. The model has one coefficient per consequence class and no additional intercept. Its class-mean tests concern zero, while the exome-wide annotation F-test concerns differences explained by annotations.

        \[\begin{aligned}\widehat\beta_{ig}&=\sum_{c\in\mathcal C}\theta_{c,g}\,\mathbf 1(C_{ig}=c)+\varepsilon_{ig},\\ \widehat\beta_{ig}&=\sum_{c\in\mathcal C}\theta^{\mathrm{adj}}_{c,g}\,\mathbf 1(C_{ig}=c)+\eta_{H,g}(H_{ig}-\bar H_g)+\eta_{A,g}(A_{ig}-\bar A_g)+\varepsilon_{ig},\\ \Delta_{M,L}&=\theta_{M,g}-\theta_{L,g}.\end{aligned}\]

    Here, Cig is the verified consequence class, and the class set contains strict missense (M), synonymous, observed HC LoF (L), and other coding consequences. The indicator is one for membership in a class and zero otherwise. The coefficients θ summarize class means; θadj describes means at common annotation reference values. The barred H and A are the frozen gene baseline averages, η are the additional annotation slopes, and ΔM,L is the missense-minus-HC-LoF contrast. Both models use the same reconstructed uncertainties and residual-scaled t intervals as the primary framework.

    Consequence classVariantsMean effect95% CINominal pHolm p (12 tests)
    Strict missense1590.06620.0162 to 0.1160.009660.058
    Synonymous430.0454-0.0704 to 0.1610.441
    Observed HC LoF12-0.0957-0.156 to -0.03520.002060.0207
    Other coding consequences60.7770.214 to 1.340.007090.0496

    The increasing-effect interpretation remains tentative. The missense coefficient has Holm-adjusted p = 0.058 across all 12 reported baseline class-mean and contrast tests. At common HMM and AlphaMissense values, its adjusted mean is 0.0489 (95% CI -0.0215 to 0.119; p = 0.172). The adjusted missense-minus-LoF contrast remains positive (0.144; p = 0.00232), while the adjusted missense-minus-synonymous contrast is -0.00265 (p = 0.969). Thus the positive mean gives limited evidence for a missense-specific increasing effect relative to synonymous variation.

    The AlphaMissense moderator describes a separate severity gradient. Its single-annotation slope is -0.0824 (p = 0.00816), and its slope in the joint HMM–AlphaMissense–LoF model is -0.0291 (p = 0.388). These negative slopes and their adjustment dependence must accompany the positive missense-class mean. The published positive missense/LC burden uses a broader consequence mask and a different aggregate estimand. None of these coefficients estimates the cognitive effect of editing a variant in a person.

    MYEF2 consequence-class means, contrasts, severity slopes, and allele-frequency sensitivities
    MYEF2 focused secondary consequence-class moderator analysis. A: all four class means versus zero. B: missense-minus-LoF and missense-minus-synonymous contrasts, before and after adjustment for centered HMM/AlphaMissense. C: the original AlphaMissense-only and joint severity slopes. D: missense and HC-LoF class means across nested AF subsets. Nominal p-values and residual-scaled 95% t intervals are shown; 12-test within-AF Holm values accompany the tables. All source variants and QC flags remain included. Positive means do not establish cognitive enhancement.

    Rodent studies connect Myef2 to myelin-gene regulation through DNA-binding repression of Mbp and RNA interactions with Mbp and Mag in oligodendrocytes (Haas et al., 1995; Wang et al., 2021). Whether these mechanisms explain the human association remains unestablished.

    Individual loss-of-function estimates need not all share the aggregate sign. Among the observed high-confidence loss-of-function variants, KDM5B has 14 positive and 40 negative estimates; MYEF2 has 7 positive and 5 negative estimates. Precision, allele frequency, and weighting matter. We therefore show individual effects alongside the source burden direction rather than assigning one sign to every variant in a gene.

    The descriptive high-confidence-LoF average for MYEF2 is especially concentrated: one negative splice-donor variant, 15-48152233-C-A, contributes 98.37% of the inverse-variance information among the 12 observed HC variants. It carries a retained interval-QC flag and imputed HMM and AlphaMissense values, and has no defensible exact residue placement. Its AF is 0.00472, so it is absent from both rarer subsets. This fraction describes our single-variant IVW summary; it does not apportion the published Genebass burden test or the joint F-test.

    Loss-of-function effect directions, source burden tests, and adjusted meta-regression coefficients
    A: published Genebass pLoF burden beta and p-value, with no reconstructed confidence interval. B: HC LoF coefficient adjusted for HMM and AlphaMissense, with the model residual-scaled 95% t interval. A and B are distinct scales, variant masks and null hypotheses. C: signs of observed individual HC LoF variant effects; counts do not reflect precision. Known annotations only are used for counts; imputed fractional LoF values do not become known variants. Rows separate FWER primary, 5% baseline joint-model BH-FDR secondary, and nominal illustrative genes. BH FDR uses only the baseline joint-model testable-gene family. Available-family BH classifications are conditional on the available source coverage.
    GeneVariantsJoint F-test pPlanned Bonferroni pSource LoF burden directionInterpretation
    ANKRD128905.14 × 10-119.96 × 10-7Lower scorePrimary threshold passed
    KDM5B6306.24 × 10-70.0121Lower scorePrimary threshold passed
    MYEF22204.25 × 10-50.822Lower scoreNominal / illustrative

    Baseline AF ≤ 0.05. Correction covers the planned primary family of 19,373 genes. Individual moderator and sensitivity tests are not additional independent discoveries.

    A secondary view at 5% false discovery rate

    Benjamini–Hochberg q-values use the 19,215 retrieved testable genes in the joint-model, AF ≤ 0.05 family. These secondary classifications are conditional on the available source data; 86 target genes remain untested. The primary Bonferroni denominator retains all planned genes. At q ≤ 0.05, 3 genes meet the secondary criterion; 2 of these also pass the primary planned-familywise threshold, while 1 is an additional FDR candidate. The table below includes the full secondary set and labels the overlap explicitly.

    GeneVariantsJoint F-test pBH qPlanned Bonferroni pSource LoF burden directionEvidence
    ANKRD128905.14 × 10-119.88 × 10-79.96 × 10-7Lower scorePrimary + secondary
    KDM5B6306.24 × 10-70.005990.0121Lower scorePrimary + secondary
    FBRSL19223.91 × 10-60.02510.0758Higher scoreSecondary FDR only

    FBRSL1 illustrates the secondary set. Its source Genebass LoF burden sign is positive (p = 0.0246), and its adjusted LoF contrast is 0.0369 (95% CI 0.019 to 0.0547). Among 31 observed HC LoF variants, 24 estimates are positive and 7 negative. It does not meet the primary planned-familywise criterion (Bonferroni p = 0.0758). The joint-test evidence weakens from p = 3.91 × 10-6 at AF ≤ 0.05 to 0.0314 at AF < 0.001 and 0.0263 at AF < 0.0001. This candidate therefore depends substantially on the broader baseline variant set.

    Adjusted HMM, AlphaMissense, and loss-of-function moderator coefficients
    Joint-model HMM, AlphaMissense and HC LoF coefficients with residual-scaled 95% t confidence intervals. All coefficients adjust for the other two moderators and an intercept. Predictor units and axis ranges differ; coefficient magnitudes are not importance scores. Rows are selected by baseline joint-model FWER or BH-FDR status plus requested illustrative anchors, never by single-moderator p-values. BH FDR uses only the baseline joint-model testable-gene family. Available-family BH classifications are conditional on the available source coverage.

    The primary Bonferroni criterion and the secondary BH-FDR criterion serve different purposes. The first is our stricter primary discovery threshold across all planned genes. The second identifies a broader set for exploration under the BH procedure and the model’s calibration assumptions. A secondary FDR candidate is labeled as such throughout the figures and structural views. HMM-only, AlphaMissense-only, LoF-only, and rarer-AF analyses are separate test families and do not enter this secondary set.

    Testing whether candidate genes converge on pathways

    We selected all 48 genes with raw joint-model p < 0.001 at AF ≤ 0.05, then tested Reactome release 98 human pathways. 30 selected genes map to Reactome. The background is the 11,259 testable genes represented in that same mapping, drawn from 19,215 testable genes overall. We tested 1,753 pathways containing 10–500 background genes and included zero-overlap pathways in the multiple-testing family. The F-test is unsigned, so this gene set mixes higher- and lower-score associations.

        \[X_P\sim\operatorname{Hypergeom}(N_B,K_P,N_F),\qquad p_P=\Pr(X_P\ge x_P)=\sum_{j=x_P}^{\min(K_P,N_F)}\frac{\binom{K_P}{j}\binom{N_B-K_P}{N_F-j}}{\binom{N_B}{N_F}}.\]

    For pathway P, NB is the mapped background size, NF the mapped selected-gene count, KP the pathway’s background-gene count, and xP the observed selected-gene overlap. XP is the random overlap under sampling from that background; j indexes possible overlaps, and the binomial coefficients count combinations. The upper-tail probability pP tests over-representation. We apply Benjamini–Hochberg correction across all eligible pathways to obtain qP. Fold enrichment is the observed selected-gene pathway fraction divided by the background fraction.

    No pathway meets BH q ≤ 0.05. The leading nominal pathway is Presynaptic depolarization and calcium channel opening, with 2 selected genes among 11 background members: CACNA1A, CACNA1E. Its nominal p is 3.72e-04, and its BH q-value is 0.652. The corrected result determines the strength of this pathway evidence.

    18 selected genes lack a Reactome mapping in this release: RC3H2, DNAAF6, PUM3, ANKRD12, MYEF2, PRDM13, FBRSL1, SLC35A5, TBX19, LNPK, MAP11, GIGYF1, CIART, CCDC112, NLRP7, EDDM3A, CGB7, RNF148. They remain in the foreground file and gene-level analysis. The enrichment uses the mapped subset. Pathway overlap, annotation coverage, gene size and unequal detection power further limit a thresholded enrichment analysis.

    Several leading nominal pathways repeatedly contain the same genes. These shared overlaps generate related pathway labels and should be read as correlated annotations of the gene list.

    Reactome pathwayOverlapFold enrichmentNominal pBH qSelected genes
    Presynaptic depolarization and calcium channel opening2/1168.23.72e-040.652CACNA1A, CACNA1E
    Transport of Ribonucleoproteins into the Host Nucleus2/3223.50.003241KPNB1, NUP88
    Nuclear import of Rev protein2/3422.10.003651KPNB1, NUP88
    Host Interactions of HIV factors3/1189.540.003711PSMC5, KPNB1, NUP88
    Interactions of Rev with host cellular proteins2/3720.30.004311KPNB1, NUP88
    NS1 Mediated Effects on Host Pathways2/4317.50.005791KPNB1, NUP88
    Cell surface interactions at the vascular wall3/1417.990.00611SLC7A8, CXADR, CD47
    Apoptotic execution phase2/5214.40.008381KPNB1, PLEC
    Apoptosis3/1656.820.009391PSMC5, KPNB1, PLEC
    Disorders of transmembrane transporters3/1656.820.009391SLC12A3, PSMC5, NUP88
    Reactome enrichment of 48 genes with joint-model p below 0.001; no pathway passes FDR correction
    Reactome human pathway over-representation for genes with joint baseline raw p<0.001. The universe comprises all testable joint-model genes mapped to Reactome; pathways with 10–500 background genes are tested. BH correction covers the entire eligible family, including zero overlaps. The 12 leading nominal pathways are shown with corrected q-values, overlap counts and fold enrichment. The gene set is unsigned and does not identify intelligence-increasing pathways.

    Putting variants back on the protein

    The interactive viewers are self-contained HTML downloads. Download a viewer from Dropbox and open it in a browser to rotate the reference structure and inspect the mapped variants.

    The 3D views connect association statistics with physical positions on the protein. Missense variants are placed at their mapped amino-acid residues in authentic AlphaFold reference models, with sequence and isoform identity checked before mapping. Color shows the sign of the variant’s fluid-intelligence association; separate symbols and controls distinguish consequence class, nominal variant evidence, annotation values, and missing annotations.

    We include the primary set, the secondary FDR set, and the requested MYEF2 example, with the evidence category stated for each view. Low local confidence or uncertain positioning between domains limits structural interpretation; these displays do not establish a confidently folded global structure or a functional pocket.

    For KDM5B, the main view uses the 1,580-residue UniProt isoform Q9UGL1-2, matching the source protein ENSP00000356233.2 and transcript ENST00000367264.7. 613 of 630 baseline variants map after exact reference-amino-acid checks. A validated sequence crosswalk supports the separate 1,544-residue canonical view.

    For MYEF2, 216 of 220 variants map to the Q9P2K5 reference. Three splice variants and R426Q remain unmapped: the current model has Q426 where the source substitution expects R. These variants retain their association statistics and annotations in the tables.

    FBRSL1 uses the source-matched 1,045-residue Q9HCM7 model; 913/922 baseline variants map. Its actual model confidence is low overall: mean C-alpha pLDDT 45.6, with 823/1045 residues (78.8%) below 50. We therefore use it as a sequence-anchored location display and do not infer a confident global fold or functional pocket.

    ANKRD12 uses the source-matched Q6UB98 reference with 2,062 residues; 885/890 baseline variants map. Mean C-alpha pLDDT is 41.15, with 77.1% of residues below 50. Five splice-only variants remain unmapped. This low-confidence model supports a reference-location display, not a confident global fold or functional pocket.

    In these views, a positive effect means associated with a higher score in this dataset. It does not establish a causal gain of function or an intelligence-enhancing mutation. AlphaMissense pathogenicity predicts the consequences of missense variation; it does not predict the direction of a fluid-intelligence association.

    ANKRD12: authentic reference coordinates and observed variant directions
    ANKRD12 (ENSG00000101745); Primary planned-familywise association. Available-family BH values are conditional on the available source coverage. Reference positions are descriptive, with local confidence and mapping limits; positive beta is higher-score association, not causal enhancement, a mutant fold, or a spatial association test. Source-matched Q6UB98 canonical isoform, AlphaFold v6, 2,062 residues; protein ENSP00000262126.3, transcript ENST00000262126.9. 885/890 baseline variants map and five splice-only variants remain unassigned. Mean C-alpha pLDDT 41.15; 77.1% of residues below 50. Global-fold and pocket inference are unsupported. Evidence label: Primary planned-familywise association; joint baseline p=5.143e-11, BH q=9.882e-07. BH q and evidence category are conditional on the available source coverage.

    Download ANKRD12 interactive 3D viewer (HTML)

    KDM5B: authentic reference coordinates and observed variant directions
    KDM5B (ENSG00000117139); Primary planned-familywise association. Available-family BH values are conditional on the available source coverage. Reference positions are descriptive, with local confidence and mapping limits; positive beta is higher-score association, not causal enhancement, a mutant fold, or a spatial association test. Main view: source-matched Q9UGL1-2, AlphaFold v6, 1,580 residues; 613/630 baseline variants map. Source protein ENSP00000356233.2, transcript ENST00000367264.7. The canonical 1,544-residue view uses an explicit validated sequence crosswalk. Evidence label: Primary planned-familywise association; joint baseline p=6.235e-07, BH q=0.00599. BH q and evidence category are conditional on the available source coverage.

    Download KDM5B interactive 3D viewer (HTML)

    FBRSL1: authentic reference coordinates and observed variant directions
    FBRSL1 (ENSG00000112787); Secondary BH-FDR candidate. Available-family BH values are conditional on the available source coverage. Reference positions are descriptive, with local confidence and mapping limits; positive beta is higher-score association, not causal enhancement, a mutant fold, or a spatial association test. Q9HCM7, AlphaFold v6, source ENST00000434748.2/ENSP00000396160.2; 913/922 baseline variants map. 7 HC splice variants and 2 alternative-protein-accession records remain unmapped. Actual PDB C-alpha mean pLDDT 45.6; 823/1045 residues (78.8%) are below 50. Global-fold and pocket interpretation are unsupported by this low-confidence model. Evidence label: Secondary BH-FDR candidate; primary planned-familywise threshold not passed; joint baseline p=3.912e-06, BH q=0.02506. BH q and evidence category are conditional on the available source coverage.

    Download FBRSL1 interactive 3D viewer (HTML)

    MYEF2: authentic reference coordinates and observed variant directions
    MYEF2 (ENSG00000104177); Nominal MYEF2 case study. Available-family BH values are conditional on the available source coverage. Reference positions are descriptive, with local confidence and mapping limits; positive beta is higher-score association, not causal enhancement, a mutant fold, or a spatial association test. Q9P2K5, AlphaFold v6; 216/220 variants map. Three splice variants and R426Q remain unmapped. The retained QC-flagged splice donor 15-48152233-C-A supplies 98.37% of descriptive observed-HC inverse-variance information; this is not its contribution to the source burden test. Evidence label: Requested nominal case study; primary and secondary criteria not passed; joint baseline p=4.246e-05, BH q=0.1243. BH q and evidence category are conditional on the available source coverage.

    Download MYEF2 interactive 3D viewer (HTML)

    Loss-of-function variants require a different visual interpretation. A stop-gain or frameshift can be anchored at its affected residue when the transcript mapping is unambiguous. That anchor marks the beginning of the change in the reference protein; it does not depict the folded structure of a truncated protein. Splice variants without a defensible residue mapping remain in the variant table. Model confidence and mapping coverage accompany every structural view.

    These maps are a way to inspect candidate mechanisms and plan experiments. Displaying several variants near one another does not constitute a statistical test of 3D clustering. The current association model uses annotations, not spatial proximity.

    Keeping missing annotations visible

    Missing annotations were imputed rather than used to discard a variant. Within each gene’s AF ≤ 0.05 baseline set, a missing transformed HMM or AlphaMissense value receives the mean of the observed transformed values. Missing loss-of-function status receives the observed high-confidence-LoF proportion. If a gene has no observed values for an annotation, we use the phenotype-wide observed mean. Known labels remain unchanged.

    Imputation preserves association records but cannot replace missing annotation information. If a gene has no observed LoF labels, its imputed LoF column is constant, so a separate LoF coefficient cannot be identified. Recovering original source annotations can restore a varying moderator while leaving the variant effects and their uncertainties unchanged. Annotation coverage is therefore part of the evidence behind each gene’s result.

    This procedure fills 601,693 HMM, 2,107,720 AlphaMissense, and 98,898 LoF entries in the baseline analysis. Imputed fractional LoF values are modeling values, not newly assigned biological LoF labels. Counts and direction summaries for observed LoF variants use known high-confidence labels only. The supplied HMM track covers autosomes; genes on X and Y remain included, with a constant imputed HMM term that cannot support a separate HMM coefficient.

    Interval and site QC flags do not exclude variants. Variant association p-values were not used to select the inputs, and no minimum allele-count filter was applied. The baseline contains 314,274 known interval-QC flags and 0 records with unknown interval-QC status. The downloadable tables and 3D tooltips retain these flags so that readers can inspect the evidence behind a result.

    How much depends on allele frequency?

    The primary analysis includes variants with AF ≤ 0.05. We also fit the same models at AF < 0.001 and AF < 0.0001, holding each gene’s baseline imputation means fixed. These are sensitivity analyses; the baseline and two rarer subsets are not three independent replications.

    For ANKRD12, the joint-test p-values are 5.14 × 10-11 at AF ≤ 0.05, 2.24 × 10-11 at AF < 0.001, and 1.34 × 10-10 at AF < 0.0001. For KDM5B, the joint-test p-values are 6.24 × 10-7 at AF ≤ 0.05, 2.59 × 10-6 at AF < 0.001, and 6.34 × 10-6 at AF < 0.0001. For MYEF2, the joint-test p-values are 4.25 × 10-5 at AF ≤ 0.05, 0.0875 at AF < 0.001, and 0.179 at AF < 0.0001. A change across frequency thresholds can reflect a small number of influential variants, reduced information, or a different set of functional changes. It is useful evidence for follow-up, not a reason to combine the thresholds into one stronger claim.

    Sensitivity of gene associations to nested allele-frequency thresholds
    Joint-model raw p-values and variant counts at three nested AF thresholds. Outlined cells pass the same planned 19,373-gene threshold at each AF cutoff. Row labels distinguish baseline FWER primary and 5% BH-FDR secondary genes; only the testable baseline joint-model family determines these labels. Rows include every baseline FWER/BH-FDR gene, KDM5B, MYEF2, and up to eight strongest baseline tests. BH FDR uses only the baseline joint-model testable-gene family. Available-family BH classifications are conditional on the available source coverage. These overlapping subsets are sensitivities, not independent replications.

    What would make an increasing-effect candidate convincing?

    The search yields an uneven picture. Stronger evidence connects loss of function in ANKRD12 and KDM5B with lower scores. MYEF2 provides a positive nominal missense-class mean and a clear contrast with its LoF class, but its adjustment and synonymous comparisons limit an increasing-effect interpretation. FBRSL1 provides a positive LoF-direction candidate whose evidence depends strongly on the broader AF set. These distinctions are central to the search for genes and variants associated with higher fluid intelligence.

    Several limits remain. The public single-variant endpoints do not supply standard errors, so we reconstruct them as |β̂| / Φ−1(1 − p/2), as requested. This is a normal-Wald reconstruction; it need not equal the original model-based uncertainty if the reported p-value came from a different calibration. The model treats variant errors as independent, because LD and carrier-overlap covariance are unavailable. It also treats imputed annotations as fixed and does not propagate their uncertainty. Retaining QC-flagged variants broadens inclusion and makes the recorded QC context particularly relevant when interpreting a candidate.

        \[\widehat s_{ig}\approx\frac{|\widehat\beta_{ig}|}{\Phi^{-1}(1-p_{ig}/2)}.\]

    These are exploratory analyses of an observational, population-based phenotype. A convincing increasing-effect candidate will require replication outside the discovery data, consistent directional evidence across appropriate sensitivity analyses, and functional experiments. Predicted pathogenicity alone cannot determine the direction of a fluid-intelligence effect. We make the statistics, annotation provenance and residue mappings available so those questions can be pursued directly.

    Explore the analysis: Catalogue of all planned genes (TSV.gz) · Saved per-gene model tests (TSV.gz) · Saved moderator coefficients (TSV.gz) · Unretrieved genes and coverage (TSV) · Code and provenance (ZIP) · Reactome enrichment table · MYEF2 consequence-class moderators.

    References

    1. Karczewski KJ, Solomonson M, Chao KR, et al. Systematic single-variant and gene-based association testing of thousands of phenotypes in 394,841 UK Biobank exomes. Cell Genomics. 2022;2:100168. doi:10.1016/j.xgen.2022.100168. Genebass is the source resource; the fluid-intelligence analysis here uses the phenotype-specific sample of 128,302 participants.
    2. Lauer L, Rivas MA. Unified meta regression models for rare variant association studies. bioRxiv. January 24, 2025. Preprint. doi:10.1101/2025.01.23.634522.
    3. Aguilar O, Rivas M, Rivas MA. A unified meta-regression model identifies genes associated with epilepsy. medRxiv. Version 3, October 23, 2025. Preprint. doi:10.1101/2024.06.27.24309590, v3.
    4. Rivas MA, Beaudoin M, Gardet A, et al. Deep resequencing of GWAS loci identifies independent rare variants associated with inflammatory bowel disease. Nature Genetics. 2011;43:1066–1073. doi:10.1038/ng.952. Visual palette adapted from colors sampled from the original figures; the effect-direction and annotation color assignments in this article are new.
    5. Cheng J, Novati G, Pan J, et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science. 2023;381:eadg7492. doi:10.1126/science.adg7492. AlphaMissense scores are from the public GRCh38 annotation release.
    6. Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596:583–589. doi:10.1038/s41586-021-03819-2.
    7. Bertoni D, Tsenkov M, Magana P, et al. AlphaFold Protein Structure Database 2025: a redesigned interface and updated structural coverage. Nucleic Acids Research. 54:D358–D362; published online November 22, 2025, January 2026 issue. doi:10.1093/nar/gkaf1226. Structural coordinates are from the database’s version-6 reference predictions; exact files and sequence checks are recorded with the viewers.
    8. Chen C-Y, Tian R, Ge T, et al. The impact of rare protein coding genetic variation on adult cognitive function. Nature Genetics. 2023;55:927–938. doi:10.1038/s41588-023-01398-8.
    9. Yamane K, Tateishi K, Klose RJ, et al. PLU-1 is an H3K4 demethylase involved in transcriptional repression and breast cancer cell proliferation. Molecular Cell. 2007;25:801–812. doi:10.1016/j.molcel.2007.03.001.
    10. Haas S, Steplewski A, Siracusa LD, et al. Identification of a sequence-specific single-stranded DNA binding protein that suppresses transcription of the mouse myelin basic protein gene. Journal of Biological Chemistry. 1995;270:12503–12510. doi:10.1074/jbc.270.21.12503.
    11. Wang Y, et al. PARP1-mediated PARylation activity is essential for oligodendroglial differentiation and CNS myelination. Cell Reports. 2021;37:109695. doi:10.1016/j.celrep.2021.109695.
    12. Reactome Pathway Database. Release 98, human Ensembl gene-to-pathway mapping at all hierarchy levels; retrieved November 9, 2026. Source data; over-representation analysis documentation. Exact source hashes, tested background and gene overlaps accompany this analysis.

    Analysis provenance: figures use the frozen analysis of 19,287 retrieved Genebass genes. The 86 source gaps remain documented, and available-family BH values are conditional on this coverage. Conceptual artwork is labeled separately from statistical and structure-based evidence.

  • Detecting localized genetic effects

    A variant protects against several coronary diagnoses. Another has a large effect on macular degeneration. A third changes metabolic disease risk in opposite directions. Which of these effects should we call localized?

    That question matters when we move from a genetic association to a biological hypothesis. A concentrated disease profile may point to a shared process across related diagnoses. A distributed profile can expose connections between systems, including tradeoffs that deserve attention in target discovery. Both are useful. We need to say precisely what the data support.

    In SPHERE, we represent phenotype association profiles as fields on a sphere and summarize their spatial structure with spherical harmonics. The maps make concentrated and broad patterns visible. We have now added a complementary comparison that works directly with complete effect estimates and standard errors for a fixed disease panel. It asks whether a disease family has an effect beyond a broad background, then asks how tightly we can bound effects outside that family.[1]

    The examples are the best way to see why these questions differ.

    Toy profiles showing broad, local, and broad plus local effects, with nested counts of 539 selected variants, 488 local scan passes, and one strict panel localized example.
    Figure 1. Three toy profiles across eight disease families and nested counts in the saved comparison. Of 539 selected candidates, 488 pass a local scan, including the one strict panel-localized example. A broad profile has moderate effects throughout. A localized profile has a strong effect in one family and small effects elsewhere. A broad-plus-local profile has a similar peak alongside substantial outside effects. These are simulated teaching examples, not observed variants. Purple marks the candidate family and teal the broad component.

    From a striking map to a complete phenotype profile

    Our archived additive catalogue contains 15,047 variants, 2,349 phenotypes and 35,119 reported variant–phenotype pairs. That is only 0.0994% of the possible matrix. A missing cell in this selected catalogue is an unreported estimate, not evidence of a zero effect.

    For the new comparison, we retrieved unthresholded effect estimates and standard errors for 539 selected variants across 32 binary endpoints. Four endpoints belong to each of eight families: coronary disease, inflammatory disease, airways, metabolic disease, neurological disease, digestive and biliary disease, eye disease, and cancer. Every candidate has all 32 estimates, giving 17,248 pairs. More than 12,000 of those estimates have original P values of at least 0.05.

    Completing the panel makes the outside-effects question possible. It does not erase candidate selection, cover all phenotypes, or turn this into a genome-wide discovery experiment.

    Two questions, two levels of evidence

    The local-increment test compares a broad phenotype pattern with a model that also allows a family-specific addition. For intuition, start with a constant background and add an effect to the four coronary endpoints. The actual model can accommodate signed broad patterns, so broad effects need not have the same direction in every family. An omnibus family test also permits the four inside endpoints to differ in direction.

    A passing local scan establishes a departure from the chosen broad model. It does not establish that outside effects are small. For a stricter label, we additionally choose a margin of log(1.10), approximately 0.0953 in log-odds units. At least one inside effect must be demonstrably larger than that margin, while all 28 outside effects must be bounded below it in magnitude. The bounds use simultaneous intervals across the 32 endpoints for a fixed variant.

    Thus “panel-localized” means evidence for a meaningful inside effect and sufficiently small measured outside effects, conditional on this panel, margin and uncertainty model. It does not mean biological effects exist nowhere else.

    MFGE8: a coronary pattern that is easy to recognize

    The MFGE8 insertion rs534125149 is a particularly compelling starting point. Previous work associated the insertion with protection against coronary atherosclerosis and related diagnoses. That clinical coherence gives the map a biological interpretation: several endpoints may reflect a shared vascular process.[2] The insertion is not established here as a loss-of-function allele, and a protective association alone does not establish the consequences of inhibiting the protein.

    In the saved 32-endpoint profile, the insertion has negative log-odds effects for coronary heart disease, myocardial infarction, angina and coronary revascularization. The corresponding odds ratios are approximately 0.78, 0.73, 0.77 and 0.70. These are related diagnoses moving in a consistent direction, rather than a peak driven by one isolated endpoint.

    SPHERE gives this variant a concentrated fitted field: effective area 0.130 and 10-degree cap energy enrichment 12.63-fold. The complete-panel comparison also identifies a coronary-family increment, with within-variant scan-adjusted P approximately 2.02 × 10⁻³⁷.

    The stricter outside-effects claim fails. The largest outside upper bound is 0.2985, above the 0.0953 margin. Part of the problem is uncertainty: the bronchiectasis estimate is −0.0492 with standard error 0.0788. There are also nonzero estimates outside the coronary family, including obesity at +0.1063. The correct conclusion is strong agreement on coronary concentration and a local increment. We cannot establish small effects across every measured outside endpoint.

    ARMS2: an eye peak that still does not prove exclusivity

    The ARMS2-annotated rs10490924 provides a second agreement example. It is the A69S variant at the well-known 10q26 age-related macular degeneration susceptibility locus.[3] This background makes the eye-family peak interpretable, while leaving regional fine-mapping and effector-gene questions open.

    In this panel, the macular-degeneration endpoint has log-odds effect +0.4188, an odds ratio of about 1.52. Cataract and retinal-detachment estimates are smaller, at +0.0715 and +0.0794. The four-endpoint omnibus therefore allows a much stronger macular signal alongside weaker or near-zero effects within the same family.

    The fitted spherical field is concentrated, with effective area 0.285 and cap enrichment 7.48-fold. The eye-family scan is exceptionally strong, approximately 3.87 × 10⁻²⁶⁴. Yet the outside upper bound is 0.1330, again above the strict margin. A highly significant inside association cannot compensate for insufficiently tight bounds elsewhere. This is why we retain the label “local increment.”

    MFGE8 coronary and ARMS2 eye endpoint forest plots showing original ALT log odds ratios, simultaneous intervals, odds ratios, and the small effect margin.
    Figure 2. MFGE8 coronary and ARMS2 eye-family estimates. Points show original ALT-allele log-odds estimates, with intervals using 3.162818 times the reported standard error, simultaneously across 32 endpoints for a fixed variant. The chosen magnitude margin is log(1.10). Both loci agree across methods on concentration and a family increment. Neither meets the strict outside bound. These are empirical saved-fit results.

    GCKR: one locus, several metabolic directions

    GCKR rs1260326 makes the broad-plus-local category biologically intuitive. Earlier studies of the Pro446Leu locus documented opposing glucose and triglyceride associations.[4] Those quantitative traits are external biological context, not endpoints used to produce our 32-disease classification.

    The saved results are oriented to the ALT C allele. Type 2 diabetes has a positive log-odds estimate of +0.0549, while lipoprotein disorders and hypercholesterolemia have estimates of −0.0560 and −0.0557. Obesity is close to zero. Averaging these four metabolic endpoints into one signed number would conceal much of the structure.

    The covariance-aware metabolic family mean is only 0.004, with standard error 0.005. The common signed-shift test has an unadjusted P value of about 0.045, because positive and negative effects largely cancel. The omnibus metabolic increment is strong, with scan-adjusted P approximately 7.51 × 10⁻³⁸. The distributed broad screen also passes, at P = 0.00870. Outside the metabolic family, the profile includes gallstone disease at +0.0673 and inflammatory bowel disease at −0.0663. SPHERE nevertheless classifies the fitted field as localized, with effective area 0.338.

    GCKR metabolic endpoint forest plot with opposing effects, alongside the PCLO annotated variant profile across 32 endpoints and its fitted broad component.
    Figure 3. GCKR and a broad-compatible comparator. The GCKR metabolic forest plot preserves opposing disease directions that nearly cancel in a family average. The PCLO-annotated variant shows all 32 original estimates and the fitted broad component. Its distributed screen passes without a passing local increment. Continuous glucose and triglyceride traits discussed in the text are external biological context, not inputs to these plots.

    This is an informative difference between the summaries. A concentration measure describes how fitted signal is distributed over the atlas. A distributed-evidence screen asks whether effects are supported across multiple outside families. A variant can satisfy both. For biological interpretation, we should keep the metabolic peak and the wider pattern in view.

    The PCLO-annotated rs2522833 supplies a useful broad-compatible comparison: its distributed screen passes at P = 0.000231, while its local scan does not, at P = 0.123. This describes supported distribution across measured families without a detected family increment under the chosen model. It is not a universal pleiotropy verdict.

    F5: a disagreement caused by the questions we did not measure

    Factor V Leiden, rs6025, is a useful stress test for any comparison of methods. Its association with activated protein C resistance provides a direct coagulation mechanism.[5]

    The spherical catalogue profile is localized and coagulation-centered. The 32-endpoint comparison is unresolved, with scan-adjusted P = 0.0553. It would be tempting to treat this as a failed replication of localization. But this fixed panel contains no coagulation family. It cannot directly interrogate the phenotypes that make F5 compelling.

    F5 therefore teaches a design lesson: a complete panel can still be incomplete for a particular biological question. The 32 estimates are all present, but the relevant outcomes are absent. A fairer follow-up would prespecify a broader panel including thrombotic and coagulation endpoints and then repeat the analysis.

    IL7R: broad geometry can coexist with a local contrast

    IL7R rs6897932 illustrates the reverse comparison. Its fitted field has effective area 0.901 and meets the spherical broad criteria. The variant also has established functional context: prior work linked it to IL7R splicing and multiple sclerosis.[6]

    In the complete panel, however, the winning contrast is a positive cancer-family increment relative to the fitted broad background, with scan-adjusted P approximately 4.06 × 10⁻¹¹. The saved broad model predicts a cancer-family baseline of −0.0347; the common signed addition is +0.0623. The raw cancer estimates are much smaller, between +0.0147 and +0.0374. Several immune and airway estimates are negative, whereas the cancer estimates are positive. Changing the broad basis could change which family wins this comparison. None of the cancer endpoints establishes an inside magnitude beyond the strict 10% margin under the simultaneous intervals. The outside bound also fails.

    The result is a local increment, not strict cancer localization. It does not establish a new cancer-specific mechanism for IL7R or make the prior immune biology irrelevant. It shows how the same profile can have broad spatial structure and a detectable relative departure in one measured family.

    POU5F1B: the strongest specificity example in this panel

    The POU5F1B-annotated rs6998061 at 8q24 is the only candidate that meets the strict panel-localization rule in the saved run. The locus has prior prostate and colorectal cancer association evidence, so this is not a claim to have discovered a new cancer locus.[7] The annotation also does not establish POU5F1B as the causal gene.

    The ALT A allele has log-odds estimates of −0.1532 for prostate cancer and −0.0958 for colorectal cancer, corresponding to odds ratios of approximately 0.86 and 0.91. Breast cancer is +0.0299 and non-small-cell lung cancer is −0.0372. These four cancers should not be collapsed into a claim of uniform protection. The omnibus family test preserves their different effects and yields scan-adjusted P approximately 7.83 × 10⁻³⁷.

    POU5F1B annotated cancer endpoint forest plot and all 28 outside upper effect bounds, each below the declared margin.
    Figure 4. The POU5F1B-annotated example. The four cancer effects are shown separately, alongside absolute-effect upper bounds for all 28 non-cancer endpoints. The inside lower bound is 0.1158, the margin is 0.0953 and the maximum outside upper bound is 0.0899. The claim concerns the 32 measured endpoints and ALT A orientation. It does not imply zero outside effects or universal cancer protection.

    What makes this example special is the outside evidence. The strongest inside lower magnitude bound is 0.1158, beyond the 0.0953 margin. The largest upper magnitude bound across all 28 non-cancer endpoints is only 0.0899, below that margin. The prostate association supplies the decisive inside bound. Every measured outside endpoint clears the corresponding small-effect requirement.

    SPHERE assigns this variant to “neither,” which includes support-filter exclusions and intermediate shapes. Only 10 catalogue endpoints are reported, below the geometric classifier’s minimum support of 15. Its fitted shape also does not meet the localization thresholds. This is therefore an informative difference in coverage and estimand, not a clean demonstration that one method has made a false-negative error.

    What did the complete comparison find?

    Among the 539 preselected candidates, 488 pass the local-increment scan. The exclusive labels are 473 local increments, 14 broad plus local, one panel-localized, four broad-compatible, five dispersed-hit profiles and 42 unresolved. Thus a local departure is common in this selected set, while the stronger outside-bounds requirement is demanding.

    An additional covariance-shrinkage sensitivity leaves 505 of 539 labels unchanged. That stability is useful, but it is not full validation. The reference correlation is an estimated proxy that may include genetic signal. Conditional simulations check the local scan with fixed known nuisance quantities, not the entire estimated-nuisance classification pipeline. The within-variant scan correction covers eight families and two test forms. It does not itself correct across the 539 selected candidates.

    Using the maps and the presentation together

    The presentation follows the same sequence: toy profiles, the measured panel, nested models, uncertainty bounds, then the empirical cases. For each variant, ask three questions. Which diagnoses drive the pattern? What does each method actually summarize? Which additional measurements would strengthen the biological interpretation?

    The website provides the maps, comparison tables and exports needed to inspect those answers. We see the greatest value in using spatial structure to organize hypotheses and complete effect profiles to test stronger statements about specificity. A convincing inside signal and informative outside bounds contribute different pieces of evidence. Keeping both visible makes the biology more useful.

    Explore the SPHERE method comparison: https://sphere-phenome-atlas.marivascruz.chatgpt.site/method-comparison/

    Open the companion presentation (ChatGPT access may be required)

    Methods and provenance

    All numerical case studies and counts above come from the saved SPHERE comparison summarized in the updated manuscript dated 8 November 2026. The teaching walkthrough uses a fixed broad basis for clarity. The empirical results retain the saved estimated basis and covariance proxy and have not been rerun under a replacement basis. The chromosome-exclusion choice used for nuisance estimation in that saved fit is not a definition of phenotype localization, independent replication, or a calibration guarantee. Simultaneous endpoint bounds apply to a fixed variant under working normal approximations. Prior candidate ascertainment and uncertainty in estimated nuisance quantities remain limitations.

    References

    [1] SPHERE preprint and the current comparison implementation. Original preprint: https://doi.org/10.1101/2025.10.27.684843 . The November 2026 comparison results discussed here are saved exploratory work associated with the current Site and updated manuscript, not a claim that the original preprint contained this complete analysis.

    [2] Ruotsalainen et al. Inframe insertion and splice site variants in MFGE8 associate with protection against coronary atherosclerosis. Communications Biology (2022). https://doi.org/10.1038/s42003-022-03552-0

    [3] Kanda et al. A variant of mitochondrial protein LOC387715/ARMS2, not HTRA1, is strongly associated with age-related macular degeneration. PNAS (2007). https://doi.org/10.1073/pnas.0703933104

    [4] Orho-Melander et al. Common Missense Variant in the Glucokinase Regulatory Protein Gene Is Associated With Increased Plasma Triglyceride and C-Reactive Protein but Lower Fasting Glucose Concentrations. Diabetes (2008). https://doi.org/10.2337/db08-0516

    [5] Bertina et al. Mutation in blood coagulation factor V associated with resistance to activated protein C. Nature (1994). https://doi.org/10.1038/369064a0

    [6] Gregory et al. Interleukin 7 receptor alpha chain (IL7R) shows allelic and functional association with multiple sclerosis. Nature Genetics (2007). https://doi.org/10.1038/ng2103

    [7] Cavazos et al. Assessment of genetic susceptibility to multiple primary cancers through whole-exome sequencing in two large multi-ancestry studies. BMC Medicine (2022). https://doi.org/10.1186/s12916-022-02535-6

  • The molecular architecture of epilepsy

    Using the latest Google Argon model we continued our analysis of the epi25.broadinstitute.org exome sequencing data.

    We posted our updated dissection of the molecular architecture of epilepsy here.

    Unified Ultra-Rare Variant Meta-Regression (AF < 0.0001) Integrating Deep-Learning Structural Pathogenicity and
    Base-Pair Population Constraint Uncovers Chromatin
    Remodeling, Ion Channel, GATOR1–mTORC1, and
    Synaptic Architectures Across 20,979 Epilepsy Cases and
    33,444 Controls.