The American Journal of Human Genetics

Allele frequency (AF) is among the most frequently applied lines of evidence in variant classification, yet the ACMG/AMP criteria that use it (BA1, BS1, PM2) are still applied at fixed defaults while computational predictors have been systematically recalibrated. Population frequencies are shaped by selection, ascertainment, and gene-level demography at once, and few genes carry enough classified variants to set a threshold directly. Extending the calibration approach applied to computational predictors, we used inheritance mode and gene-level missense constraint as stratification axes and pooled variants within each stratum. ClinVar missense variants annotated against gnomAD v4.1.1 were stratified along both, and gene-normalized kernel density estimates were fit to pathogenic and benign variants within a sliding window along the constraint axis. Thresholds were placed where the likelihood ratio crossed ACMG/AMP evidence strengths at a prior of 0.0441. Derived thresholds varied systematically with constraint and differed between inheritance modes, departing from the fixed defaults in both directions. On held-out genes, stratified cutoffs reached 96.7% accuracy against 90.1% unstratified. Restricted to the 73 ClinGen expert panel genes with autosomal dominant or recessive inheritance, the derived cutoffs reached 91.0% accuracy at 69.5% variant coverage, against 88.8% accuracy at 86.2% coverage for the panel-specified cutoffs. AF thresholds for these criteria are not constant across genes, and inheritance mode and missense constraint capture much of that variation. The resulting cutoffs are empirically derived, carry explicit uncertainty, and deploy as a lookup table across thousands of genes no expert panel currently covers.

Non-random participation in genetic studies can bias associations between genetic variants and outcomes. Existing methods to detect ascertainment bias often require individual-level data, thus limiting their broad applicability. Here, we introduce a summary-statistics-based method to detect and quantify ascertainment bias in large-scale genetic studies. Our method estimates a parameter,{theta} , which captures deviations in the mean polygenic score (PGS) of an ascertained sample relative to its expectation across non-ascertained or differentially ascertained references. We show through extensive simulations that our method is robust to population stratification and reference misspecification unlike naive mean PGS comparison. When applied to 21 traits across 11 large-scale biobanks, our method recapitulates known patterns of ascertainment and detects new evidence of ascertainment on genetic susceptibility to depression, height and blood pressure in many biobanks. Overall, our framework enables systematic assessment of ascertainment directly from summary statistics and provides a scalable tool for evaluating representativeness in large scale genetic studies.

The gene-sex association quality control (QC) metric was developed during the era of early small-sized genome-wide association studies and typically filter variants based on controls alone. While this practice had minimal impact in early studies, in contemporary large-scale settings they can introduce systematic bias by disproportionately discarding variants with pronounced sex differences in allele frequencies (AF), potentially removing real gene-sex interaction signals. To address this limitation, we introduce Sex-Prevalence Adjusted Allelic Difference Estimates (SPADE) metric, an X chromosome-inclusive QC framework that not only preserves potentially informative variants exhibiting gene-sex interaction but also enables a rapid and exploratory scan for such interaction. SPADE adjusts for sex-specific disease prevalence, maintains correct type I error control, and reduces the risk of falsely excluding variants compared to existing QC approaches. Extensive simulations across diverse disease architectures demonstrated that SPADE remained well calibrated, whereas the controls-only approach exhibited massively inflated type I error when variants were associated with disease. We further developed an open-source command-line software implementation to facilitate its application in large-scale genetic studies. Applying SPADE to an autism spectrum disorder (ASD) case-control cohort (6,873 cases and 8,981 controls) comprising of the Autism Speaks MSSNG, Simons Simplex Collection (SSC), and Simons Powering Autism Research (SPARK) datasets, we demonstrate that SPADE is well calibrated relative to a controls-only approach. Notably, the complementary exploratory gene-sex interaction scan implicates a sex antagonistic region encompassing RBMX2, SLC25A14, and BCORL1 (lead SNP rs150885581: A>G, p = 6.19 * 10 ^ -9), providing candidate genes for future functional investigation.

Motivation: Horizontal pleiotropy remains a major challenge for Mendelian randomization (MR), particularly as increasingly large genome-wide association studies identify many genetic instruments for complex traits. Existing pleiotropy-robust methods rely on assumptions regarding the proportion or distribution of valid instruments that may be violated when pleiotropy is widespread. Results: We developed CaPri, a Bayesian MR framework that jointly estimates causal and pleiotropic effects. CaPri allows a potentially arbitrary proportion of instruments to contribute to a shared directional pleiotropic component and estimates the posterior probability and magnitude of the causal effect, instrument-specific posterior inclusion probabilities for pleiotropy, and the shared pleiotropic effect. Across simulations, CaPri maintained appropriate false positive control and accurate causal effect estimation regardless of the proportion of pleiotropic instruments, while accurately identifying pleiotropic instruments and estimating their shared effect. Performance remained robust across varying causal and pleiotropic effect sizes and exposure-outcome confounding, although sufficiently large outcome GWAS sample sizes were important for stable inference and limited prior dependence. In real data analyses, CaPri reported low posterior probabilities of causality across negative controls and strong support for established positive control relationships. Application to alcohol consumption and coronary artery disease further demonstrated its ability to identify potentially meaningful pleiotropic instruments. CaPri provides a complementary framework for strengthening causal inference and triangulation across MR methods with different identifying assumptions. Availability and implementation: CaPri is available at https://github.com/OuRen0104/CaPri.

Genetic diagnosis requires attributing a patient's disease to a variant. We develop the Probability of Necessity (PN), a counterfactual estimand giving the probability that disease would not have occurred absent a germline variant, given both are observed; and we use this formalism to explore the logic and limitations of attributing a disease to genetic variants. We show PN dissociates from penetrance: identical penetrance yields different PN depending on baseline disease prevalence. Case selection biases PN estimation, favoring population-scale cohorts over ascertained case series. PN falls as competing-cause prevalence rises, formalizing why necessity differs from penetrance under causal heterogeneity. We introduce the Probability of Mediated Necessity (PMN), updating PN using mediator biomarkers proxying a variant's operative pathway. Using UK Biobank data, we estimate PN and PMN for rare LDLR variants in ischemic heart disease (IHD, ICD-10 I25), using hyperlipidemia (E78) as mediator, stratifying by ClinVar class and computational pathogenicity score. PN resolves heterogeneity within the variant of unknown significance (VUS) class beyond classification alone. Replicating in GBA1 carriers with Parkinson's disease, a mechanistically distinct, low-penetrance system, reproduces this ordering and shows high PN despite modest penetrance. These results support further development of causal necessity as a framework in genetic diagnostics with implications for policy and clinical decisions in both monogenic and common complex diseases.

Polygenic scores (PGS) are a primary output of large-scale genetic studies and are being deployed in clinical and non-clinical settings. However, current PGS assume simple additive models that ignore context-specific genetic effects, which likely reduce their accuracy and robustness. To address this, we developed PGSC, a PGS framework to incorporate locus-specific gene-context interaction effects (GxC). Simulations show PGSC is robust under the additive model and outperforms PGS in realistic settings. Using sex, age, and statin treatment status as contexts in UK Biobank, we find that PGSC outperforms PGS on average across 48 traits, with substantial improvement in some cases, such as GxSex for testosterone, GxAge for bilirubin, and GxStatins for LDL cholesterol. PGSC consistently outperforms a simple genome-wide GxC model, ampPGS, which only outperforms PGS when a context uniformly amplifies all genome-wide additive effects. Critically, PGSC improvements replicate across ancestries in the UK Biobank and in an external cohort, the Mount Sinai Million Health Discovery Program. Finally, we test robustness to log-scale phenotypes and find that ampPGS gains vanish, while the locus-specific GxC components in PGSC persist. Overall, PGSC is a simple, robust framework that demonstrates GxC effects can improve out-of-sample PGS prediction and is a step toward precision treatment.

As genomic sequencing evolves beyond rare disease diagnostics toward population screening and precision medicine, clinical variant interpretation is increasingly challenged by variants whose clinical consequences depend on biological context. Current frameworks, including the ACMG/AMP guidelines, generally assign a single classification to each variant regardless of inheritance state or genetic context, potentially failing to communicate context-dependent clinical consequences. Here, we address this issue using loss-of-function variants in LPL as a uniquely informative model system in which residual physiological LPL activity can be directly quantified in vivo. By systematically integrating published biallelic LPL genotypes, physiological measurements, functional studies, and clinical phenotypes, we identified a biologically meaningful transition at approximately 10% residual physiological LPL activity. Activity below this level was predominantly associated with classical childhood-onset familial chylomicronemia syndrome (FCS), whereas higher activity was associated with phenotypic attenuation and modifier-dependent clinical expression. Furthermore, heterozygous loss-of-function variants exhibited an estimated penetrance of 5-7% for severe hypertriglyceridemia. We therefore propose a context-dependent framework in which biallelic complete- or near-complete loss-of-function genotypes are interpreted as causative for FCS, whereas heterozygous variants are interpreted as predisposing to severe hypertriglyceridemia while retaining recognition of FCS carrier status. Together, our findings demonstrate that clinical variant interpretation should integrate available biological context--including, where relevant, allelic configuration, residual biological function, and penetrance--rather than rely on the intrinsic molecular consequence of the variant alone. More broadly, this framework provides a conceptual model for interpreting variants across the continuum from Mendelian disease to genetic predisposition in the era of precision medicine.

Evolutionary theory offers a unifying explanation for the genetics of complex traits and diseases by modeling how a trait and its genetic basis are maintained by evolutionary forces. A growing body of evidence suggests that linkage disequilibrium (LD) couples a variant's evolution to that of the genetic background it resides on. However, widely used single-variant theories omit this coupling between variants and the background, or hold the background fixed. We close this gap by developing a model of linkage LD evolution accounting for stabilizing selection, recombination, and genetic drift. The model deduces the genetic architecture of individual variants subject to the co-evolving genetic background. More efficient selection in larger populations is partially offset by slower decay of existing LD and increased generation of new LD. Conversely, noisier LD in smaller populations exposes the variants to episodes of stronger selection. We ask whether ancestry differences in the relationship between effect-size magnitude and minor allele frequency require different selection strengths or can arise from differences in demographic history alone. Stabilizing selection needs to be more than three times stronger in Africans than in Europeans under the constant-size demography, whereas an Out-of-Africa demography alone can explain most of the observed frequency-dependence gap. As this example indicates, selection strength is in general not identifiable from the frequency--effect-size relationship alone, because different combinations of selection and demographic history can produce the same pattern in data with further redundancy added by LD.

Polygenic scores (PGS) are increasingly common in human genetics research, yet their performance and portability vary substantially across traits and populations. We systematically evaluate prediction accuracy and cross-ancestry portability for 3,994 published PGSs across the phenome in the ancestrally diverse CCPM biobank. We identify widespread trait-specific and cross-trait associations, demonstrating both the utility and breadth of existing polygenic predictors. However, predictive effects frequently differ across ancestry groups, highlighting persistent challenges for portability that threaten equity in downstream applications. We further find that accuracy and portability depend strongly on whether the PGS is continuous or thresholded. Finally, through hierarchical mixed modeling, we identify score-development, genome features, and cohort characteristics associated with improved PGS accuracy and portability. These results highlight key factors supporting future development of robust and portable scores for research and clinical translation.

Background Short-read genome sequencing generates the signal necessary to detect balanced reciprocal translocations at single-nucleotide resolution, but standard structural variant callers represent them as breakend pairs. Structural variant callers produce thousands of breakend records per genome, the majority of which are false positives arising from segmental duplications, repetitive elements, and reference assembly artifacts. This specificity problem has prevented balanced translocation detection from entering routine genome sequencing workflows. We developed a highly specific pipeline to resolve balanced reciprocal translocations and applied it across five clinically distinct cohorts to characterize the prevalence and diagnostic contribution of balanced reciprocal translocations. Methods Short-read genome sequencing data from 16,131 individuals across five cohorts were analyzed: healthy participants from the 1000 Genomes Project (n=3,202), probands with structural birth defects and their family members (CHOP BDB cohort; n=2,875, including 1017 probands), probands enrolled in the GREGoR consortium (GREGoR cohort; n=3,574 probands), probands with undiagnosed disease and their family members enrolled in the Undiagnosed Diseases Network (UDN cohort; n=6,071 including 2,027 probands) and neonates who underwent rapid genome sequencing (CHOP Baby Eagle cohort; n=409). A computational pipeline was developed to characterize breakend events from standard structural variant callers to high-confidence balanced translocation calls by exploiting well-characterized properties of true balanced translocation signatures in short-read data. Translocation breakpoints were resolved to single-nucleotide precision and interrogated for disruption of known disease genes or potential regulatory consequence otherwise. A subset of the identified breakpoints was confirmed by Sanger sequencing. Results Balanced reciprocal translocations were identified in 0.196% of structural birth defect probands (2/1,017), 0.25% of the GREGoR probands (9/3,574), 0.39% of the UDN probands (8/2,027) and 0.24% of neonates who underwent rapid genome sequencing (1/409). Parental carrier frequency in the birth defects cohort was 0.269% (5/1,858), consistent with previously reported estimates in similarly ascertained families, and 0.098% (3/3,040) in the UDN cohort. All six translocations identified in the birth defects cohort were orthogonally validated using Sanger sequencing, confirming pipeline specificity. Approximately 31% of translocations across the three affected cohorts disrupted a known disease gene at the breakpoint, establishing or refining a molecular diagnosis in five cases previously unresolved. The single translocation identified in the newborn cohort disrupted no disease gene and was classified as a variant of uncertain significance.

Protein coding variants span the gamut of structural and functional effects on proteins, and they have similarly heterogeneous effects on traits. Their functional effects can be predicted and categorized, particularly using machine learning models, and integrating these predictors benefits gene discovery. However, it remains unclear how well known structural features and predicted functional scores explain the heterogeneous effects of protein coding variants on human complex traits. We introduce the Kernel Allelic Series Model (KASM), which integrates multiple functional annotations to form a kernel describing how allelic effects within a gene covary. We apply it to exome data from 26 traits and 786,871 individuals from UK Biobank and All of Us, integrating 35 annotations describing protein structure and predicting allelic effects. We estimate that rare missense heritability is at least 1.27{+/-}0.09% on average, 1.83{+/-}0.11% under our most flexible model, exceeding pLoF heritability (0.87{+/-}0.05%) and far exceeding missense burden heritability as previously estimated (0.54{+/-}0.04%). AlphaMissense is the single most enriched missense predictor, but a composite score doubles its enrichment. A series of structural features is enriched for heritability, from amino acid chemical features and structural motifs to residue contacts and protein-protein interaction interfaces. Examining pairs of variants, we find that most missense effects are modestly correlated, but an exception is when residues are close in 3D structure but distant in 1D sequence; such pairs also covary across species. Finally, we analyze individual genes and identify gene-specific enrichments, like a 95-fold enrichment among PIEZO1 hydrogen bond residues, that nominate functional hypotheses.

Mendelian randomization (MR) is widely used to infer causal relationships using genetic variants as instrumental variables, yet the selection of genetic instruments is not always given sufficient attention. Many MR studies rely on default linkage disequilibrium (LD) clumping parameters (r2 <0.001, 10,000 kb), as implemented in commonly used tools, without assessment of their suitability for specific exposures. We investigated whether this approach yields optimal instruments or whether a more pragmatic strategy yields stronger instruments. Using UK Biobank data, we examined three distinct exposure types-circulating amino acids, body mass index (BMI), and major depressive disorder (MDD). For each phenotype, we systematically varied LD clumping thresholds (r2 and genomic distance) and evaluated each instrument via both their average strength (F-statistic) and total strength (R2). Across all phenotypes, optimal instruments differed from default parameters and varied by exposure. For amino acids and BMI, more stringent LD thresholds (r2=0.00001) combined with larger clumping windows improved instrument strength, whereas for MDD, a highly polygenic, binary trait, smaller windows with stringent r2 maximized variance explained while maintaining F-statistics above the desired threshold (>10). Notably, increasing the number of SNPs did not consistently improve instrument quality, highlighting a trade-off between instrument strength and potential pleiotropy. We demonstrate that universal reliance on default LD clumping parameters can lead to suboptimal instruments. We propose a pragmatic framework for instrument selection based on empirical evaluation of strength metrics, improving the robustness and transparency of MR analyses across different exposure types.

Pathogenic variants in SGCA, encoding -sarcoglycan, cause an autosomal recessive limb-girdle muscular dystrophy, LGMDR3/2D, yet clinical interpretation of SGCA variants remains challenging due to the high prevalence of rare missense variants. -sarcoglycan is an essential component of the sarcoglycan complex at the muscle cell membrane, and pathogenic variants frequently impair its membrane localization. Here, we systematically assess the effects of all possible single-nucleotide variants across the SGCA coding sequence using a saturation mutagenesis-based experimental assay that quantifies -sarcoglycan surface expression. We generate a comprehensive functional atlas that distinguishes tolerated and damaging variants, aligning with independent genetic and clinical evidence, and reveals domain-specific properties of the cytoplasmic region, in which C-terminal truncating variants retain membrane localization, suggesting possible pathogenic mechanisms beyond impaired trafficking. This work provides a scalable functional framework to support genetic diagnosis and variant interpretation in sarcoglycanopathies. Graphical AbstractSchematic overview of saturation mutagenesis-based functional mapping of SGCA. O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=187 SRC="FIGDIR/small/740214v1_ufig1.gif" ALT="Figure 1"> View larger version (41K): org.highwire.dtl.DTLVardef@1dc3d7corg.highwire.dtl.DTLVardef@48a213org.highwire.dtl.DTLVardef@88bec7org.highwire.dtl.DTLVardef@1a51bea_HPS_FORMAT_FIGEXP M_FIG C_FIG

Background Gene-level constraint and conservation scores are widely used to prioritise disease genes and to benchmark genomic language models. Purifying selection acts weakly on phenotypes manifesting after reproduction, so these scores are expected to carry less information about late-onset disease. An onset-dependent gradient has been demonstrated for human allele frequencies but not for the gene-level scores used in practice. Results We built an age-of-onset axis for 2,885 Mendelian diseases from Orphanet and fitted cross-species conservation (241-way phyloP) and human-population constraint (shet) on the same 5,322 gene-disease pairs against matched controls, to test whether one decays faster with onset. Their onset slopes differed by 0.23 logit units (97.5% confidence interval -0.10 to 0.56), equivalent to 0.024 units of area under the receiver operating characteristic curve (AUROC) in a design resolving 0.049; the interval excludes constraint losing discriminative power faster than conservation by more than 0.011 AUROC units, widening to 0.033 when somatic associations are dropped. The usual explanation, that early-onset disease genes are developmental, pleiotropic and haploinsufficient by construction, did not account for the contrast; conditioning barely moved it. The answer did, however, depend on the constraint metric: three alternatives yielded larger contrasts excluding zero, though paired comparison could not distinguish the estimates. The metrics also diverged on recessive disease genes, with shet ranked below the gene pool (AUROC 0.487) and loss-of-function depletion above it (0.533), though, against matched controls, both ranked them below (0.457 and 0.450). Evolutionary scores improved a non-evolutionary baseline by 17 to 23%, with no onset gradient. Conclusions Conservation and constraint do not lose their value at measurably different rates as disease onset rises, and with only 283 late-onset Mendelian diseases annotated genome-wide, the question cannot be settled more finely. The gene-class explanation does not account for the contrast. Metrics that rank genes near-identically can still yield opposite onset-stratified answers, so a result reported by one metric carries a sensitivity that its confidence interval does not show. No constraint metric identifies recessive disease genes against matched controls, so a pipeline with recessive disease in scope needs a correction that switching metric does not supply.

Local genetic correlation is estimated as a ratio of a local genetic covariance to the geometric mean of two local heritabilities. LAVA, the standard implementation, returns that ratio only where both traits clear a univariate heritability test at P < 0.05, a rule its documentation presents as a stabiliser. We show that it is a selection event acting on the denominator, and characterise what it does to the estimates that survive. Under the noncentral Wishart model LAVA's own interval routine assumes, the gate-conditional mean of the estimator is affine in the true correlation, with slope and intercept available in closed form from quantities every run already reports, that is, the components retained, the two univariate P-values, and the sample-overlap correlation. Attenuation is therefore a property of the analysis, not of the method. Across six psychiatric trait pairs the measured slope runs from 0.579 to 0.923 and the closed form predicts each to within 0.007 with nothing fitted to the panel, while within a single pair it runs from 0.371 to 0.971, so no single factor can correct it. Overlap and the filter jointly manufacture correlation where the truth is zero, most where power is lowest. We invert the closed form into a gate-aware correction, which reduces mean absolute bias from 0.247 to 0.026 over 48 simulation cells, and pair it with a diagnostic that decides whether the correction is identified. Three pairs fail it, all three sharing one trait whose local heritabilities are indistinguishable from a complete local null.

Psoriasis is associated with increased cardiovascular morbidity, but the extent and structure of the shared genetic architecture remain incompletely understood. We integrated large-scale European-ancestry genome-wide summary statistics for psoriasis and seven cardiovascular diseases to quantify genetic overlap, model shared liability and prioritize candidate mechanisms. Psoriasis showed modest genetic correlations with all cardiovascular outcomes, but mixture modelling estimated that more than half of psoriasis-associated non-zero variants were shared with five cardiovascular diseases. Genomic structural equation modelling of a latent psoriasis-CVD factor identified 167 independent risk loci, including 50 not genome-wide significant in any component trait. Multimodal prioritization nominated 473 candidate genes and 90 approved compounds with therapeutic relevance, including shared targets involving PDE4B, GLP1R, KEAP1 and PRDM16. Mediation and Mendelian randomization analyses implicated cardiometabolic traits, although no single trait accounted for more than half of the shared liability. These results prioritize genetic mechanisms and pharmacological hypotheses for psoriasis-associated cardiovascular multimorbidity.

Principal component analysis (PCA) is routinely used to visualize population structure, yet how sparse sampling and rare-variant depletion affect low-dimensional plots remains poorly understood. Using spatial simulations, we show that these factors interact to distort visualization of genetic landscapes, producing triangular and three-ray patterns, artificial outliers and misleading clines. We develop an objective-guided manifold-learning framework that searches across genotype normalization, PCA representation and dimensionality, distance metrics, and UMAP, densMAP and PHATE parameters. High-dimensional classic PC scores consistently outperform the eigenvectors used in population genetics, but other optimal parameters and ranking objectives depend on data quality, sampling and SNP ascertainment. Across six human and animal datasets, optimized embeddings recover fine-scale structure obscured by PCA and supported by independent genetic evidence. In ancient Eurasia, optimized PHATE resolves Slavic-associated structure corroborated by haplotype-sharing communities, qpAdm, and Y-chromosome lineages. These results call for caution in interpreting PCA plots and establish optimized manifold learning as a hypothesis-generating approach.

Different regions of genes perform distinct functions and vary in their importance to human health. Evolutionary intolerance provides a powerful means of identifying regions where disruptive mutations are under strong purifying selection, informing genetic disease discovery and variant interpretation. However, estimating intolerance in small sub-genic regions from population variation alone is underpowered and unstable. We present PRIME, a Bayesian model that stabilizes estimates of regional missense intolerance by sharing information hierarchically across regions. Importantly, PRIME produces a full joint posterior across all genes, allowing complex inferential questions that are difficult or impossible to address with existing approaches to be answered. We utilize this to identify regions enriched for pathogenic and experimentally deleterious missense variants, improve prioritization of Mendelian disease genes by focusing on their most intolerant regions, and uncover conserved patterns of purifying selection across protein families. Integrating PRIME with existing computational variant predictors improves pathogenicity prediction, demonstrating that regional missense intolerance provides complementary information for clinical variant interpretation.

Unresolved rare disease is a major public health challenge affecting ~300 million people worldwide. At least 50% of these individuals remain genetically unresolved after applying exome sequencing and/or whole genome sequencing. One source of these missing diagnoses is the presence of rare variants within the non-coding genome that are detected but not interpreted by whole genome sequencing. In order to systematically evaluate candidate pathogenic non-coding variants, we created the Genomic Analysis of Variants in Unresolved Rare Disease (GAVURD) system. GAVURD leverages trio whole genome sequencing alignment data to produce a short list of putative pathogenic non-coding variants for a given proband. GAVURD uses best practices for de novo and rare inherited variant identification, links variants to human disease genes harnessing topologically associated domain (TAD) data, and rank prioritizes variants based on phenotypic overlap. As a proof-of-concept, we applied GAVURD to ten probands with unresolved rare disease and implicated six potentially causal non-coding variants based on a confluence of evidence supportive of pathogenicity. The GAVURD system serves an important role in prioritizing candidate non-coding causal variants for unresolved rare disease that can serve as the high value and informed focus of additional functional follow-up studies.

Including recently admixed populations in genome-wide association studies (GWAS) is important for equitable and ancestry-resolved genetic discovery. The existing popular method, Tractor, estimates ancestry-specific effects from individual-level data but cannot leverage external GWAS summary statistics due to mismatches in underlying model parameters. We introduce TLS-Tractor, a transfer-learning method that uses the generalized method of moments to integrate external GWAS summary statistics with internal individual-level data for local ancestry-aware association analysis. In simulations, TLS-Tractor controlled type I error, accurately estimated ancestry-specific effects, and increased power relative to the internal-only Tractor. Analyses integrating African-European admixed participants from All of Us with Million Veteran Program summary statistics corroborated these gains and showed that local ancestry adjustment can improve calibration, localization, and interpretation, whereas standard GWAS meta-analysis often provides greater power. We introduce an efficient tlstractor R package that achieves over 200x faster local ancestry tract extraction and 4-32x faster association testing than the original Tractor implementation.