Fine-mapping methods for complex traits: essential adaptations for samples of related individuals

Abstract

Fine-mapping causal variants from genome-wide association studies (GWAS) loci is challenging in populations with substantial relatedness, such as livestock, as standard methods often assume unrelatedness, leading to poor fine-mapping accuracy. Here, we introduce a comprehensive Bayesian framework to address this. Our approach features BFMAP-Shotgun Stochastic Search for individual-level data, which uses a linear mixed model (LMM) and shotgun stochastic search with simulated annealing. For summary statistics, we develop FINEMAP-adj and SuSiE-adj, novel strategies that directly use standard FINEMAP and SuSiE for samples of related individuals by employing LMM-derived inputs (particularly a relatedness-adjusted linkage disequilibrium matrix). Furthermore, genomic-feature posterior inclusion probability (PIP), implemented here as gene-level PIP (PIP_gene), is proposed to enhance detection power by aggregating variant signals. Extensive simulations based on pig genotypes across diverse heritability levels and population structures (pure-breed and multi-breed) show our methods substantially outperform existing tools (FINEMAP, SuSiE, FINEMAP-inf, SuSiE-inf, and GCTA-COJO) in samples of related individuals, achieving notable improvements in fine-mapping accuracy (e.g. up to several-fold increases in the area under the precision-recall curve). Multi-breed populations greatly enhance fine-mapping accuracy compared to single-breed populations. Additionally, PIP_gene markedly improves candidate gene identification. Application to Duroc pig traits demonstrates practical utility, with functional enrichment analysis confirming our methods’ superior identification of biologically relevant variants. This work provides robust, validated methods and associated software for accurate fine-mapping in populations with complex relatedness.

Introduction

Genome-wide association studies (GWAS) have successfully discovered many genetic associations with complex traits and diseases [1]. However, these studies primarily identify associated genomic regions rather than pinpoint the actual causal variants, largely due to linkage disequilibrium (LD), whereby noncausal variants are correlated with true causal variants [2, 3]. Such associated regions often contain many SNPs, the majority of which have no direct effect on the trait [4], thereby complicating downstream functional validation. Fine-mapping aims to address this by statistically prioritizing the most likely causal variants within a GWAS-identified locus, assuming at least one such variant exists [5] and thus narrowing the field for functional studies.

Statistical fine-mapping methods typically account for LD structure to identify a small set of putative causal variants, effectively treating the problem as variable selection in a regression model [6, 7]. Given m SNPs in a region, 2^m possible causal models exist, making exhaustive evaluation computationally prohibitive for densely genotyped regions. Early approaches like conditional stepwise selection (e.g. GCTA-COJO) [8] identify independent signals but can be suboptimal, particularly when multiple causal variants in LD are present or when a noncausal variant strongly tags multiple causal variants [5, 7]. Methods like CAVIAR [7] and its extension CAVIARBF [9] improved upon this by jointly modeling multiple variants and exhaustively evaluating all possible models with up to a pre-specified small number of causal variants, but faced computational limitations with larger regions or higher numbers of true causal variants. To address this challenge, newer approaches were developed. FINEMAP [10] employs shotgun stochastic search (SSS) [11] to efficiently explore the vast model space, while SuSiE uses Iterative Bayesian Stepwise Selection with the “Sum of Single Effects” model [12, 13] to search multiple effects. These methods have become widely adopted due to their improved computational efficiency. Recent extensions like FINEMAP-inf and SuSiE-inf [14] further refine these by attempting to model infinitesimal effects within the candidate region.

Despite these advancements, a major limitation of most existing fine-mapping methods, particularly those based on summary statistics, is their foundation in linear regression models that assume samples of unrelated individuals. The use of related individuals violates the underlying model assumption of unrelatedness, potentially distorting LD and subsequent causal variant identification. While it has become standard to use summary statistics derived from linear mixed models (LMMs) that account for relatedness and/or infinitesimal background at the GWAS stage [15, 16], the downstream fine-mapping tools often still do not correctly incorporate this relatedness structure into their own models. This critical issue of appropriately adapting fine-mapping for related individuals has particularly been overlooked in livestock genetics, where populations typically exhibit substantial and complex inter-individual relatedness due to breeding practices.

This study introduces a comprehensive Bayesian fine-mapping framework specifically designed to address these challenges in samples of related individuals. We aim to demonstrate that existing summary-statistics-based fine-mapping methods cannot be naively applied to samples of related individuals and develop comprehensive methodological solutions for related samples. We build upon our previously developed BFMAP approach, which uses an LMM for individual-level data and originally employed forward selection (BFMAP-Forward) [17]. Here, we enhance the BFMAP framework by implementing a more robust model exploration strategy, BFMAP-Shotgun Stochastic Search (BFMAP-SSS), to better handle scenarios with highly correlated causal variants. Recognizing the widespread use of summary statistics, we further develop novel adaptation strategies, FINEMAP-adj and SuSiE-adj. These approaches enable the direct use of standard FINEMAP and SuSiE software with summary statistics obtained from related individuals by inputting LMM-derived associations and a relatedness-adjusted LD matrix. Finally, to improve detection power in genomic regions with challenging LD patterns, as is common in livestock, we propose a general framework of genomic-feature posterior inclusion probability that aggregates variant-level evidence across genomic features. We implement and validate the use of this framework through gene-level posterior inclusion probability (PIP_gene).

Through simulations based on real livestock genotypes and application to economically important traits in Duroc pigs, we demonstrate both the inadequacy of naively applying existing summary-statistics methods to related individuals and the substantial improvements offered by our comprehensive methodological solutions (BFMAP-SSS, FINEMAP-adj, SuSiE-adj, and PIP_gene) in fine-mapping performance. This work provides novel methodologies and associated software essential for robust identification of causal variants and trait-associated genomic features in samples with complex genetic relatedness.

Materials and methods

Fine-mapping methodology

We developed a framework for fine-mapping complex traits in samples of related individuals, addressing analyses with both individual-level data and summary statistics. Figure 1 provides a schematic representation of this framework, illustrating the relationships between existing methods and the novel approaches developed herein. Detailed mathematical derivations for all methods are provided in Supplementary Materials.

Figure 1.

Overview of fine-mapping methods for unrelated and related individuals. Overview of fine-mapping approaches for samples of unrelated (A, C, and E) and related (B, D, and F) individuals. (A) Linear regression provides the foundation for fine-mapping in samples of unrelated individuals. (B) BFMAP employs LMMs with GRM for fine-mapping in samples of related individuals. (C) Standard summary-statistics methods (FINEMAP and SuSiE) are derived from the linear regression framework. (D) Adapted methods (FINEMAP-adj and SuSiE-adj) use relatedness-adjusted correlations () and effective sample size (ñ) derived from the BFMAP framework. The red dashed arrow indicates that standard FINEMAP and SuSiE software can be directly applied with these adapted inputs. Through the approximation D ≈ Iñ, these methods achieve compatibility with standard FINEMAP and SuSiE implementations. (E) FINEMAP-inf and SuSiE-inf extend their respective base methods by modeling infinitesimal effects within candidate regions. (F) all methods produce variant-level posterior inclusion probability (PIP; or PCIP for BFMAP-forward) as output. Gene-level PIP, available for methods that provide posterior model probabilities (BFMAP-forward, BFMAP-SSS, FINEMAP-adj, and FINEMAP), aggregates variant-level evidence for prioritizing causal genes.

Typical fine-mapping tools, such as FINEMAP [10] and SuSiE [13], usually use summary statistics. Their underlying models are derived from linear regression assuming samples of unrelated individuals, which limits their direct applicability or accuracy when substantial relatedness is present (Fig. 1A and C). To overcome this, we developed BFMAP, a Bayesian fine-mapping approach first introduced for individual-level livestock data using forward selection (BFMAP-Forward) [17]. BFMAP employs an LMM that explicitly accounts for whole-genome infinitesimal effects and genetic relatedness through a genomic relationship matrix (GRM) (Fig. 1B). In this study, alongside the established BFMAP-Forward, we introduce BFMAP-SSS, a novel BFMAP implementation coupling SSS with simulated annealing for enhanced model space exploration.

Building upon the BFMAP framework, we developed a strategy to leverage its advantages within a summary statistics context. The BFMAP model, initially designed for individual-level data, can be transformed into an equivalent model based on summary statistics (Fig. 1D). Through an approximation (), analogous to the one used in GRAMMAR-Gamma, this transformed model yields a functional form identical to that underlying standard FINEMAP and SuSiE. This key structural equivalence enables the direct adaptation of these existing tools for accurate fine-mapping in related individuals. This adaptation involves a two-step process: first, calculating relatedness-aware inputs [including standard LMM-derived z-scores (z), an effective sample size (ñ), and a relatedness-adjusted LD matrix (), as detailed in Supplementary Methods]; and second, providing these as the conventional z-score, sample size, and LD matrix inputs to FINEMAP or SuSiE. We term these adapted methods FINEMAP-adj and SuSiE-adj.

It is noteworthy that recently proposed fine-mapping methods, FINEMAP-inf and SuSiE-inf [14], are conceptually related to our LMM-based approaches (BFMAP, FINEMAP-adj, and SuSiE-adj) in that both incorporate terms to account for infinitesimal effects (Fig. 1E). However, a key distinction is that FINEMAP-inf and SuSiE-inf model the infinitesimal effects of variants within the candidate fine-mapping region, whereas our methods model whole-genome infinitesimal effects via GRM.

Finally, recognizing the challenges of fine-mapping to single-variant resolution in populations with extensive LD, such as livestock, we propose the use of genomic-feature posterior inclusion probabilities, which aggregate variant-level evidence to assess the overall evidence for any defined genomic feature harboring at least one non-zero-effect variant. In this work, we focus specifically on gene-level posterior inclusion probability (PIP_gene) as our primary implementation of this concept, aggregating evidence to evaluate whether a gene contains at least one causal variant and offering a balance between localization precision and detection power (Fig. 1F). In the subsequent sections, the performance of our novel contributions (BFMAP-SSS, FINEMAP-adj, and SuSiE-adj), alongside our previously developed BFMAP-Forward approach [17], is evaluated in comparison with existing approaches (FINEMAP, SuSiE, FINEMAP-inf, SuSiE-inf, and GCTA-COJO) for fine-mapping in related individuals.

Data simulation

We evaluated fine-mapping methods using genotype data from 31 301 Duroc pigs provided by Smithfield Premium Genetics (Roanoke Rapids, NC, USA). The animals were genotyped with the PorcineSNP60 BeadChip (Illumina Inc., San Diego, CA, USA) and genotypes were imputed to 34 615 361 autosomal variants using SWIM 1.0 [18]. This full set of imputed variants was then filtered for minor allele frequency (MAF) ≥1%, Hardy–Weinberg equilibrium (P ≥ 1 × 10⁻⁸), and IMPUTE2 info score ≥ 0.8, resulting in ~11.7 million quality-controlled variants on autosomes. From these quality-controlled variants, we randomly selected a 4-Mb region on chromosome 1, followed by minimal LD pruning (-indep-pairwise 2000 200 0.999), resulting in 6504 variants for the simulation study. This region exhibited complex LD structure characteristic of livestock populations, with 15.7% of all variant pairs showing an LD r² of 0.3 or higher (Supplementary Fig. S1), creating challenging conditions where causal variants can be highly correlated with each other and with many noncausal variants.

Quantitative traits were simulated using MPH [19] based on an LMM (Supplementary Model 5) with heritability levels of 0.05, 0.3, and 0.5 to represent low-, moderate-, and high-heritability traits commonly observed in livestock populations. The GRM (G) for the trait simulation was computed using 10 000 randomly selected, quality-controlled chip SNPs. We positioned 1, 2, or 3 causal variants within the central 2-Mb region, collectively explaining either 1% or 4% of phenotypic variation (PVE). Simulations used samples of 5000 or 10 000 randomly selected individuals, generating 100 replicates for each of the 36 scenarios (three heritability levels × three causal variant counts × two PVE levels × two sample sizes).

To assess Type I error control, we conducted simulations under the null hypothesis in which no causal variants were present while polygenic background effects were retained. We generated 1000 replicates of 10 000 individuals under three heritability levels (h² = 0.05, 0.3, and 0.5) and performed GWAS to evaluate false positive rates at different significance thresholds. We estimated the genome-wide Type I error rate as the proportion of replicates with ≥1 association exceeding a given threshold. Additionally, we applied fine-mapping to replicates showing spurious associations to assess each method’s propensity to assign high PIPs under the null.

We further evaluated method performance across genetically diverse populations by extending our simulation framework to include multi-breed scenarios combining Duroc and Yorkshire pig genotypes, with Yorkshire data provided by AcuFast LLC (Navasota, TX, USA). Multi-breed datasets were constructed by merging genotype data from both breeds. Quantitative traits were simulated using the same LMM (Supplementary Model 5) as in single-breed scenarios, with the heritability set to 0.5. Causal variants were sampled from the same 2-Mb candidate region used in single-breed simulations, with 1, 2, or 3 variants collectively explaining either 1% or 4% of phenotypic variance, yielding six simulation scenarios. To account for breed-specific allele frequencies, the additive genetic variance of a variant in a structured population was calculated as , where and are the pooled allele frequencies for the variant, is the additive effect, and quantifies the standardized variance in allele frequencies among subpopulations following Wright’s formulation [20]. The term follows the classical derivation of per-variant additive variance [21], with accounting for population structure. For each simulation replicate, combined datasets included equal numbers of individuals from each breed (5000 Duroc and 5000 Yorkshire pigs), with breed identity included as a covariate in subsequent analyses. This approach enabled evaluation of fine-mapping performance in multi-breed studies common in livestock genomics.

For gene-based fine-mapping evaluation, we partitioned the 4-Mb candidate region into 100 equal blocks to serve as gene proxies. This block size approximates the average pig gene length of ~33 kb derived from the Sscrofa11.1 genome annotation (Ensembl release 113) [22]. To generate the required summary statistics for fine-mapping from each simulated dataset, we conducted association tests for variants in the candidate region using an LMM implemented in GCTA-MLMA (v1.94.1). GCTA-MLMA used the same GRM as in the trait simulation. Furthermore, required LD matrices were generated from each simulation sample’s genotype data: both the standard SNP correlation matrix (Σ) as in Supplementary Equation (3) and the relatedness-adjusted correlation matrix () as in Supplementary Equation (9). The standard matrix (Σ) was used by GCTA-COJO, FINEMAP, SuSiE, FINEMAP-inf, and SuSiE-inf, while the relatedness-adjusted matrix () was used by FINEMAP-adj and SuSiE-adj.

Performance evaluation in simulations

The performance of various fine-mapping approaches was benchmarked using the datasets generated across multiple simulation scenarios. For single-breed (Duroc) populations, we evaluated 36 scenarios (three causal variant counts × two PVE levels × two sample sizes × three heritability levels). For multi-breed (Duroc–Yorkshire) populations, we evaluated six additional scenarios (three causal variant counts × two PVE levels, with h² = 0.5 and total sample size = 10 000). It is important to note that while 100 replicates were generated for each scenario, not all replicates resulted in a significant association signal within the candidate region when analyzed by GCTA-MLMA. Specifically, only those simulation replicates where at least one variant in the candidate region achieved a GCTA-MLMA P-value <1 × 10⁻⁵ were retained for subsequent fine-mapping evaluation. This yielded a minimum of 63 such replicates for any given scenario, with detailed counts available in Supplementary Table S1.

The methods evaluated included our novel approaches (BFMAP-SSS, FINEMAP-adj, SuSiE-adj, and PIP_gene), our previously described BFMAP-Forward method [17], and several well-established reference methods: FINEMAP v1.4.2 [10], SuSiE (as implemented in susieR v0.12.35) [12, 13], FINEMAP-inf v1.3, SuSiE-inf v1.4 [14], and GCTA-COJO (--cojo-slct in GCTA v1.94.1) [8, 23]. BFMAP-Forward and BFMAP-SSS were implemented in BFMAP v0.90, while FINEMAP-adj and SuSiE-adj directly used FINEMAP and SuSiE, respectively, with adapted inputs designed for related individuals. For consistency across all applicable methods, a causal variant’s prior PVE and the maximum number of causal variants were set to 0.01 and 5, respectively, in all simulation datasets. Furthermore, the same GRM, constructed from the 10 000 quality-controlled chip SNPs as detailed in the trait simulation, was used by all our LMM-based fine-mapping methods requiring such an input, specifically BFMAP (both BFMAP-Forward and BFMAP-SSS) and our -adj approaches (FINEMAP-adj and SuSiE-adj).

We applied consistent P-value-based prefiltering for fine-mapping across methods. Specifically, for GCTA-COJO (--cojo-slct), the significance threshold in its stepwise procedure (--cojo-p) was set to 1 × 10⁻³. Similarly, for all other fine-mapping tools evaluated (FINEMAP, SuSiE, FINEMAP-inf, SuSiE-inf, and our BFMAP and -adj approaches), only variants with a GWAS P-value <1 × 10⁻³ from the respective input summary statistics (GCTA-MLMA) were included in the fine-mapping analysis. This initial filtering threshold was chosen to reduce the computational burden of model searching for each tool while aiming to be sufficiently inclusive of potentially causal variants. For SuSiE and SuSiE-adj analyses, we set estimate_residual_variance = TRUE to enable proper variance estimation since we used in-sample LD matrices computed from the same individuals as the association analysis. Unless otherwise specified, all other parameters for these software tools were maintained at their default values.

We applied the fine-mapping framework to multi-breed data by treating combined populations as a single analysis unit while accounting for breed-specific effects through covariates. This approach assumes causal variants are shared across breeds with identical substitution effects, while noncausal variants exhibit different apparent effects due to breed-specific LD structures. This differential pattern enhances our ability to distinguish causal from non-causal variants in multi-breed contexts [24–26].

To evaluate fine-mapping accuracy, variants from each method were designated as putatively causal based on their respective PIPs (or PCIPs for BFMAP-Forward) using thresholds ranging from 0.0 to 1.0. Performance was primarily quantified using recall (also known as sensitivity or power) and precision (also known as positive predictive value). Recall is the proportion of true causal variants correctly identified [calculated as TP / (TP + FN)], and precision is the proportion of designated putatively causal variants that are indeed true causal [calculated as TP / (TP + FP)]. Here, TP, FP, and FN represent the number of true positives, false positives, and false negatives, respectively, determined at a given PIP (or PCIP) threshold and averaged over all simulation replicates retained for fine-mapping evaluation per scenario (Supplementary Table S1). Precision and recall were also calculated for gene-level PIPs using the same approach, with genes rather than variants as the unit of evaluation.

Given that GCTA-COJO (--cojo-slct) does not compute PIPs, a modified evaluation strategy was necessary. Variants identified by GCTA-COJO were ranked by first prioritizing those from its jma output, followed by variants in its cma output, with the latter ranked by their conditional P-values. This P-value-based ranking was then used to assess GCTA-COJO’s ability to prioritize true causal variants relative to the PIP/PCIP-based rankings from other methods. Specifically, we evaluated its recall by selecting the top k-ranked variants (where k ranged from 1 to 20) as putatively causal. For each k, recall was calculated as the proportion of all true causal variants that were included in this top k selection, averaged across all qualifying simulation replicates per scenario.

Duroc pig trait analysis

This study also analyzed several economically important traits from the same cohort of Duroc pigs described in the Data simulation section. Phenotype data, collected by Smithfield Premium Genetics between 2015 and 2021, included growth traits for ~27 000 animals [off-test body weight (WT), back fat thickness (BF), and loin muscle depth (MS)] and reproduction traits for 3290 sows [number of piglets born alive (NBA), born dead (NBD), and weaned (NW)]. The three growth traits were collected at the end of the performance test, as detailed in Bergamaschi et al. [27]. All phenotypes were pre-adjusted to remove systematic environmental effects, including farm and physiological sources of variation. For association and fine-mapping analyses, we utilized the ~11.7 million quality-controlled autosomal variants (derived as detailed in the Data simulation section).

GWAS was performed using SLEMM-GWA v0.89.5 [28], which provides accuracy comparable to GCTA-MLMA but with orders of magnitude greater computational efficiency (Supplementary Fig. S2). Association peaks were identified by visual inspection of the Manhattan plot for each trait, with a P-value threshold of 1 × 10⁻⁵. For each retained peak, candidate region boundaries were established by extending 2 Mb upstream and downstream from the position of the minimum P-value variant. Variants with P < 1 × 10⁻³ within each defined candidate region were then retained for subsequent fine-mapping analyses. The same 10 000 quality-controlled chip SNPs selected for the trait simulation were used to construct the GRM for methods requiring it (e.g. SLEMM and our BFMAP and -adj approaches).

To evaluate whether fine-mapped variants were enriched in functional genomic features, we employed GEMRICH (https://github.com/jiang18/gemrich), which estimates functional enrichment of causal variants across disjoint annotation categories. Functional annotations were derived from the Ensembl Sscrofa11.1 gene annotation (release 113). We analyzed enrichment in regions comprising coding sequences (CDS), promoters (defined as 3 kb upstream of transcription start sites), and untranslated regions (UTRs), collectively termed “CPU” regions. For this analysis, enrichment was estimated separately with fine-mapping results from each method (BFMAP-SSS, BFMAP-Forward, FINEMAP, FINEMAP-adj, FINEMAP-inf, SuSiE, SuSiE-adj, and SuSiE-inf). FINEMAP and FINEMAP-adj were run with 100 000 iterations, and all FINEMAP-based methods (FINEMAP, FINEMAP-adj, and FINEMAP-inf) were set to save 1 000 000 configurations to ensure robust statistical inference. SuSiE-based methods (SuSiE, SuSiE-adj, and SuSiE-inf) were run with min_abs_corr = 0 to retain all credible sets.

Results

Simulation benchmarks

To rigorously evaluate the fine-mapping approaches, we conducted simulation studies based on real pig genotypes across diverse scenarios. Single-breed (Duroc) simulations varied by heritability (0.05, 0.3, or 0.5), number of causal variants (1, 2, or 3) within the candidate region, total proportion of variance explained (PVE) by these causal variants (1% or 4%), and sample size (5000 or 10 000 individuals). Multi-breed (Duroc–Yorkshire) simulations varied by number of causal variants (1, 2, or 3) and PVE (1% or 4%), with h² = 0.5 and 10 000 individuals. For each scenario, 100 replicates were generated, though only those yielding significant GWAS associations (GCTA-MLMA P < 1 × 10⁻⁵) were retained for fine-mapping evaluation, ensuring robust downstream analysis.

Impact of relatedness adjustment on genotype correlations

Summary-statistics-based fine-mapping methods critically rely on an accurate LD matrix reflecting correlations between genotypes. In samples of related individuals, one might naively compute a standard LD matrix directly from the observed genotypes. However, as we have theoretically established in Supplementary Methods (“Fine-mapping with related individuals: summary statistics”), a relatedness-adjusted LD matrix () is required to properly account for the genetic relatedness among individuals in fine-mapping. To empirically illustrate the impact of this adjustment, we compared elements of the standard LD matrix (Σ) with those of the relatedness-adjusted LD matrix () within our 4-Mb candidate region for simulations (Supplementary Fig. S3). This comparison revealed substantial differences between the two types of correlations. Notably, the relatedness adjustment (transforming Σ to ) often led to a shrinkage of correlation magnitudes, particularly for variants that exhibited strong positive or negative correlations in the standard Σ matrix. Consequently, applying fine-mapping methods with a standard, unadjusted LD matrix in samples of related individuals, rather than the appropriate relatedness-adjusted version, can severely compromise fine-mapping performance and the accuracy of causal variant identification.

Comparative performance of proposed and existing fine-mapping methods

Overall, GCTA-COJO’s performance in prioritizing causal variants was generally comparable to or slightly poorer than established PIP-based methods across the simulation scenarios (Supplementary Fig. S4). Consequently, while its results are available, our primary comparative focus shifts to the PIP/PCIP-based methods.

Our methods designed for related individuals (BFMAP-Forward, BFMAP-SSS, FINEMAP-adj, and SuSiE-adj) consistently and substantially outperformed the established approaches (FINEMAP, SuSiE, FINEMAP-inf, and SuSiE-inf) across all 42 simulation scenarios, spanning different heritability levels, sample sizes, and population structures, as illustrated by the precision-recall curves (Figs 2 and 3) and the area under the precision-recall curve (AUPRC) values (Table 1). This performance advantage for our methods was particularly pronounced in scenarios with larger sample sizes, higher PVE, or multi-breed contexts. For single-breed Duroc populations with h² = 0.5 and 10 000 individuals, when AUPRCs were averaged across the six scenarios (three causal variant counts × two PVE levels) with 10 000 individuals, BFMAP-SSS (average AUPRC = 0.239) showed approximately a 3.9-fold higher AUPRC than the best-performing reference method, SuSiE using LMM-derived association statistics (average AUPRC = 0.049) (Table 1). Specifically, FINEMAP-adj and SuSiE-adj, which incorporate relatedness-adjusted LD matrices and LMM-derived association statistics, achieved remarkable improvements over their counterparts (FINEMAP and SuSiE) that use standard LD matrices, with FINEMAP-adj yielding a 5.6-fold higher average AUPRC than standard FINEMAP, and SuSiE-adj a 3.1-fold higher average AUPRC than standard SuSiE. These substantial gains were achieved when standard FINEMAP and SuSiE were provided with LMM-derived association statistics under these conditions with 10 000 individuals (Table 1). The consistency of performance advantages across heritability levels confirms the robustness of our approaches. Although the performance differential narrowed with declining heritability, our approaches maintained superiority even at h² = 0.05. For example, with 10 000 individuals, BFMAP-SSS achieved average AUPRC of 0.212 compared to 0.137 for the best-performing reference method (SuSiE) at h² = 0.05, and 0.220 compared to 0.059 at h² = 0.3. Remarkably, multi-breed populations enhanced the absolute performance of our methods, with BFMAP-SSS achieving a 0.7-fold increase in average AUPRC in multi-breed contexts (0.408) compared to single-breed populations (0.239), demonstrating that genetic diversity from multiple breeds improves fine-mapping accuracy.

Figure 2.

Precision-recall curves for identifying causal variants with PIPs across simulation scenarios in single-breed Duroc populations. GWAS summary data were derived from GCTA-MLMA (LMM). (A) Comparison of standard fine-mapping methods (FINEMAP and SuSiE) and their relatedness-adjusted adaptations (FINEMAP-adj and SuSiE-adj). (B) Comparison of BFMAP implementations (BFMAP-SSS and BFMAP-forward) and infinitesimal-effect methods (FINEMAP-inf and SuSiE-inf). Results are shown for Duroc purebred populations with heritability h² = 0.5 across different sample sizes (5000 and 10 000 individuals), PVE levels (1% and 4%), and numbers of causal variants (1–3). Each panel contains the same 12 simulation scenarios for direct comparison across methods.

Figure 3.

Precision-recall curves for identifying causal variants with PIPs across simulation scenarios in multi-breed Duroc-Yorkshire populations. GWAS summary data were derived from GCTA-MLMA (LMM) with breed included as covariate. (A) Comparison of standard fine-mapping methods (FINEMAP and SuSiE) and their relatedness-adjusted adaptations (FINEMAP-adj and SuSiE-adj). (B) Comparison of BFMAP-SSS and infinitesimal-effect methods (FINEMAP-inf and SuSiE-inf). Results show performance in combined Duroc-Yorkshire populations (5000 individuals per breed with N = 10 000 total) across different PVE levels (1% and 4%) and numbers of causal variants (1–3). Each panel contains the same 6 simulation scenarios for direct comparison across methods.

Table 1.

AUPRC for causal variant identification by fine-mapping methods across simulation scenarios.

Population	h 2	Sample size	Method	Number of causal variants						Average
				1		2		3
				PVE = 0.01	PVE = 0.04	PVE = 0.01	PVE = 0.04	PVE = 0.01	PVE = 0.04
Duroc	0.05	5000	BFMAP-SSS	0.126	0.339	0.051	0.192	0.019	0.101	0.138
			BFMAP-Forward	0.115	0.306	0.040	0.077	0.014	0.041	0.099
			FINEMAP-adj	0.126	0.337	0.053	0.150	0.021	0.069	0.126
			SuSiE-adj	0.118	0.339	0.046	0.078	0.019	0.052	0.109
			FINEMAP	0.114	0.336	0.044	0.089	0.016	0.076	0.112
			FINEMAP-inf	0.081	0.330	0.046	0.146	0.016	0.091	0.118
			SuSiE	0.112	0.333	0.042	0.081	0.019	0.052	0.106
			SuSiE-inf	0.054	0.164	0.021	0.051	0.007	0.027	0.054
		10 000	BFMAP-SSS	0.191	0.511	0.056	0.336	0.031	0.148	0.212
			BFMAP-Forward	0.197	0.517	0.017	0.126	0.014	0.081	0.159
			FINEMAP-adj	0.191	0.506	0.049	0.217	0.027	0.120	0.185
			SuSiE-adj	0.193	0.506	0.035	0.168	0.021	0.079	0.167
			FINEMAP	0.173	0.212	0.036	0.104	0.016	0.056	0.100
			FINEMAP-inf	0.152	0.187	0.037	0.175	0.016	0.084	0.109
			SuSiE	0.186	0.428	0.033	0.088	0.018	0.065	0.137
			SuSiE-inf	0.086	0.098	0.018	0.046	0.010	0.027	0.048
	0.3	5000	BFMAP-SSS	0.136	0.355	0.053	0.176	0.024	0.096	0.140
			BFMAP-Forward	0.120	0.323	0.040	0.114	0.007	0.043	0.108
			FINEMAP-adj	0.136	0.348	0.052	0.161	0.023	0.093	0.135
			SuSiE-adj	0.112	0.358	0.048	0.103	0.022	0.062	0.117
			FINEMAP	0.083	0.098	0.022	0.042	0.011	0.020	0.046
			FINEMAP-inf	0.064	0.081	0.023	0.057	0.015	0.035	0.046
			SuSiE	0.108	0.257	0.035	0.041	0.010	0.016	0.078
			SuSiE-inf	0.024	0.051	0.006	0.024	0.004	0.009	0.019
		10 000	BFMAP-SSS	0.195	0.524	0.058	0.354	0.030	0.162	0.220
			BFMAP-Forward	0.176	0.522	0.027	0.194	0.015	0.112	0.174
			FINEMAP-adj	0.191	0.516	0.058	0.291	0.030	0.129	0.202
			SuSiE-adj	0.190	0.529	0.040	0.210	0.028	0.100	0.183
			FINEMAP	0.070	0.046	0.016	0.036	0.010	0.022	0.033
			FINEMAP-inf	0.050	0.024	0.018	0.045	0.011	0.029	0.029
			SuSiE	0.153	0.099	0.023	0.042	0.016	0.024	0.059
			SuSiE-inf	0.021	0.020	0.018	0.014	0.005	0.011	0.015
	0.5	5000	BFMAP-SSS	0.128	0.419	0.052	0.224	0.030	0.113	0.161
			BFMAP-Forward	0.124	0.379	0.032	0.143	0.011	0.061	0.125
			FINEMAP-adj	0.126	0.408	0.053	0.204	0.031	0.092	0.152
			SuSiE-adj	0.120	0.415	0.036	0.155	0.025	0.072	0.137
			FINEMAP	0.067	0.064	0.024	0.020	0.013	0.013	0.034
			FINEMAP-inf	0.033	0.021	0.016	0.012	0.010	0.011	0.017
			SuSiE	0.120	0.192	0.032	0.032	0.019	0.022	0.070
			SuSiE-inf	0.021	0.027	0.009	0.016	0.005	0.009	0.015
		10 000	BFMAP-SSS	0.186	0.469	0.097	0.382	0.076	0.227	0.239
			BFMAP-Forward	0.140	0.469	0.053	0.250	0.061	0.157	0.188
			FINEMAP-adj	0.186	0.474	0.095	0.309	0.074	0.218	0.226
			SuSiE-adj	0.176	0.477	0.086	0.238	0.064	0.157	0.200
			FINEMAP	0.046	0.056	0.023	0.037	0.015	0.027	0.034
			FINEMAP-inf	0.024	0.028	0.011	0.017	0.006	0.018	0.017
			SuSiE	0.109	0.051	0.040	0.038	0.024	0.032	0.049
			SuSiE-inf	0.018	0.016	0.012	0.012	0.009	0.008	0.013
Duroc and Yorkshire	0.5	10 000	BFMAP-SSS	0.410	0.753	0.179	0.632	0.063	0.410	0.408
			FINEMAP-adj	0.410	0.742	0.182	0.596	0.063	0.368	0.393
			SuSiE-adj	0.417	0.746	0.156	0.573	0.061	0.314	0.378
			FINEMAP	0.153	0.133	0.061	0.116	0.021	0.052	0.089
			FINEMAP-inf	0.113	0.128	0.044	0.159	0.029	0.052	0.088
			SuSiE	0.217	0.280	0.087	0.211	0.035	0.069	0.150
			SuSiE-inf	0.010	0.012	0.013	0.001	0.007	0.010	0.009

Note: AUPRC values were calculated by varying PIP thresholds (or PCIP for BFMAP-Forward) from 0 to 1 and computing the area under the resulting precision-recall curves. GWAS summary data were derived from GCTA-MLMA (LMM). Results are shown for different sample sizes (5000 and 10 000 individuals), PVE levels (1% and 4%), and numbers of causal variants (1–3).

Among our LMM-based approaches, the overall differences in fine-mapping performance were relatively small (Table 1, Figs 2 and 3). BFMAP-SSS, employing SSS, performed best overall. BFMAP-Forward, which uses a forward selection strategy, showed performance comparable to or slightly less favorable than FINEMAP-adj and SuSiE-adj that implement more exhaustive model space exploration.

Genome-wide Type I error rates were stable across h² levels and ranged from 0.637 to 0.673 at P < 1 × 10⁻⁵ to around 0.05 at P < 5 × 10⁻⁷ (Table 2). Based on these results, we recommend using P < 5 × 10⁻⁷ as the genome-wide significance threshold for prioritizing loci for downstream analysis in pig datasets. For our fine-mapping evaluation, we used P < 1 × 10⁻⁵ as an inclusive threshold to broaden the set of candidate regions.

Table 2.

Genome-wide type I error rates calculated from 1000 simulations under the null hypothesis across significance thresholds and heritability levels.

P-value Threshold	h 2 = 0.05	h 2 = 0.3	h 2 = 0.5
1 × 10−7	0.011	0.011	0.011
5 × 10−7	0.061	0.059	0.053
1 × 10−6	0.115	0.113	0.112
5 × 10−6	0.437	0.419	0.428
1 × 10−5	0.673	0.656	0.637

When fine-mapping was applied to false positive signals under the null hypothesis (replicates with P ≤ 5 × 10⁻⁷), our proposed methods demonstrated superior specificity (Table 3). On average, our methods assigned high PIPs (≥0.9) to substantially fewer variants than standard methods (0–0.136 vs. 0.246–1.396 variants per replicate). This pattern was consistent across all heritability levels, indicating that our methods are less prone to false discoveries even when spurious associations occur.

Table 3.

False positives produced by fine-mapping methods under the null hypothesis.

Method	h 2	Number of significant replicates	Mean N of fine-mapped variants per replicate
			PIP ≥ 0.1	PIP ≥ 0.5	PIP ≥ 0.9
BFMAP-SSS	0.05	61	1.754	0.164	0.000
BFMAP-forward			1.836	0.213	0.049
FINEMAP-adj			1.803	0.180	0.000
SuSiE-adj			1.721	0.148	0.000
FINEMAP			2.361	0.852	0.623
SuSiE			2.393	0.770	0.623
FINEMAP-inf			0.967	0.311	0.246
SuSiE-inf			1.984	0.393	0.328
BFMAP-SSS	0.3	59	2.051	0.288	0.017
BFMAP-forward			1.966	0.390	0.136
FINEMAP-adj			2.034	0.322	0.000
SuSiE-adj			1.729	0.322	0.000
FINEMAP			3.949	1.492	0.780
SuSiE			2.542	0.949	0.576
FINEMAP-inf			2.441	0.864	0.576
SuSiE-inf			1.932	0.814	0.593
BFMAP-SSS	0.5	53	2.000	0.358	0.000
BFMAP-forward			1.849	0.472	0.132
FINEMAP-adj			2.132	0.377	0.019
SuSiE-adj			2.170	0.377	0.019
FINEMAP			4.811	2.019	1.321
SuSiE			2.925	0.962	0.566
FINEMAP-inf			3.604	1.358	0.755
SuSiE-inf			2.604	1.642	1.396

Note: Fine-mapping was applied to replicates with spurious genome-wide significant associations (P ≤ 5 × 10⁻⁷) from simulations under the null hypothesis. Values in the right three columns represent mean number of variants per replicate assigned PIPs above specified thresholds.

Fine-mapping challenges in livestock populations and impact of simulation parameters

It is crucial to note the inherent difficulty of fine-mapping in livestock populations, characterized by strong LD and extensive genetic relatedness among individuals. Even with methods explicitly accounting for relatedness and whole-genome infinitesimal effects, performance can be constrained. For instance, in the simulation scenario with three causal variants, 10 000 individuals, h² = 0.5, and 1% PVE, the best-performing reference method (SuSiE, using its optimal input) yielded an AUPRC of only 0.024 (Table 1). While our novel methods offered a substantial improvement (e.g. BFMAP-SSS achieved an AUPRC of 0.076 in the same scenario), these results underscore that fine-mapping in such populations remains a formidable task.

The simulation parameters also highlighted expected trends. Performance for most methods improved with fewer causal variants or larger PVE for the candidate region. A key observation was the differential response to increased sample size: while our relatedness-aware approaches (BFMAP-Forward, BFMAP-SSS, FINEMAP-adj, and SuSiE-adj) uniformly benefited from larger sample sizes, the methods originally designed for unrelated individuals (FINEMAP, SuSiE, and their -inf variants) showed limited improvement or, in some instances, even a decline in performance as sample size increased when applied to these related samples (Table 1). This underscores the critical importance of appropriately modeling genetic relatedness, especially in larger cohorts of related individuals.

Gene-level posterior inclusion probabilities increase detection power in fine-mapping

To address the challenges of fine-mapping at single-variant resolution, particularly in populations with extensive LD like the Duroc pigs in this study, we evaluated the utility of gene-level posterior inclusion probability (PIP_gene), a specific implementation of our genomic-feature aggregation framework. Since PIP_gene calculation requires posterior model probabilities for variant sets, which are not standard outputs for all fine-mapping tools, we computed PIP_gene using the outputs from BFMAP-Forward, BFMAP-SSS, and FINEMAP-adj, as these methods provide the necessary model-level information. This selection of methods is sufficient to demonstrate the advantages of gene-level aggregation over variant-level PIPs, an orthogonal comparison to the inter-method performance benchmarks detailed previously.

The application of gene-level fine-mapping resulted in a substantial improvement in the ability to correctly identify causal genes (i.e. our defined gene proxy blocks containing one or more true causal variants) for all three methods evaluated. Precision-recall curves consistently show superior performance for gene-level PIPs compared to variant-level PIPs across all simulation scenarios including single-breed (Fig. 4) and multi-breed contexts (Fig. 5), with AUPRC values markedly higher for gene identification than variant identification across all methods and scenarios (Table 4). Notably, the magnitude of this performance gain was often greatest in the more challenging simulation scenarios. For instance, with three causal variants, 1% PVE, h² = 0.5, and 10 000 individuals (a setting where variant-level fine-mapping proved particularly difficult), BFMAP-Forward, BFMAP-SSS, and FINEMAP-adj achieved AUPRCs of 0.061, 0.076, and 0.074, respectively. In contrast, when evidence was aggregated to the gene level using PIP_gene, these same methods achieved AUPRCs of 0.305, 0.423, and 0.414, respectively (Table 4). This represents an approximate 4- to 4.6-fold increase in AUPRC, highlighting the substantial power gained by shifting the inferential focus from individual variants to gene units in complex fine-mapping contexts. Multi-breed populations further enhanced gene-level performance compared to single-breed contexts.

Figure 4.

Precision-recall curves for identifying causal genes with PIP_gene across simulation scenarios in single-breed Duroc populations. Results show the ability of three methods (BFMAP-forward, BFMAP-SSS, and FINEMAP-adj) to identify causal genes (defined as gene proxy blocks containing true causal variants) in Duroc purebred populations with heritability h² = 0.5 under different sample sizes (5000 and 10 000 individuals), PVE levels (1% and 4%), and numbers of causal variants (1–3). A–L represent the 12 simulation scenarios.

Figure 5.

Precision-recall curves for identifying causal genes with PIP_gene across simulation scenarios in multi-breed Duroc–Yorkshire populations. Results show the ability of two methods (BFMAP-SSS and FINEMAP-adj) to identify causal genes (defined as gene proxy blocks containing true causal variants) in combined Duroc–Yorkshire populations (5000 individuals per breed with N = 10 000 total) under different PVE levels (1% and 4%) and numbers of causal variants (1–3). A–F represent the six multi-breed simulation scenarios.

Table 4.

AUPRC for causal gene identification by fine-mapping methods across simulation scenarios.

Population	h 2	Sample size	Method	Number of causal variants						Average
				1		2		3
				PVE = 0.01	PVE = 0.04	PVE = 0.01	PVE = 0.04	PVE = 0.01	PVE = 0.04
Duroc	0.05	5000	BFMAP-SSS	0.540	0.797	0.328	0.566	0.243	0.468	0.490
			BFMAP-Forward	0.495	0.667	0.193	0.321	0.140	0.261	0.346
			FINEMAP-adj	0.537	0.793	0.308	0.520	0.228	0.413	0.467
		10 000	BFMAP-SSS	0.611	0.861	0.485	0.752	0.335	0.582	0.605
			BFMAP-Forward	0.586	0.732	0.285	0.459	0.218	0.379	0.443
			FINEMAP-adj	0.610	0.859	0.447	0.610	0.315	0.517	0.560
	0.3	5000	BFMAP-SSS	0.570	0.793	0.303	0.532	0.228	0.459	0.481
			BFMAP-Forward	0.455	0.649	0.197	0.391	0.106	0.273	0.345
			FINEMAP-adj	0.567	0.789	0.295	0.515	0.222	0.443	0.472
		10 000	BFMAP-SSS	0.622	0.859	0.436	0.759	0.337	0.611	0.604
			BFMAP-Forward	0.585	0.725	0.311	0.463	0.203	0.412	0.450
			FINEMAP-adj	0.620	0.854	0.435	0.681	0.328	0.561	0.580
	0.5	5000	BFMAP-SSS	0.581	0.834	0.402	0.682	0.261	0.533	0.549
			BFMAP-Forward	0.535	0.780	0.277	0.548	0.149	0.371	0.443
			FINEMAP-adj	0.582	0.826	0.402	0.649	0.257	0.494	0.535
		10 000	BFMAP-SSS	0.687	0.857	0.465	0.779	0.423	0.691	0.650
			BFMAP-Forward	0.618	0.828	0.341	0.585	0.305	0.527	0.534
			FINEMAP-adj	0.689	0.854	0.464	0.727	0.414	0.672	0.637
Duroc and Yorkshire	0.5	10 000	BFMAP-SSS	0.869	0.966	0.597	0.943	0.389	0.851	0.769
			FINEMAP-adj	0.865	0.960	0.596	0.930	0.382	0.836	0.761

Note: AUPRC values were calculated by varying PIP_gene thresholds from 0 to 1 and computing the area under the resulting precision-recall curves. Results show the ability of three methods (BFMAP-Forward, BFMAP-SSS, and FINEMAP-adj) to identify causal genes (defined as gene proxy blocks containing true causal variants) under different sample sizes (5000 and 10 000 individuals), PVE levels (1% and 4%), and numbers of causal variants (1–3).

Prior fine-mapping studies often adopt thresholds such as PIP ≥ 0.9 for high-confidence causal variants [14], PIP ≥ 0.5 for a broader set of candidate variants [29], and a lenient threshold of PIP ≥ 0.1 for exploratory analyses [30]. In our analyses, precision increased substantially with stricter PIP thresholds. For variant-level fine-mapping using BFMAP-SSS in pure-breed Duroc populations (averaged across six scenarios with three heritability levels and 10 000 individuals), PIP ≥ 0.1 yields relatively high discovery (recall = 0.298) but modest precision (0.210); PIP ≥ 0.5 gives more balanced results (precision = 0.692, recall = 0.095); and PIP ≥ 0.9 achieves high precision (0.954) at the cost of low recall (0.047). Gene-level analysis using the same BFMAP-SSS framework consistently outperformed variant-level approaches, with PIP ≥ 0.5 achieving superior precision (0.756) and recall (0.416). Thus, we suggest PIP ≥ 0.1 for broad discovery, PIP ≥ 0.5 for candidate identification, and PIP ≥ 0.9 for high-confidence variants.

Computational performance

We evaluated the runtime of our LMM-based fine-mapping methods using simulated datasets with two causal variants (4% PVE) for 5000 and 10 000 individuals, focusing on the same 4-Mb candidate region (3837 and 4371 prefiltered SNPs, respectively) as used in our main simulation benchmarks. Analyses were conducted on an Intel Xeon Gold 6230 CPU using 20 threads. Key fine-mapping parameters, such as the prior PVE per causal variant (0.01) and the maximum number of assumed causal variants (5), were kept consistent with the main simulation benchmarks.

Wall-clock times for all methods, representing the analysis of a single simulation replicate, are detailed in Supplementary Table S2. The three fastest methods (SuSiE-adj, BFMAP-Forward, and FINEMAP-adj) showed comparable speed, requiring 0.22–0.28 min for 5000 individuals and 0.33–0.51 min for 10 000 individuals. The reported times for FINEMAP-adj and SuSiE-adj include precalculation of the relatedness-adjusted LD matrix. BFMAP-SSS, which achieved the highest fine-mapping accuracy, required moderately longer runtimes of ~0.95 min (5000 individuals) and 1.70 min (10 000 individuals). Overall, all methods demonstrated practical computational efficiency, with even BFMAP-SSS completing analyses in under 2 min for datasets with over 4000 variants and 10 000 individuals.

Application to Duroc pig traits

We applied our novel fine-mapping methods to analyze six economically important traits in the Duroc pig population previously described. To identify candidate genomic regions for fine-mapping, we first performed GWAS for each trait using the mixed-model approach implemented in SLEMM-GWA. All six traits produced high-quality Manhattan plots, five of which displayed association peaks with an inclusive significance threshold of P < 1 × 10⁻⁵ (Supplementary Fig. S5), yielding 48 candidate region-trait pairs: 45 for performance traits and three for reproduction traits (Supplementary Table S3).

We also compared our summary-statistics-based method FINEMAP-adj against individual-level BFMAP-SSS by re-analyzing the candidate regions for Duroc pig traits. Overall, FINEMAP-adj and BFMAP-SSS showed high concordance in variant-level PIPs across most regions. However, evaluation of FINEMAP-adj with different iteration settings revealed that the default iteration settings were insufficient for optimal performance in these complex livestock regions. With default iterations (which ranged from 100 to 1717 across regions), the mean correlation between FINEMAP-adj and BFMAP-SSS PIPs across the candidate regions was 0.959, but systematic increases in iteration numbers improved concordance: correlations increased to 0.971 (10 000 iterations) and reached a plateau at 0.982 (100 000 iterations), with no further improvement at 200 000 iterations (Supplementary Fig. S6A). For regions where the correlation remained below 0.9 even with extended iterations, we observed that FINEMAP-adj showed insufficient exploration of complex models compared to BFMAP-SSS. Specifically, FINEMAP-adj exhibited a systematic bias toward single-variant models, while BFMAP-SSS more thoroughly explored configurations with two or three variants (Supplementary Fig. S6B). This suggests that FINEMAP-adj’s (or FINEMAP’s) search algorithm may struggle to adequately explore complex model spaces in some specific cases, even with extended iterations.

Functional enrichment analysis assessed whether fine-mapped variants were enriched in CPU regions (collectively comprising CDS, promoter, and UTR) (Fig. 6). Our methods achieved enrichment values of 8.54–12.13 (BFMAP-SSS: 12.13 ± 2.59; SuSiE-adj: 11.18 ± 2.36; FINEMAP-adj: 8.54 ± 2.06; BFMAP-Forward: 8.79 ± 2.70) compared to 4.56–6.63 for standard methods (FINEMAP: 4.56 ± 1.45; SuSiE: 6.63 ± 2.00; FINEMAP-inf: 5.92 ± 1.80; SuSiE-inf: 4.74 ± 1.26). The consistent enrichment advantage of all our methods over standard approaches suggests superior ability to identify functionally relevant variants.

Figure 6.

Functional enrichment of fine-mapped variants in CDS, promoter, and UTR regions. Fold-enrichment (± SE) of fine-mapped variants in CDS, promoter, and UTR regions (CPU) for each method across all analyzed Duroc pig traits.

To identify putative causal variants and genes within these candidate regions, we initially employed BFMAP-SSS due to its strong performance in simulations. This fine-mapping identified 35 variants with a PIP > 0.1 and 2 variants with a PIP > 0.5 across all region-trait pairs (Supplementary Table S4). We then calculated the gene-level PIP (PIP_gene) for all annotated genes within each candidate region, with gene boundaries defined using Ensembl annotations and extended by 3 kb upstream and downstream. This approach identified 93 gene-trait pairs with PIP_gene > 0.1, of which 26 pairs had a PIP_gene > 0.5 (Supplementary Table S5). These results underscore the utility of this gene-based strategy for prioritizing candidate genes.

The power of PIP_gene was particularly evident in regions where individual variant PIPs were low. For example, a candidate region on chromosome 1 (chr1:51 300 453–55 300 453) associated with back fat thickness (BF) (Fig. 7A) contained no variants with a PIP greater than 0.03 (Fig. 7B), making variant prioritization challenging. However, the PIP_gene analysis revealed that MRAP2 emerged with a strong signal (PIP_gene = 0.669, distinctly higher than other genes in the region; Fig. 7C). The MRAP2 gene encodes melanocortin receptor accessory protein 2, which regulates melanocortin receptor activity and energy homeostasis [31]. Previous studies have identified loss-of-function MRAP2 variants as pathogenic for monogenic hyperphagic obesity in humans [32, 33], and Mrap2-knockout mice develop severe early-onset obesity with increased fat mass and adipocyte hypertrophy [33]. These findings strongly support MRAP2 as a biologically plausible candidate for the BF association observed in our study.

Figure 7.

Fine-mapping results for two candidate regions. BFMAP-SSS fine-mapping analysis of two genomic regions associated with back fat thickness: chr1:51 300 453–55 300 453 and chr6:144 948 358–148 948 358. (A) and (D) show locus zoom plots displaying local association signals within each interval. (B) and (E) show the PIPs of individual variants, and (C) and (F) show the PIPs of genes.

A second illustrative example for BF is the candidate region on chromosome 6 (chr6:144 948 358–148 948 358; Fig. 7D). In this locus, fine-mapping with BFMAP-SSS identified three prioritized variants (each with PIP > 0.1) that are located within the LEPR gene. This gene encodes the leptin receptor, which mediates leptin signaling in energy homeostasis and body weight regulation [34]. The LEPR gene showed a strong gene-level signal (PIP_gene = 0.867), consistent with its well-established role in regulating energy homeostasis and body composition [35–37]. Mutations in this gene can cause severe early-onset obesity, hyperphagia, and hypogonadotropic hypogonadism in humans [38, 39]. While PIPs of all other variants in the candidate region were < 0.05 (Fig. 7E), our PIP_gene analysis highlighted an additional candidate gene, DNAJC6 (PIP_gene = 0.878; Fig. 7F). This gene encodes auxilin-1, a DNAJ/HSP40 family protein that regulates resting metabolic rate [40] and has been shown to suppress adipogenesis and energy metabolism in adipocytes [41]. These findings demonstrate the ability of PIP_gene to enhance detection power for relevant genes whose signals might be diffuse at the individual variant level.

Discussion

Fine-mapping complex traits in populations with extensive relatedness, such as livestock, presents significant statistical challenges. Standard fine-mapping methodologies, predominantly developed for human genetics studies involving largely unrelated individuals, can falter or yield unreliable results when applied naively to these populations. This study addresses this critical gap by introducing a comprehensive framework. This includes BFMAP (a Bayesian approach for individual-level data, encompassing our previously developed BFMAP-Forward [17] and the novel BFMAP-SSS implementation), innovative summary-statistics adaptations (FINEMAP-adj and SuSiE-adj), and a gene-level aggregation metric (PIP_gene), all designed to improve fine-mapping under such complex conditions of relatedness.

A key contribution of this work is the clear demonstration, both theoretically and through extensive pig genotype-based simulations, that appropriately modeling relatedness is critical for achieving accurate fine-mapping. Our results unequivocally show that standard fine-mapping tools (GCTA-COJO, FINEMAP, SuSiE, and their -inf variants), even when provided with summary association statistics from mixed models, exhibit substantially poorer performance than our relatedness-aware methods (BFMAP-Forward, BFMAP-SSS, FINEMAP-adj, and SuSiE-adj) designed to correctly utilize relatedness information (Table 1, Figs 2 and  3). The performance gap, reflected in markedly poorer precision-recall profiles and lower AUPRC values for existing methods (GCTA-COJO, FINEMAP, SuSiE, and their -inf variants) (Table 1, Figs. 2 and  3), underscores the potential for a high proportion of false positives in previous fine-mapping studies in livestock that applied human-centric fine-mapping methods not designed for samples of related individuals (e.g. [42–46]). It is important to note that simply using LMM-derived summary association statistics was insufficient to overcome the fundamental limitations of these methods when applied to samples of related individuals. This highlights the critical need for approaches that explicitly integrate relatedness, primarily through a relatedness-adjusted LD matrix, as implemented in our -adj methods, or via direct LMM modeling as in BFMAP.

Among our proposed relatedness-aware methods, BFMAP-SSS generally emerged as the most accurate, likely due to its sophisticated SSS with simulated annealing allowing for a more thorough exploration of the model space. FINEMAP-adj and SuSiE-adj also performed well, trailing BFMAP-SSS and slightly outperforming BFMAP-Forward (Table 1 and Fig. 2). This performance hierarchy aligns with theoretical expectations: the forward selection strategy of BFMAP-Forward is inherently less exhaustive than the search algorithms in BFMAP-SSS or the variable selection approaches underpinning SuSiE (and thus SuSiE-adj). The slight performance edge of BFMAP-SSS over the -adj methods may stem from the GRAMMAR-Gamma-like approximation employed to allow the FINEMAP and SuSiE programs to be directly applied using appropriately transformed summary statistics (LMM-derived z-scores and effective sample size ñ) and a relatedness-adjusted LD matrix (). While developing new programs for FINEMAP-adj and SuSiE-adj from scratch to circumvent this approximation is feasible, the current -adj approaches offer substantial practical advantages by leveraging well-established, existing software, especially given their performance close to BFMAP-SSS.

The choice among our developed methods in practice will depend on a balance of factors. BFMAP-SSS offers the highest accuracy but is the most computationally intensive, though remains highly practical. For studies where computational resources are a constraint, FINEMAP-adj, SuSiE-adj, and BFMAP-Forward provide faster alternatives with only a modest trade-off in accuracy compared to BFMAP-SSS (Fig. 2 and Supplementary Table S2). FINEMAP-adj, BFMAP-SSS, and BFMAP-Forward also readily provide the posterior model probabilities necessary for calculating PIP_gene, a feature not directly available from SuSiE-adj (which builds on SuSiE) without further methodological development. Furthermore, users should be aware that FINEMAP-adj (which builds on FINEMAP) can sometimes require extended iterations for convergence in complex regions. We therefore recommend BFMAP-SSS when maximal accuracy is paramount and resources permit; FINEMAP-adj when gene-level insights are desired with good speed; SuSiE-adj for fast variant-level fine-mapping without immediate need for PIP_gene; and BFMAP-Forward as a rapid option for individual-level data.

A significant challenge in livestock genomics is the extensive and complex LD structure, usually making fine-mapping at single-variant resolution exceedingly difficult. Our proposed PIP_gene metric addresses this by aggregating variant-level evidence, thereby enhancing the power to detect trait-associated genes, albeit with a trade-off in localization precision from variant to gene. As demonstrated in our simulations, the improvement in detection power (AUPRC) using PIP_gene was substantial, particularly in challenging scenarios with low PVE and multiple causal variants, where gene-level AUPRCs were up to 14.6-fold higher than variant-level AUPRCs (Table 4 and Fig. 4). This utility was further underscored in our analysis of Duroc pig traits, where PIP_gene successfully highlighted biologically relevant genes such as MRAP2 and DNAJC6 for back fat thickness, even when underlying variant PIPs were diffuse or individually weak (Fig. 7). Furthermore, this approach led to the identification of other promising candidate genes, such as DYRK4 as a novel candidate for back fat thickness, whose high PIP_gene values warrant further investigation [47]. These findings collectively highlight the value of our fine-mapping framework for enhancing candidate gene discovery in livestock. Beyond gene-based applications, the aggregation framework underlying PIP_gene represents a broader methodological advance in genomic-feature posterior inclusion probabilities. While we focus on genes due to their well-defined boundaries and clear biological relevance, the same mathematical approach can be applied to any functionally defined genomic region, including regulatory elements, noncoding RNAs, or transposable elements, providing a flexible foundation for functional interpretation across diverse genomic contexts.

Despite the advancements presented, this study has limitations. A primary limitation is our simulation design. First, our evaluations focused on a single 4-Mb candidate region where we simulated only one to three variants as having larger causal effects. This, combined with an infinitesimal background modeled by 10 000 genome-wide chip SNPs, represents one specific model of genetic architecture; real traits exhibit greater heterogeneity in the number, effect size distribution, and genomic location of causal variants. Furthermore, the GRM used to model this infinitesimal background was constructed directly from these 10 000 known infinitesimal-effect SNPs and was then supplied as input to GRM-dependent fine-mapping methods. This perfect concordance is an idealization not reflective of real-world scenarios, where GRMs are estimated from available marker data (often LD-pruned) and may not fully capture true genetic covariance or perfectly align with the underlying polygenic architecture. We further investigated the impact of using a more “practical” GRM, constructed from 10 000 randomly selected quality-controlled genome-wide variants (from the ~11.7 million). Re-analyzing all simulation scenarios with FINEMAP-adj and SuSiE-adj using this practical GRM yielded fine-mapping performance highly concordant with that obtained using the “simulation-derived” GRM (Supplementary Fig. S7 and Supplementary Table S6). This result suggests that while GRM misspecification remains a theoretical concern, our -adj approaches demonstrate robustness to this aspect of GRM construction, at least with reasonably dense marker-based GRMs. While our study effectively demonstrates the advantages of relatedness-aware methods under the tested architectures, further research exploring diverse genetic complexities will enhance understanding of performance boundaries in real-world scenarios. Secondly, the current implementations, particularly those involving matrix operations on the GRM or relatedness-adjusted LD, may face scalability challenges with datasets involving hundreds of thousands of individuals, a scale becoming increasingly common, requiring future algorithmic refinements. Thirdly, our current PIP_gene formulation, which sums posterior model probabilities for models including any variant within a gene region, can be systematically influenced by gene length and variant density. Under the standard fine-mapping prior that assigns equal a priori causal probability to each variant, longer genes with more variants inherently have greater opportunities to accumulate signal, leading to higher PIP_gene values that may not solely reflect biological enrichment for causality. For applications requiring correction, post-fine-mapping adjustment can be performed by regressing PIP_gene values against gene length or variant count, excluding genes with PIP_gene above a threshold (e.g. 0.05) to preserve potential true biological signals, and using the regression residuals as length-adjusted PIP_gene values. Future refinements could explore incorporating more sophisticated priors that account for gene size or variant density to further enhance the specificity of gene-level inference. Fourthly, our real data analysis relied on imputed sequence data for SNPs and indels; other structural variations were not considered, and differential imputation quality across variants could introduce unmodeled uncertainty.

Future research could focus on several exciting avenues. Incorporating functional annotation data into the prior model probabilities for variants or genes [17, 48–50] holds considerable promise for further enhancing the statistical power and accuracy of our methods. Our framework’s inherent handling of population stratification through the GRM and fixed effects provides a foundation for extensions to complex structured populations. While our current single-population approach for multi-breed populations assumes shared causal variants with identical effects, leveraging breed-specific LD patterns to discriminate causal variants, future extensions could relax the assumption of identical substitution effects by modeling different effect sizes across breeds, or account for breed of origin of alleles using crossbred animals [51]. Additionally, while this study focused on fine-mapping GWAS signals for complex traits, the principles of accounting for relatedness (particularly through adjusted LD matrices and mixed models) are broadly applicable. Investigating the performance and adaptation of our framework for other types of genetic analyses in samples of related individuals, such as fine-mapping expression quantitative trait loci (eQTLs) [52–55], represents a valuable direction.

In conclusion, this study provides a robust suite of fine-mapping tools (BFMAP-Forward, BFMAP-SSS, FINEMAP-adj, and SuSiE-adj) and a novel gene-level aggregation strategy (PIP_gene) specifically designed for and validated in populations with complex relatedness, such as livestock. By properly accounting for genetic relationships and whole-genome infinitesimal effects, these methods offer substantial improvements in accuracy and power over existing approaches, paving the way for more reliable identification of causal variants and genes underlying complex traits in these important populations.

Key Points

• BFMAP-SSS enables accurate fine-mapping in related individuals using individual-level LMMs.

• Novel -adj methods enable FINEMAP/SuSiE for related samples using adjusted LD and LMM-derived inputs.

• Gene-level PIPs improve detection power, vital for high-LD livestock genomes.

• Multi-breed populations improve fine-mapping accuracy over single-breed analyses.

Author contributions

J.J. conceived and designed the study and developed the associated methodologies and software tools. J. Wang conducted the simulation studies and pig trait analyses. F.T., Y.H., G.S., C.S. and C.M. contributed to data preprocessing. J. Wang, J.J. and J. Wei performed data visualization. J.J. and J. Wang wrote the manuscript. All authors reviewed and approved the final manuscript.

Conflict of interest: Y.H. is employed by Smithfield Premium Genetics, and G.S. and C.S. are employed by AcuFast LLC. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Funding

This work is supported by the Agriculture and Food Research Initiative (AFRI) Foundational and Applied Science Program, project award no. 2023-67015-39260, and the Research Capacity Fund (HATCH), project award no. 7008128, from the U.S. Department of Agriculture’s National Institute of Food and Agriculture.

Data availability

BFMAP can be downloaded at https://github.com/jiang18/bfmap. Tutorials for FINEMAP-adj and SuSiE-adj are available at https://github.com/JJWang259/FineMapping-RelatedIndividuals. The Duroc pig data used in this study are the property of Smithfield Premium Genetics (Rose Hill, NC, USA), and the Yorkshire pig data are the property of AcuFast LLC (Navasota, TX, USA). These data were used under license for the current study and are not publicly available due to access restrictions. Data access requests may be considered under negotiated research agreements and should be directed to Kent Gray, General Manager (kgray@smithfield.com) for Duroc data and Garrett See (Garrett.See@acufastswine.com) for Yorkshire data.

Biographical note

The authors represent expertise in quantitative genetics and animal breeding from academic institutions (North Carolina State University, University of Florence, and Wayne State University) and industry partners (Smithfield Premium Genetics and AcuFast LLC).