Over the past two decades, Genome-Wide Association Studies (GWASs) have become one of the most powerful tools in human genetics, successfully linking thousands of genetic variants - known as Single Nucleotide Polymorphisms (SNPs) - to complex traits and diseases such as obesity, diabetes, heart disease, and Alzheimer's disease. The basic idea of a GWAS is to scan the genomes of large numbers of individuals and identify regions where genetic variation correlates with a particular trait of interest.
However, a major challenge remains: identifying which specific genetic variant within an associated region is actually responsible for influencing the trait - that is, identifying the true causal variant. This challenge arises largely because of a phenomenon called Linkage Disequilibrium (LD). LD refers to the tendency of nearby variants on the same chromosome to be inherited together across generations. As a result, many variants in a genomic region tend to be strongly correlated with one another, making it difficult to distinguish the truly causal variant from its neighbors that are merely "along for the ride."
To tackle this problem, a suite of statistical methods known as genetic fine-mapping has been developed. These methods go beyond simple association testing and aim to assign well-calibrated probabilities - called Posterior Inclusion Probabilities (PIPs) - to each variant in a candidate region, reflecting the likelihood that the variant is the true causal one. Fine-mapping methods also typically output credible sets: small groups of variants that together have a high probability (e.g., 95%) of containing at least one causal variant. The smaller and more precise these credible sets are, the more actionable and interpretable the results become for downstream biological investigation.
Despite significant progress in fine-mapping methodology, most existing approaches treat genetic fine-mapping as a purely statistical problem - relying solely on the patterns of LD and association statistics. A growing body of evidence, however, suggests that integrating functional genomic information can substantially improve fine-mapping accuracy. In particular, research has shown that trait-associated SNPs are significantly more likely to be expression Quantitative Trait Loci (eQTLs) - that is, variants that influence the expression level of nearby genes. This creates an opportunity: by leveraging eQTL data as a functional prior, we can guide fine-mapping toward variants that are more biologically plausible causal candidates.
This is precisely the motivation behind SuSiE², introduced in this paper by Zhang, Jiang, and Zhao from Yale University's Department of Biostatistics. SuSiE² is a new two-layer statistical framework that integrates eQTL information into genetic fine-mapping, built on top of the widely used Sum of Single Effects (SuSiE) model.
The SuSiE Foundation
Before describing SuSiE², it is important to understand the underlying SuSiE model (Wang et al., 2020), upon which the new method is built. SuSiE stands for Sum of Single Effects and represents a novel approach to Bayesian variable selection in linear regression - a problem that is central to genetic fine-mapping.
In fine-mapping, the goal is to identify which of the many candidate SNPs in a region are causally related to the trait. Traditional methods like CAVIAR and FINEMAP model this directly, but at significant computational cost. SuSiE takes a different approach: it models the total genetic effect on a trait as the sum of multiple "single effects," where each single effect captures exactly one causal signal. This structure leads to an efficient Iterative Bayesian Stepwise Selection (IBSS) algorithm that is both computationally fast and statistically interpretable.
The output of SuSiE includes a PIP for each variant (representing the probability that it is causal) and a set of credible sets - one per inferred causal signal - each containing a small group of highly correlated candidate variants. A key feature of SuSiE is that it can handle regions with multiple independent causal signals simultaneously, without requiring the user to specify the exact number of signals in advance.
By default, SuSiE assigns uniform prior probabilities to all variants - meaning that before seeing the data, every variant is considered equally likely to be causal. SuSiE² improves on this by replacing the uniform prior with an informative prior derived from eQTL data.
The SuSiE² Two-Layer Framework
SuSiE² connects two SuSiE models in a principled two-stage pipeline:
• Layer 1 - eQTL-based SuSiE: In the first layer, SuSiE is applied to gene expression data, treating the expression levels of a candidate risk gene as the response variable and the genotypes of nearby SNPs as predictors. This eQTL analysis produces PIPs for each variant reflecting their probability of influencing gene expression. These are referred to as PIPe (eQTL PIPs).
• Layer 2 - Trait-based SuSiE with eQTL priors: In the second layer, a standard SuSiE model is run on the GWAS summary statistics for the complex trait of interest. Crucially, rather than using uniform priors, the PIPe values from Layer 1 are fed in as prior inclusion probabilities. This means that variants with strong eQTL signals are given a head start - they are considered more likely to be causal before even examining the GWAS data for the trait.
The authors prove formally that, under the assumption that genetic variants influence the phenotype through gene expression, the final PIPs produced by SuSiE² (PIPt) are exactly equal to the posterior probabilities of causality given both the phenotype and gene expression data jointly. This gives SuSiE² a strong theoretical justification beyond being a heuristic combination of two models.
An important practical advantage of the two-layer design is its flexibility. Because the two stages are modular, they can use different LD reference panels and different sample populations - unlike most multi-trait fine-mapping methods that require all traits to be measured in the same individuals with the same LD matrix. SuSiE² can also work with summary statistics rather than individual-level data, making it applicable to the growing catalog of publicly available GWAS summary statistics.
Simulation Studies
To evaluate SuSiE², the authors conducted a comprehensive set of simulations designed to mimic real fine-mapping studies. Genotype data was drawn from 10,000 randomly selected European individuals from the UK Biobank (UKBB), and fine-mapping regions from chromosome 1, each containing 5,000 SNPs, were used as the simulation landscape.
Three scenarios were tested: (a) "All causal SNPs are eQTLs" - where all trait-associated variants also influence gene expression; (b) "Some causal SNPs are eQTLs" - a more realistic scenario where only a subset of causal SNPs are eQTLs; and (c) a high-complexity scenario with ten risk genes and thirty total causal SNPs. Performance was measured across three criteria: Power (the proportion of true causal SNPs captured in at least one credible set), Coverage (the proportion of credible sets containing at least one true causal variant), and Average CS Size (the average number of variants per credible set).
SuSiE² was compared against four other methods: single-trait SuSiE (baseline), mvSuSiE, flashfm, and fastPAINTOR. The results were consistent and favorable for SuSiE² across both the in-sample LD matrix setting and when using an external reference panel (the 1000 Genomes Project Phase 3 European samples).
With the in-sample LD matrix, SuSiE² achieved 15–40% higher power than single-trait SuSiE and outperformed all competing multi-trait methods, including a 5% improvement over the second-best method, mvSuSiE. It maintained coverage close to the target 95% level and reduced the average credible set size by approximately 36%, compared to only 5% for flashfm and 34% for mvSuSiE. fastPAINTOR produced the smallest credible sets but at the cost of severely degraded coverage and power, suggesting it was overfitting.
With an external reference panel, all methods showed reduced performance - as expected when the LD information is imprecise - but SuSiE² remained the most competitive. It maintained higher power than single-trait SuSiE across all heritability settings and showed better coverage than mvSuSiE, which suffered from inflated false discoveries under inconsistent LD. An additional simulation showed that even when eQTL data came from a separate, smaller sample (2,000 individuals), SuSiE² still improved both power and coverage over standard SuSiE, demonstrating its robustness to differences in study design.
Real Data Applications
Body Mass Index (BMI)
The authors applied SuSiE² to fine-map genetic variants associated with Body Mass Index (BMI) using summary statistics from the UK Biobank GWAS (approximately 360,000 participants). eQTL priors were derived from GTEx v8 gene expression data in two adipose tissues: subcutaneous adipose (ADS) and visceral adipose (ADV), with sample sizes of 581 and 469 individuals, respectively.
From 714 candidate fine-mapping regions, standard SuSiE identified 449 credible sets with an average size of 4.80 SNPs. When SuSiE² was applied with ADS eQTL priors, the number of credible sets increased to 480, the count of single-SNP credible sets rose from 138 to 165, and the average credible set size dropped to 3.86. Results with ADV were similar (461 total credible sets, average size 3.82), with the median size reduced from 3 to 2 SNPs in both tissue cases. These results held consistently across chromosomes, demonstrating that eQTL information from relevant adipose tissues systematically improves fine-mapping resolution for BMI.
Two illustrative examples were highlighted. In a region on chromosome 1 containing the MRPS14 and CACYBP genes - both previously associated with BMI - standard SuSiE identified only one credible set. After integrating ADS eQTL data, SuSiE² identified two additional credible sets pointing to specific functional variants in CACYBP and MRPS14. In a region on chromosome 6 near SNRPC, SuSiE produced a broad credible set of 28 SNPs with no clear leading variant. SuSiE² narrowed this to just 4 SNPs with a well-defined lead variant (rs9462015), substantially increasing the precision and interpretability of the result.
Alzheimer's Disease (AD)
SuSiE² was also applied to Alzheimer's disease using summary statistics from a large-scale meta-analysis of over 1.1 million individuals. eQTL priors were estimated using the ROSMAP dataset, which provided bulk RNA-sequencing data from 473 individuals with both genotype and gene expression measurements in brain tissue.
The analysis focused on 20 AD risk loci with known likely causal mediators (functional SNPs predicted from single-cell epigenomic data). SuSiE² successfully identified 9 out of 20 mediators within a credible set, compared to just 5 identified by standard SuSiE - a substantial improvement in biological discovery. In the non-coding region specifically, SuSiE² captured 6 mediators versus 3 for SuSiE. Additionally, SuSiE² produced more credible sets overall (29 vs. 27) with a smaller average size (8.0 vs. 9.6 SNPs), and it reduced the 75th percentile of credible set size from 13 to 11 SNPs - suggesting it avoided producing extremely large, uninformative credible sets.
Notable examples included the PICALM locus, where eQTL priors from the brain tissue data allowed SuSiE² to elevate the PIP of the true mediator enough to include it in a credible set - something standard SuSiE failed to do. In contrast, at loci where eQTL signals were weak or the sample size was too small to produce informative priors (e.g., TNFRSF21, TMEM139, NGFR), SuSiE² produced results identical to standard SuSiE, showing that the method gracefully falls back to the baseline when eQTL information is uninformative.
Strengths, Limitations, and Future Directions
Key strengths of SuSiE² include:
• Statistically principled integration: The two-layer design is formally justified as computing the posterior probability of causality given both GWAS and eQTL data.
• Flexibility: The two stages can use different datasets, sample sizes, and LD reference panels - unlike multi-trait methods that require matched samples.
• Robustness: SuSiE² maintains good performance even with small eQTL datasets and with external LD reference panels, and degrades gracefully when eQTL information is weak.
• Compatibility with summary statistics: The method works with publicly available GWAS summary statistics, making it broadly applicable without requiring access to individual-level data.
Limitations and open challenges acknowledged by the authors include:
• Gene and tissue selection: Users must decide which genes and tissues to include when computing eQTL priors, which may introduce subjectivity or require additional preprocessing steps.
• Parameter sensitivity: The parameter K (number of assumed causal signals per region) can affect results in real data applications, and better automated selection strategies are needed.
• Limited to gene expression: The current framework uses only eQTL information, but other molecular QTLs (e.g., protein QTLs, chromatin accessibility QTLs) could further enhance performance.
Future directions suggested include extending SuSiE² to jointly integrate eQTLs from multiple tissues simultaneously, and incorporating additional types of molecular QTL data to capture a broader range of disease mechanisms.
Conclusion
SuSiE² represents a significant methodological advance in the field of genetic fine-mapping. By embedding eQTL functional information directly into the fine-mapping prior in a statistically rigorous way, it bridges the gap between genetic association and biological causation more effectively than prior approaches. Evaluated across a range of simulation scenarios and applied to two important complex traits - BMI and Alzheimer's disease - SuSiE² consistently outperformed standard SuSiE and competing multi-trait methods on power and precision, while maintaining appropriate statistical coverage.
For the broader genomics and computational biology community, this work underscores a key principle: leveraging multi-omics data in a principled, layered fashion can unlock discoveries that single-data-type analyses miss. As eQTL resources like GTEx continue to grow in depth and tissue coverage, frameworks like SuSiE² are poised to become increasingly valuable tools in the translation of GWAS findings into biological and clinical insights.