Imagine trying to identify a single suspect in a crowd where everyone is dressed identically and moves in perfect sync. This is essentially the challenge geneticists face when they try to pinpoint which specific gene variant causes a disease. Nearby variants on a chromosome are so tightly linked that they almost always show up together, making it nearly impossible to tell which one is truly responsible and which ones are simply "guilty by association." A new statistical method called SuSiE-RSS is helping researchers cut through this genetic noise, and it does so using only publicly shared summary statistics rather than hard-to-obtain patient-level data. Developed by a team at the University of Chicago and Columbia University, this approach is reshaping how scientists narrow down the true genetic culprits behind complex diseases.
The Problem With Genetic Neighbors
When scientists run a genome-wide association study, they scan thousands of genetic variants to see which ones correlate with a trait or disease. The trouble is that variants sitting close together on a chromosome tend to be inherited as a block, a phenomenon called linkage disequilibrium. Because of this, dozens of variants in a region might all appear statistically linked to a disease, even though only one or two of them actually cause it. Separating the true causal variants from their innocent neighbors, a process known as fine-mapping, has long depended on having access to the complete genotype and phenotype data of every study participant. Getting that level of data is often difficult, both logistically and for privacy reasons, which is why many datasets are shared only as summary statistics instead.
Rebuilding SuSiE for Summary Data
A few years ago, a method called the Sum of Single Effects, or SuSiE, offered a simpler and faster way to fine-map causal variants using full individual-level data. It works by treating the overall genetic effect as a sum of several individual "single effects," each representing one likely causal variant, and then fitting these one at a time through an iterative process. The catch was that SuSiE still needed raw genotype and phenotype records to run.
The researchers behind this new work extended SuSiE so it can operate on summary data instead, calling the result SuSiE-RSS, with RSS standing for "regression with summary statistics." Their key insight was that the mathematics behind SuSiE only truly requires certain combined quantities, known as sufficient statistics, rather than the raw data itself. When those exact sufficient statistics happen to be available, SuSiE-RSS reproduces precisely the same results as running SuSiE on the original data. When they are not available, the method approximates them using widely shared resources: single-variant association scores, known as z-scores, and a correlation matrix describing linkage disequilibrium, typically estimated from a reference panel of similar genetic ancestry.
Cleaning Up Messy Data
One particularly practical contribution of this work is a set of built-in diagnostics for spotting errors before they distort the results. A common and stubborn problem in fine-mapping is the "allele flip," where the same genetic variant is recorded using opposite allele codes in the study data and in the reference panel used for linkage estimates. Left unnoticed, a single flipped allele can throw off correlation estimates for dozens of nearby variants, leading the algorithm to mistakenly flag a harmless variant as causal. The team demonstrated this danger directly: in a simulated dataset with one true causal variant among 1,002 candidates, a single mismatched allele caused the method to wrongly assign a mismatched variant a near-certain probability of being causal. By plotting each variant's observed association score against what the model expects it to be, the researchers showed that flipped alleles stand out clearly as outliers, allowing them to be caught and removed before they mislead the analysis.
The team also built in a refinement step that helps the algorithm escape false solutions it sometimes settles into, and a regularization procedure that adjusts the estimated correlation matrix when it does not perfectly match the true underlying data.
How It Stacks Up
Using genotype data from 50,000 UK Biobank participants across 200 genomic regions, the researchers compared SuSiE-RSS against three established fine-mapping tools: FINEMAP, DAP-G, and CAVIAR. When given accurate, in-sample correlation data, all four methods performed similarly well. But real-world studies almost always rely on external reference panels rather than in-sample data, and it was here that differences emerged. SuSiE-RSS matched or outperformed the alternatives while running about as fast as FINEMAP and DAP-G, and dramatically faster than CAVIAR, which became impractically slow as it checked every possible combination of causal variants. The quality of the reference panel used for correlation estimates mattered more to overall accuracy than which specific method was chosen, underscoring that a large, well-matched reference sample remains essential regardless of the statistical tool applied.
Looking Ahead
Fine-mapping remains one of the trickiest steps in translating genetic association signals into real biological insight, largely because of the entangling effects of linkage disequilibrium. The core problem is that scientists rarely have access to the complete patient-level data needed for accurate causal variant detection, forcing them to rely on shared summary statistics instead. This creates a pressing need for methods that can work reliably with incomplete information while still guarding against the data quality issues, like allele flips and mismatched reference panels, that plague real-world genetic datasets. SuSiE-RSS addresses this need directly, offering a tool that produces individual-level-quality results from summary data alone whenever the underlying data are consistent, while also flagging when they are not. For researchers hunting the genetic roots of conditions ranging from heart disease to psychiatric disorders, a faster and more trustworthy way to separate true causal variants from their genetic neighbors could meaningfully speed up the path from statistical association to biological understanding. Future work refining how these methods handle very large effect sizes and non-standard study designs, such as case-control or mixed-model analyses, will likely extend their reach even further.