🧬 genomics

SuSiNE: Genetic fine-mapping with signed functional priors and multi-basin ensembling

SuSiNE is an enhanced genetic fine-mapping method that extends SuSiE by incorporating signed functional priors and multi-basin ensembling to improve causal variant recovery, resolve linkage disequilibrium ambiguity, and provide robust diagnostics while avoiding the pitfalls of purity filtering.

Original authors: Callahan, M. G., Zhu, X.

Published 2026-08-06

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Callahan, M. G., Zhu, X.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a detective trying to solve a crime in a crowded city. You have a list of suspects, but here's the twist: many of them look exactly alike, wear the same clothes, and live in the same building. In the world of genetics, this is called "Linkage Disequilibrium." When scientists study how our DNA influences traits like height or disease risk, they find that certain genetic variants (the "suspects") are so tightly linked that they travel together from parent to child. It's like trying to figure out which twin actually stole the cookie when they both have chocolate on their faces and were in the kitchen at the same time.

To solve this, scientists use a method called "fine-mapping." Think of it as a high-tech interrogation room where they try to pinpoint the single, true culprit among the lookalikes. The current star detective in this field is a tool called SuSiE. It's fast and clever, using math to assign a "probability of guilt" to each suspect. But even the best detectives have blind spots. Sometimes, the interrogation room gets stuck on a false lead because the clues are confusing. Other times, the detective gets so focused on one theory that they miss other equally strong possibilities. And sometimes, the detective throws away a perfectly good clue just because it didn't look "clean" enough, even though it held the key to the case.

This paper introduces a new, upgraded detective team called SuSiNE. The authors realized that the old method was missing a crucial piece of information: the direction of the clues. Imagine if the witnesses didn't just say "I saw someone," but also said, "I saw someone running toward the exit" or "running away from it." SuSiNE can use these directional hints from advanced AI models (which predict how a genetic change affects the body) to narrow down the suspects. But the authors also discovered that the old detective's habit of throwing away "messy" clues was actually hurting the investigation. By combining the new directional hints with a strategy of running many different versions of the interrogation and then voting on the best result, SuSiNE solves the case much more often than the old method.

The Detective's Dilemma: Why the Old Method Stumbles

In the world of genetic fine-mapping, the goal is to find the specific DNA changes that cause a trait. The problem is that DNA variants are often correlated, like a group of friends who always hang out together. If you see one of them, you likely see the others. This makes it hard to tell who is actually responsible for the effect.

The popular tool SuSiE (Sum of Single Effects) tries to solve this by breaking the problem down. It assumes there are a few "single effects" (one culprit per effect) and tries to find them. It's great because it's fast and gives us a "Credible Set"—a shortlist of suspects that likely contains the true culprit. However, the authors found three main ways SuSiE can fail:

The "Lookalike" Trap (LD Ambiguity): When two suspects are identical twins, SuSiE might split the guilt 50/50 between them. It knows one of them did it, but it can't decide which one.
The "Stuck in a Rut" Trap (Optimizer Barrier): The math used by SuSiE is like a hiker trying to find the highest peak in a foggy mountain range. Sometimes, the hiker gets stuck on a small, local hill and thinks it's the top, missing the massive mountain nearby that is actually the true answer.
The "Too Picky" Trap (Purity Filtering): This was a surprising discovery. The old method had a rule: if a list of suspects (a "credible set") wasn't "pure" enough (meaning the suspects weren't very different from each other), throw the whole list away. The authors showed that this rule often throws away real clues. In their tests, using this rule actually made the detective worse at finding the true culprit, dropping the success rate from about 25% to 19%.

Enter SuSiNE: The Detective with a Compass

The authors created SuSiNE (Sum of Single Non-central Effects) to fix these problems. The biggest upgrade is a new way to use "functional annotations." These are clues from biological models (like AI that predicts how DNA changes affect gene expression) that tell us not just if a variant matters, but how it matters.

Imagine a witness saying, "The thief is wearing a red hat." The old method (SuSiE) could only use the fact that the thief wore a hat. It couldn't tell the difference between a red hat and a blue hat. SuSiNE, however, can use the color. If the AI predicts a variant will increase gene expression, and the data shows the trait is increasing, SuSiNE says, "Great match! This suspect is likely guilty." If the AI predicts an increase but the data shows a decrease, SuSiNE says, "Wait, that doesn't make sense. This suspect is probably innocent." The authors call this "self-gating": the model automatically checks if the clue matches the evidence before using it.

But SuSiNE doesn't just rely on one interrogation. To avoid getting stuck in a "local hill," the team runs the investigation many times with slightly different starting points and settings. This is the Ensembling part. They then use a smart voting system called Cluster-Weight Aggregation to combine all the results. Instead of picking just one "winner," they look at all the different theories that fit the data well and combine their insights. This ensures they don't miss a valid theory just because one run of the math got stuck.

What They Found: A Better Detective

The authors tested SuSiNE in two ways: with simulated data (where they knew the answer) and with real data from the GTEx Lung study.

In the Simulations:
When they used high-quality AI predictions (calibrated to match real-world performance), SuSiNE was a clear winner.

The Score: The old method (SuSiE) recovered the true causal variants with a score (AUPRC) of 0.2474. SuSiNE raised this to 0.3130.
The Gain: This is a 0.0656 improvement, which sounds small but is a massive 26.5% relative gain.
The Precision: At a high standard of 75% precision (meaning 3 out of 4 suspects are guilty), SuSiE found the true culprits 11.9% of the time. SuSiNE found them 19.3% of the time. That is a 61.7% relative increase in success.
The "Purity" Lesson: They confirmed that the old rule of throwing away "impure" lists was a mistake. Filtering by purity dropped the success rate from 0.248 down to 0.189. The authors suggest we stop using this filter as a default rule.

In the Real World (GTEx Lung Data):
They applied SuSiNE to 20 real gene locations.

Changing the Suspects: In 7 out of 20 locations, SuSiNE placed significant weight on the new AI-informed clues, changing which variants were highlighted as the most likely culprits.
The Cleanest Shift: The gene ARSA showed the clearest, most durable change. The AI clues helped the model find a better answer that stuck even when the clues were removed.
The Cautionary Tale: For the gene YDJC, the model shifted to a new suspect, but this coincided with a mismatch in the reference data (the "map" of the city didn't match the actual streets). This reminded the authors that while the method is powerful, it still needs to be careful about the quality of the background data.

The Verdict

The paper suggests that by adding "directional" clues from AI models and by running many different versions of the analysis to explore all possibilities, we can significantly improve our ability to find the true genetic causes of traits. The authors found that the old habit of discarding "messy" data was actually hurting us, and that a new, more flexible approach (SuSiNE) can find the true culprits much more often. While the results are based on simulations and a specific case study, the improvement is robust across different scenarios, suggesting that this new way of thinking about genetic clues is a promising step forward for the field.

Technical Summary of SuSiNE: Genetic Fine-mapping with Signed Functional Priors and Multi-basin Ensembling

Problem Statement
Genetic fine-mapping aims to identify causal variants within trait-associated loci but faces significant challenges due to linkage disequilibrium (LD) and the combinatorial complexity of causal configurations. The widely used SuSiE (Sum of Single Effects) method employs fast variational inference to produce posterior inclusion probabilities (PIPs) and credible sets. However, SuSiE has three distinct limitations:

LD Ambiguity: Highly correlated variants can provide interchangeable explanations, causing PIP mass to shift within a "basin" of the solution space without resolving the true causal variant.
Model-Specification Barrier: A single SuSiE fit produces a factorized posterior (an "AND-of-ORs" representation) that cannot adequately represent uncertainty across distinct, competing causal configurations (an "OR-of-ANDs" structure) if they reside in different basins of the evidence lower bound (ELBO) landscape.
Optimizer Barrier: The Iterative Bayesian Stepwise Selection (IBSS) algorithm can converge to a suboptimal local basin, missing the configuration that best fits the data.

Furthermore, the authors identify that the common practice of "purity filtering" (removing credible sets with low pairwise correlation) often discards informative signals where real effects are concentrated on weakly correlated variants, thereby reducing recovery performance.

Methodology: SuSiNE
The authors introduce SuSiNE (Sum of Single Non-central Effects), an extension of SuSiE that addresses these limitations through three core components:

Signed Functional Priors (The $\mu_0$ Channel):
Unlike existing methods that incorporate functional annotations only via prior inclusion probabilities ( $\pi$ ), which are unsigned, SuSiNE introduces a non-zero prior mean ( $\mu_0 = c \cdot a$ ) for variant effect sizes. Here, $a$ is a signed functional annotation (e.g., from sequence-to-function models like AlphaGenome) and $c$ is a scaling factor.
- Self-Gating: The single-effect Bayes factor in SuSiNE includes an alignment term. If the sign of the annotation agrees with the observed association direction, the evidence is reinforced; if they disagree, the annotation is automatically attenuated. This protects against annotation noise without requiring manual tuning.
- Preservation of Properties: The method maintains conjugacy, summary-statistic sufficiency, and monotone ELBO ascent, ensuring compatibility with existing SuSiE infrastructure.
Multi-basin Ensembling:
To overcome the optimizer and model-specification barriers, SuSiNE employs a grid-based ensembling strategy. It fits multiple models across a grid of perturbed initializations, prior scales, and prior variances. This allows the workflow to explore different regions of the ELBO landscape and discover multiple plausible basins.
Cluster-Weight Aggregation:
To summarize the ensemble without collapsing to a single "winner" (which would lose OR-of-ANDs uncertainty), the authors propose a cluster-weight aggregation rule.
- Fits are clustered based on a "credible-shift" PIP distance metric.
- Clusters are weighted by the maximum ELBO of their members, and weights are distributed within clusters using ELBO-softmax.
- This approach prevents a basin with many near-duplicate fits from dominating the summary, ensuring that distinct high-quality basins contribute to the final posterior.
New Diagnostics:
The paper introduces effect-level diagnostics to replace or supplement traditional credible set summaries:
- Concentration ( $K_\ell$ ): Effective support size based on Shannon entropy.
- Accuracy ( $A_\ell$ ): Ratio of the highest PIP among true causal variants to the global maximum PIP.
- Fitted-Basis Movement: Metrics to track whether the explanatory signal directions change across fits.

Key Results

Pathology and Purity Filtering: In controlled simulations, the authors demonstrate that purity filtering (threshold 0.50) discards genuine causal signals, reducing the Area Under the Precision-Recall Curve (AUPRC) for causal recovery from 0.248 to 0.189. They show that purity is not a reliable proxy for effect quality or causal coverage.
Baseline Performance (Single Fit): In oligogenic simulations calibrated to AlphaGenome eQTL benchmarks ( $\phi_a \approx 0.3$ ), the SuSiNE functional- $\mu$ channel (grid-searched) achieved a pooled AUPRC of 0.319, significantly outperforming the best annotation-agnostic SuSiE variant (0.251) and the functional- $\pi$ approach (0.270).
Ensemble Performance: The recommended workflow (C-CS: fixed-variance functional- $\mu$ grid with cluster-weight aggregation) raised the pooled AUPRC from 0.2474 (SuSiE-equivalent baseline) to 0.3130 (a 0.0656 absolute gain, 26.5% relative improvement). At 75% precision, recall increased from 11.9% to 19.3% (61.7% relative gain).
Robustness: Gains were robust across varying annotation qualities and alternative genetic architectures (sparse and diffuse), though gains were smaller in extremely sparse or diffuse regimes. The method showed sensitivity to "null contamination," where annotations aligned with association noise in non-causal variants could degrade performance, highlighting the importance of the self-gating mechanism.
Real Data (GTEx Lung): Applied to 20 GTEx Lung cis-eQTL loci, SuSiNE placed nontrivial weight on annotation-informed fits at 7 of 20 loci. It altered which variants received high PIPs in several cases. Notable shifts occurred at ARSA (durable shift) and YDJC (shift coinciding with reference-LD discrepancies). Internal diagnostics found little evidence of strong annotation confounding in the panel.

Significance and Claims
The paper claims that SuSiNE offers a practical framework for integrating directional functional predictions (such as those from modern AI sequence models) into fine-mapping without sacrificing the computational efficiency or theoretical guarantees of SuSiE.

Methodological Advancement: It demonstrates that signed priors can be incorporated via a prior-mean channel to provide "self-gating" against annotation noise, a capability missing in prior inclusion-based approaches.
Addressing Multimodality: By pairing grid-based exploration with cluster-weight aggregation, SuSiNE provides a mechanism to summarize uncertainty across competing causal configurations that a single variational fit cannot capture.
Re-evaluating Diagnostics: The authors argue that the standard purity filter is often counterproductive and propose effect-level diagnostics (concentration, accuracy, basis movement) for deeper model inspection.
Modest Scope: The authors emphasize that the real-data results are case studies demonstrating method behavior using reference LD, not definitive variant-level discoveries. They caution that performance gains depend on the quality of the annotations and the alignment between annotation errors and association noise, and that the method does not claim to solve the fundamental data limitations of LD.

The work is presented as a step toward utilizing AI-driven sequence models for fine-mapping, provided they are used carefully to avoid circularity and annotation-confounding pitfalls.

The Detective's Dilemma: Why the Old Method Stumbles

Enter SuSiNE: The Detective with a Compass

What They Found: A Better Detective

The Verdict

More like this