Most genetic variants linked to disease sit in non-coding DNA, the part that does not code for proteins but controls when and where genes switch on. Finding which variant in a correlated block actually causes an effect is called statistical fine-mapping. It works much better when each variant gets an informative prior, a starting probability of being causal that is built from biological annotations. The difficulty is that we now have several strong but disagreeing sources for that prior. These include sequence-to-function neural networks such as Enformer and AlphaGenome, DNA language models, models trained on reporter assays, and tissue chromatin maps. Today a pipeline picks one prior and reports one posterior probability per variant. Nobody can tell whether a confident causal call comes from the genetic data or from whichever annotation the analyst happened to choose.
We propose to replace the single prior with a set of plausible priors. Each variant gets an interval of prior probability, shaped by how much the functional predictors disagree and capped by hard biological rules. We then compute, exactly and cheaply, the lowest and highest posterior inclusion probability (PIP: the posterior chance that the variant is causal) that any prior in the set allows. A variant whose lower bound is high is a robust causal call, and a variant with a wide interval is a call that depends on the prior. The core claim is testable. At equal numbers of calls, variants ranked by the lower bound should replicate on independent functional evidence more often than variants ranked by any single-prior or averaged-prior PIP.
Core Claim and Mechanism
The closest existing work is the Expression Modifier Score study (Leveraging supervised learning for functionally informed fine-mapping of cis-eQTLs). It trains a random forest on more than 6,000 features to separate high-PIP from low-PIP GTEx eQTLs, which are variants statistically linked to a gene's expression level. It then feeds the score into SuSiE, a Bayesian fine-mapping method that models a locus as a sum of a few single causal effects, as a per-variant prior. That study motivates this proposal in two ways. It shows that informative priors matter: it reports 20,913 additional putative causal eQTLs across 49 tissues and 310 additional candidate trait genes. It also exposes the risk. Distance to the gene's start site and Basenji scores account for over 98% of the predictor's performance in whole blood. The training labels come from fine-mapping the same data, so a biased prior could manufacture confident calls that SuSiE cannot correct. We keep SuSiE and its likelihood unchanged. What changes is the prior, which becomes a credal set (a convex set of probability vectors), and the output, which becomes an interval.
The mechanism follows from the form of SuSiE's single-effect posterior. For one effect, let $b_k$ be variant $k$'s Bayes factor (how much better the data fit if $k$ is causal), and let $\pi_k$ be its prior weight. The posterior that variant $j$ is that effect is
$$\alpha_j(\pi) = \frac{\pi_j b_j}{\sum_k \pi_k b_k}.$$
This is a ratio of two linear functions of $\pi$. Suppose $\pi$ is constrained to a polytope, for example per-variant bounds $l_k \le \pi_k \le u_k$ plus $\sum_k \pi_k = 1$. Then the maximum and minimum of $\alpha_j$ are linear-fractional programs, and the standard Charnes–Cooper change of variables turns each into an ordinary linear program. For box constraints the solution is a simple greedy rule. To maximise $\alpha_j$, set $\pi_j$ to its upper bound and push the remaining mass onto the variants with the smallest Bayes factors; to minimise it, do the reverse. The extremes are therefore exact and cheap for each single effect. This sidesteps the scaling problems that make general credal inference expensive.
For SuSiE with several effects, the overall PIP is $1-\prod_l(1-\alpha_{lj})$. We bound it by solving the per-effect programs inside SuSiE's fitting loop, holding the other effects' residuals fixed. We then check that approximation against brute-force enumeration over sampled priors on small loci.
Two pieces of prior work supply the vocabulary rather than the method. Logical Credal Networks represent interval-bounded conditional probabilities and answer queries with nonlinear solvers. We borrow the idea of interval-valued beliefs, but our structure is much narrower, so we need only linear programs. Probabilistic Reasoning with Inconsistent Beliefs Using Inconsistency Measures separates non-negotiable integrity constraints from soft beliefs that may conflict. We adopt that split. Hard constraints become fixed caps; for example, a variant in chromatin that is closed in every reference epigenome for the tissue lineage (Integrative analysis of 111 reference human epigenomes) gets an upper bound on its prior. Predictor disagreement only widens the soft interval. POPE supplies the guiding question: does an optimum sit on a robust plateau or a fragile spike? Our PIP interval asks the same question about each causal call.
Research Program
The first step builds calibrated per-variant priors from $K\approx5$ sources. These are the Expression Modifier Score, Enformer variant-effect magnitudes (Effective gene expression prediction from sequence), AlphaGenome tissue-matched expression effects (AlphaGenome), and the zero-shot log-likelihood ratio from the DNA language model GPN (DNA language models are powerful predictors). The fifth source, MPAC predicted allelic skew (Identifying non-coding variant effects at scale), is used only in configurations not evaluated on MPRA labels, for the circularity reasons given under Proposed Experiments. Each source is mapped to a prior scale by fitting its enrichment among held-out-chromosome fine-mapped eQTLs. This mirrors how the Expression Modifier Score was scaled, so every source enters on the same footing.
The second step defines two credal sets, chosen so that they can be told apart. The hull set is the convex hull of the $K$ calibrated prior vectors. Because the extremes of a linear-fractional function over a polytope lie at its corners, this set needs only $K$ SuSiE runs. The box set gives each variant an interval spanning its own calibrated values across sources, widened slightly, then intersected with the hard chromatin caps. The box allows combinations that no single source proposes, which the hull does not. Whether that extra freedom helps or only adds conservatism is an empirical question, and we test it directly.
The third step computes PIP intervals genome-wide on GTEx v8 cis-eQTLs for tissues that match available reporter-assay cell lines. Liver is matched to HepG2. Whole blood is used as the closest available proxy for K562, since K562 is a leukaemia line rather than a whole-blood tissue. We then sort credible-set variants into robust, prior-driven, and data-driven groups according to interval width and location.
Baselines
The naive reference point is SuSiE with a uniform prior. It shows how much any functional prior adds. Every method is run on the same LD reference, the same summary statistics, and the same number of effects $L$.
The strongest existing method is SuSiE with the Expression Modifier Score prior, run exactly as in its paper on the same loci. The file reports discovery counts for it: 20,913 additional putative causal eQTLs across 49 tissues and 310 additional UK Biobank candidate genes. However, I found no reported result on Lacuna for its precision against independent reporter-assay validation. We will therefore measure that ourselves. For context on the predictors, AlphaGenome reports eQTL sign-prediction auROC rising from 0.75 to 0.80, and Enformer reports better causal-variant prioritisation in 47 of 48 GTEx tissues. Neither paper reports fine-mapping precision with its scores used as a prior.
The most dangerous competitor is the averaged prior: SuSiE run once with the arithmetic mean of the $K$ calibrated priors. A close relative is posterior averaging over the $K$ single-prior runs. Both use the same inputs and the same compute as our hull set.
Two ablations isolate the key ingredient. Using the hull set alone tests whether box freedom matters. Using the box set without the hard chromatin caps tests whether the integrity constraints matter.
Beating the baselines means the following. At matched numbers of calls (top 500, 1,000, and 5,000), robust calls ranked by lower-bound PIP must show significantly higher validation precision than calls ranked by the Expression Modifier Score PIP and by the averaged-prior PIP. They must also keep at least 70% of those baselines' validated discoveries. A pure precision gain obtained by calling almost nothing does not count.
Why not just…?
Why not just use a better single prior, such as AlphaGenome's predictions? A newer predictor may well rank variants better than the Expression Modifier Score. It still gives one answer with no warning when that answer is wrong. More importantly, each single-prior run is one corner of our hull set, so the comparison is built into the design. If one source's PIP ranking matches the precision of the robust lower bound at every cutoff, the interval adds nothing for ranking. Its remaining value would then be limited to flagging prior-driven calls, and we will report that outcome as such.
Why not just average the priors or the posteriors? This is the real threat. Averaging is the obvious way to combine disagreeing sources, and it may capture most of the gain. The difference is that an average hides disagreement, while the lower bound penalises it. Our decisive test compares two groups with the same averaged PIP, one with narrow intervals and one with wide intervals. If they validate at the same rate, interval width carries no information. In that case the averaged prior is the better recommendation, and the box machinery should be dropped.
Why not just run a sensitivity analysis over a handful of priors? For the hull set, that is exactly what our method reduces to, and we say so. The linear-program formulation only earns its place for the box set, whose corners cannot be enumerated: with $M$ variants there are about $2^M$ of them. The hull-versus-box ablation decides whether that larger set is worth having.
Why not just learn annotation weights from the data with hierarchical Bayes? If the data pinned down the weights tightly, a credal set would be unnecessary. However, the training labels for such weights come from the same fine-mapping. Labels in the Expression Modifier Score study are circular in exactly this way, and that circularity is what produces its proximity bias. We will report how much the PIP interval narrows when the evaluation tissue's own eQTLs are excluded from calibration. A strong narrowing would mean the data do constrain the prior and would favour this alternative.
Proposed Experiments
Validation must avoid circular labels. The primary truth set is MPRA allelic-skew calls, meaning variants whose two alleles drive measurably different reporter activity, in K562 and HepG2 from the MPAC training compendium. For context, MPAC reports a Pearson correlation of about 0.69–0.76 with allelic skew; this measures how predictable skew is from sequence and is not a fine-mapping result. MPAC is excluded from the prior in every configuration evaluated on these labels. Reporter assays miss native chromatin context, so we add two secondary checks. The first is agreement of eQTL effect direction with AlphaGenome-independent CAGI5 saturation-mutagenesis data where the variants overlap. The second is cross-tissue replication: a robust call in one tissue should reappear when a related tissue is fine-mapped with its own priors.
The decisive checks come in three parts. The first is precision at matched call counts against all baselines. The second is the matched-average-PIP comparison of narrow-interval versus wide-interval variants. The third is a proximity-bias audit, which compares how robust calls and Expression Modifier Score calls are distributed by distance to the transcription start site. If the interval mainly pulls down distal calls that only one sequence model supports, we will be suppressing long-range enhancers. Work such as A DNA foundation model predicts osteoporosis risk genes without proximity bias suggests that would be a mistake, and we would report it as a failure mode rather than a success. Computational overhead is measured against single SuSiE runs. Box bounds are greedy per effect, so the expected cost is a small multiple of one SuSiE fit per locus.
References
- Leveraging supervised learning for functionally informed fine-mapping of cis-eQTLs identifies an additional 20,913 putative causal eQTLs
- Logical Credal Networks
- Probabilistic Reasoning with Inconsistent Beliefs Using Inconsistency Measures
- Integrative analysis of 111 reference human epigenomes Open
- POPE: post optimization posterior evaluation of likelihood free models
- Effective gene expression prediction from sequence by integrating long-range interactions
- AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model
- DNA language models are powerful predictors of genome-wide variant effects
- Identifying non-coding variant effects at scale via machine learning models of cis-regulatory reporter assays
- A DNA foundation model predicts osteoporosis risk genes without proximity bias