Positional cloning finds the gene behind a trait by following which stretches of chromosome travel with the trait from parents to offspring. Crossovers during meiosis cut those stretches into smaller pieces, and the search narrows until one gene is left. The method often stalls a step short of that. Where crossovers rarely happen, or where the family cannot be enlarged, the search ends with a window of roughly five to twenty genes. Every variant in that window is inherited together with the trait, so statistics cannot tell them apart. Choosing which gene to knock out or sequence first then comes down to lab intuition or to ranking tools that depend on known gene–disease links, and those tools fail exactly when the gene is new.
We propose to settle that last step with a Delphi panel made of language-model agents. A Delphi panel is a structured way of polling experts: each one rates the options alone and anonymously, sees only a summary of the group's ratings plus anonymized reasons, and then revises or defends their rating over a few rounds. Our version adds two constraints. First, each agent sees only one kind of evidence, such as expression, model-organism phenotypes, protein constraint and structure, or predicted variant consequence. Second, the genes are anonymized, so agents must reason from the evidence and cannot recall a published answer. The claim is that this blinded, evidence-split Delphi ranks the causal gene higher than knowledge-graph tools on novel genes. It must also beat a simple ensemble of independent model samples given the same compute. Finally, the spread of the panel's final ratings should tell the lab when the shortlist can be trusted.
Why this gap, and what is already known
The mapping literature in the file shows where the search stalls, but it offers no method for the stalled step. The Hairy glume gene in wheat sits at the distal end of chromosome 1AS, a region of chromosome variation and suppressed crossing over (Fine mapping of the Hairy glume (Hg) gene in a chromosome variation region at the distal terminus of 1AS). Human pedigrees such as the consanguineous high-myopia family (Clinical and linkage study on a consanguineous Chinese family with autosomal recessive high myopia.) or the multi-generation otosclerosis families (A pathogenic deletion in Forkhead Box L1 (FOXL1) identifies the first otosclerosis (OTSC) gene) cannot be made larger. These papers establish the problem. They do not supply a method for it.
The strongest evidence for the gap comes from Simulation of undiagnosed patients with novel genetic conditions. It builds 42,680 simulated patients whose candidate lists average about 14 genes, which is the size of a stalled interval. It then tests Phrank, LIRICAL and HiPhive against a knowledge base frozen at 2015. The tools work on known conditions but degrade on novel ones: for a novel disease with a novel gene, the best method placed the causal gene at an average rank of 5.99. That cohort is our main testbed, and its frozen knowledge base is what makes it usable.
Two agentic genomics papers supply technique but stop short of our question. AIVARI Agent runs one agent per case over many gene–disease hypotheses, with retrieval from OMIM and ClinVar. It reaches group sensitivity of 0.956, but its row precision of only 0.351 floods reviewers with false alarms, and it never challenges its own inclusions. VarLitBench and VarLitAgent shows that models can tell whether published evidence points toward pathogenic or benign (F1 0.979 with o4-mini) but cannot grade how strong it is. Both work with variants that have already been named. Neither does discovery within an interval, and neither uses a panel.
Two multi-agent papers constrain the design. Voting or Consensus? Decision-Making in Multi-Agent Debate finds three things. Independent first drafts, which correspond to Delphi round one, add 3.3% on average. Consensus protocols help on factual tasks. Long debates drift from the task. Emergence of Biased Consensus in Multi-Agent LLM Debates shows that groups of agents at low sampling temperature lock into a biased answer within one or two rounds. Mixing temperatures between 1.3 and 1.7, or mixing model families, prevents this. We adopt both mitigations. Neither paper touches biological evidence, so whether these dynamics carry over is itself something we test.
The Impact of Stability Considerations on Genetic Fine-Mapping ranks variants by whether their posterior stays high across five GEUVADIS populations, and finds that stable variants score better on 361 of 378 functional annotations. This motivates our use of agreement across independent evidence views as a signal of causality. That method, however, needs many unrelated individuals in which variants can be separated. A stalled Mendelian interval has a single shared haplotype, so all its variants co-segregate and that kind of statistic does not apply.
Core Claim
The claim has three parts.
- Ranking. In a stalled interval with a novel causal gene, a blinded Delphi panel ranks the causal gene better than knowledge-graph prioritizers.
- Feedback matters. The gain comes from the feedback rounds, not only from pooling independent opinions. Full Delphi should beat round-one-only aggregation at equal token budget.
- Calibrated stopping. The panel's final disagreement on the top gene is a calibrated stopping signal. It sets a shortlist size that keeps sensitivity close to AIVARI-like levels while greatly improving precision.
If the second part fails, the honest conclusion is that ensembling suffices. We would report that.
Research Program
Evidence dossiers. For each candidate gene in each case we build a dossier from time-frozen sources matching the benchmark's 2015 freeze. Each dossier has four parts:
- tissue expression;
- orthology and model-organism phenotypes;
- constraint, protein-domain and structure summaries;
- the predicted consequence of the observed variant.
Gene symbols, protein names and any literature identifiers are replaced by random codes. Phenotype descriptions stay as free text for the patient but are never linked to a gene by name.
The panel. The panel has eight to twelve agents. Each agent is assigned one dossier part, so two or three agents share each evidence view. Agents are drawn from at least two model families and use temperatures spread across roughly 1.3 to 1.7, following the heterogeneity result above.
The rounds.
- Round one: each agent rates every candidate from 1 to 9 for its likelihood of causing the phenotype and writes a short rationale. No agent sees any other agent's output.
- Coordination between rounds: a coordinator program (not an agent) computes the median and interquartile range for each gene and compresses the rationales into anonymized arguments for and against.
- Rounds two and three: agents see their own prior rating next to the group summary. Any agent whose rating falls outside the interquartile range must either revise it or justify it in writing.
We cap the panel at three rounds because of the drift finding.
Retention rule. Genes are retained using a content-validity ratio, the standard Delphi rule for keeping an item that enough panelists consider essential:
$$\mathrm{CVR}_g = \frac{n_g - N/2}{N/2}$$
Here $N$ is the number of panelists, $n_g$ is the number who rate gene $g$ at 7 or above in the final round, and $\mathrm{CVR}_g$ ranges from −1 to 1. The output is a ranked list ordered by median rating, with ties broken by smaller interquartile range. A shortlist consists of genes whose $\mathrm{CVR}_g$ exceeds a threshold, and the threshold is tuned on held-out cases. The interquartile range of the top gene is reported as the confidence signal.
Baselines
Do-nothing reference. We rank candidates uniformly at random and by gene length. With about 14 candidates, the expected rank under random ordering is about 7.5. This floor shows how much any method adds.
Strongest existing published method. This is the best of Phrank, LIRICAL and HiPhive, as reported by Simulation of undiagnosed patients with novel genetic conditions. That paper is the only source in the file that benchmarks prioritizers on novel genes in candidate lists of interval size. We rerun these tools on our splits of the same simulated cohort with the same 2015 knowledge base. Beating them on the novel-disease, novel-gene subset means a mean causal-gene rank below the reported 5.99, with the difference significant under a paired bootstrap over patients. We also report top-1 and top-3 accuracy.
Strongest model-based competitor. This is a single frontier model that receives all four dossier parts for all candidates. It is run with self-consistency, meaning many independent samples whose ratings are averaged. This is the cheap alternative a reviewer would reach for. The number of samples is set so that total tokens match the full Delphi run. Prompt tuning gets the same budget of held-out development cases as the panel.
Round-one-only Delphi. The same blinded, evidence-split, heterogeneous panel is stopped after round one and aggregated by median. This is the "All-Agents Drafting without feedback" condition from Kaesberg et al. It isolates the key ingredient, controlled feedback. Full Delphi must beat it at matched tokens, with the extra samples in the round-one condition making up for the feedback rounds.
Open debate with majority vote. The same agents discuss freely without anonymity or interquartile-range rules and then vote. This tests whether the Delphi structure matters beyond having several agents.
Further ablations. We remove gene masking, merge the evidence views so every agent is a generalist, and make all agents share one model family at temperature 0.3.
Shortlist comparison. We compare shortlists against AIVARI's operating point of 0.956 group sensitivity and 0.351 row precision. That figure comes from a different institutional cohort that we cannot access, so it is a reference point, not a head-to-head result. Within our data, we compare shortlist precision at fixed 0.95 sensitivity against a single-agent inclusion rule run the way AIVARI runs.
Why not just…?
Why not just grow the mapping population until one gene remains? In cold spots such as the 1AS terminus, crossovers inside the interval are rare no matter how many progeny are screened, and human pedigrees cannot be enlarged. This approach works outside cold spots. Our method is for the cases where it has already been tried and failed.
Why not just use fine-mapping statistics such as PICS on the interval? Those methods separate variants using differences in how variants are correlated across many unrelated people. In a single segregating haplotype every variant co-segregates perfectly, so posteriors are flat. The stability work (The Impact of Stability Considerations on Genetic Fine-Mapping) needs multiple populations, which a family-based interval does not have.
Why not just use Phrank, LIRICAL or HiPhive? On novel genes they reach only an average rank of 5.99, because the knowledge-graph edges they rely on do not exist yet. This baseline is in the main comparison. If the panel cannot beat it on the novel subset, the proposal fails.
Why not just ask one strong model many times and average? This is the real threat. The panel's gains might come from sampling diversity and nothing else. The matched-token self-consistency baseline and the round-one-only ablation settle the question directly. If either matches full Delphi, we would recommend the cheaper method and report that controlled feedback adds nothing for this task.
Why not just let the models debate and vote? Open debate drifts over rounds (Voting or Consensus?) and locks into early bias at low temperature (Emergence of Biased Consensus). Neither failure has yet been shown on biological evidence, so the open-debate baseline measures both directly. We track the rank of the causal gene from round to round, and we check whether panels that start with a wrong top gene ever recover.
Proposed Experiments
Main testbed: the simulated cohort. The primary benchmark is the simulated undiagnosed cohort, split by known versus novel disease and gene. Its candidate lists stand in for stalled intervals. The primary metrics are mean causal rank, top-1 accuracy and top-3 accuracy on the novel-disease, novel-gene subset. Masking is essential because the models were trained after 2015 and may know the "novel" genes. As a leakage check, we compare masked and unmasked runs. A large unmasked advantage would mean the unmasked models are recalling answers, not reasoning, and that only masked numbers count.
Plant test set. The second test set reconstructs final intervals and candidate lists from fine-mapping studies in the file:
- rice rl2 (Rolling Leaf 2);
- wheat RHT25 (Wheat plant height locus RHT25);
- watermelon dg (Genetic mapping and molecular characterization of the delayed green gene dg);
- rice CYP76M7/M8 (Deletion of Diterpenoid Biosynthetic Genes);
- eggplant anthocyanin (Fine Mapping of Candidate Gene Controlling Anthocyanin Biosynthesis);
- sorghum Tan1/Tan2 (Identification and analysis of novel recessive alleles).
This set is small and these genes are published, so it serves as a masked transfer check from human phenotypes to plant traits, not as a headline result. No reported result found on Lacuna for any prioritizer on these plant intervals.
Calibration. We bin cases by the final interquartile range of the top gene and plot top-1 accuracy per bin. The stopping rule is useful only if accuracy rises steadily as disagreement shrinks.
Decisive checks. Four results together decide whether the proposal holds:
- Full Delphi beats round-one-only aggregation and matched-token self-consistency on novel-gene rank.
- Masked performance beats 5.99 on the novel-disease, novel-gene subset.
- The homogeneous low-temperature panel shows the early lock-in that the heterogeneous panel avoids.
- Shortlists keep 0.95 sensitivity at clearly higher precision than single-agent inclusion.
References
- Fine mapping of the Hairy glume (Hg) gene in a chromosome variation region at the distal terminus of 1AS
- Clinical and linkage study on a consanguineous Chinese family with autosomal recessive high myopia.
- A pathogenic deletion in Forkhead Box L1 (FOXL1) identifies the first otosclerosis (OTSC) gene
- Simulation of undiagnosed patients with novel genetic conditions
- AIVARI Agent: An Evidence-Grounded Agentic LLM for Variant Reportability and Interpretation
- VarLitBench and VarLitAgent for Benchmarking and Automating LLM-Assisted Functional Evidence Curation in Genomic Variant Interpretation
- Voting or Consensus? Decision-Making in Multi-Agent Debate
- Emergence of Biased Consensus in Multi-Agent LLM Debates
- The Impact of Stability Considerations on Genetic Fine-Mapping
- Rolling Leaf 2 Controls Leaf Rolling by Regulating Adaxial-Side Bulliform Cell Number and Size in Rice
- Wheat plant height locus RHT25 encodes a PLATZ transcription factor that interacts with DELLA (RHT1)
- Genetic mapping and molecular characterization of the delayed green gene dg in watermelon (Citrullus lanatus)
- Deletion of Diterpenoid Biosynthetic Genes CYP76M7 and CYP76M8 Induces Cell Death and Enhances Bacterial Blight Resistance in Indica Rice ‘9311’
- Fine Mapping of Candidate Gene Controlling Anthocyanin Biosynthesis for Purple Peel in Solanum melongena L.
- Identification and analysis of novel recessive alleles for Tan1 and Tan2 in sorghum