The problem
More than 90% of disease-associated variants from genome-wide association studies sit in noncoding regulatory DNA, and explaining what they do remains a central open problem in genomic medicine. Large language models asked to interpret these variants often invent transcription factor binding changes, cite experimental support that does not exist, and read meaning into negligible signals.
In an early prototype, an LLM reported that SP1 showed "a gain in binding" at rs6983267 because the ratio of two probabilities was large. Both probabilities (0.007 and 0.0009) were far below the binding threshold: there was no binding at either allele, so nothing to gain or lose.
How it works
Deterministic science, agentic reasoning. All biological classification is done by rule-based, auditable code. The LLM may choose the next investigation step or write the final report, but it never classifies evidence or assigns a verdict.
- Predict. DeepVRegulome scores the reference and alternate alleles with 458 DNABERT transcription factor models. A deterministic classifier labels each result (loss or gain, strengthened or weakened, retained, or no confident binding) and flags saturated predictions.
- Plan. For each hypothesis, a planner picks the next evidence source based on current uncertainty: allele-specific binding from ADASTRA, motif scoring from JASPAR, or regulatory context from ENCODE cCRE. The planner can be rule-based or LLM-mediated.
- Verify. A deterministic verifier interprets each observation and updates the hypothesis. Only direct experimental evidence can mark a claim as supported, contradicted or rescued.
- Stop or abstain. When no evidence source can resolve the claim, a stopping rule assigns partial support or abstains. What ARGUS observes changes what it does next.
- Report. An LLM writes the narrative from pre-classified facts and cannot override a verdict or claim experimental confirmation the evidence does not provide.
Key results
At rs6983267, a well-studied cancer risk variant at 8q24, the same planner produced four different investigations for four transcription factors, all from real evidence queries.
| TF | Prediction | Steps | Outcome | Why |
|---|---|---|---|---|
| FOXA1 | Retained (saturated) | 3 | Rescued | ADASTRA shows significant allele-specific binding (FDR 0.030, 15 experiments), revealing a model false negative hidden by saturation. |
| KLF6 | Strengthened | 8 | Abstained | No ADASTRA data at the variant; the JASPAR motif opposes the prediction while ENCODE cCRE supports it. Mixed indirect evidence. |
| RAD21 | Retained (saturated) | 8 | Abstained | A powered but nonsignificant allelic test (FDR 0.65, 5 experiments), which does not establish the absence of an effect. |
| SP1 | Retained (saturated) | 8 | Abstained | No direct experimental evidence at this locus, so no claim is promoted to a finding. |
- FOXA1 and SP1 enter with the same saturated prediction. FOXA1 is rescued because experimental evidence exists; SP1 abstains because it does not.
- An LLM-mediated planner reached the same verdicts as the fixed planner with 9 tool calls instead of 13, by skipping evidence that could not resolve the claim.
- Frontier LLMs prompted directly gave confident answers that mixed general 8q24 biology with variant-specific claims, and none abstained (an informal observation, not a benchmark).
Papers
Team
Pratik Dutta (lead), Matthew B. Obusan, Max Chao, Rekha Sathian, Nimisha Papineni, Ramana V. Davuluri
Support
ARGUS is supported by the Anthropic AI for Science program, awarded to the Davuluri Lab at Stony Brook University.
Cite
@article{dutta2026argus,
title = {Unlocking the Regulatory Genome by ARGUS: An Evidence-Constrained
Agentic Framework for Interpreting Single Nucleotide Variants},
author = {Dutta, Pratik and Obusan, Matthew B. and Chao, Max and
Sathian, Rekha and Papineni, Nimisha and Davuluri, Ramana V.},
journal = {arXiv preprint arXiv:2610.12281},
year = {2026}
}