Unlocking the Regulatory Genome by ARGUS: An Evidence-Constrained Agentic Framework for Interpreting Single Nucleotide Variants

1 Introduction

• Noncoding regulatory variants can affect transcription factor binding, chromatin accessibility, or enhancer-promoter contacts, making functional interpretation a GWAS-to-mechanism bottleneck.A typical GWAS locus contains dozens of variants in linkage disequilibrium.Source: §1 · p. 1
• Prediction and interpretation are distinct: interpreting a predicted binding change requires tissue context, disease relevance, independent evidence, and uncertainty assessment.AlphaGenome Atlas precomputes predictions for all 9 billion possible single nucleotide variants, but does not distinguish model predictions from independent support.Source: §1 · p. 2
• An LLM prototype called SP1 binding gained from a log-odds ratio of 3.02, although both allele probabilities were below the 0.5 active-binding threshold.The reference and alternate probabilities were 0.007 and 0.0009, respectively; neither indicated binding.Source: §1 · p. 2
• ARGUS keeps biological classification in auditable rule-based code while allowing the LLM to select evidence-gathering actions; observations change the next investigation step.The LLM cannot classify evidence or assign verdicts, which are constrained by pre-classified facts.Source: §1 · p. 2

2 Architecture

• ARGUS has a deterministic prediction stage and an agentic investigation stage that subjects predictions to evidence-based scrutiny.The prediction stage is shared with the DeepVRegulome pipeline.Source: §2 · p. 2

2.1 Prediction Stage

• For each variant, ARGUS selects TF models from 458 DNABERT-based models and classifies allele probabilities using a 0.5 active-binding threshold.The classifier distinguishes threshold-crossing changes, quantitative shifts, retained binding, and no confident binding; it flags saturation when both probabilities exceed 0.95.Source: §2.1 · p. 2

2.2 Investigation Stage: The Agentic Loop

• The investigation is a stateful loop in which a planner selects evidence based on current uncertainty and a deterministic verifier updates the hypothesis state after each observation.The planner's priority policy tries ADASTRA first, then JASPAR and ENCODE cCRE if the hypothesis remains unresolved, and abstains when no admissible source remains.Source: §2.2 · p. 3
• ADASTRA tables match 445 of the 458 TF models, defining the current scope of direct experimental falsification.JASPAR motif evidence uses a directional relative-score threshold of 0.05 and cannot independently establish or refute experimental binding.Source: §2.2 · p. 4
• The verifier treats ADASTRA findings according to significance, power, direction, and saturation; JASPAR is indirect computational evidence, while ENCODE cCRE supplies regulatory context.The stopping rule can assign partially_supported or abstained, but reserves supported, contradicted, and rescued verdicts for direct experimental evidence.Source: §2.2 · p. 4

2.3 Evidence-Constrained Reporter

• The reporter generates narrative from pre-classified facts and cannot alter evidence, verdicts, state transitions, or claim experimental confirmation when disallowed.The authors note that constrained state transitions do not guarantee factuality of unrestricted downstream prose.Source: §2.3 · p. 4

3.1 Divergent Trajectories at rs6983267

• FOXA1 was rescued after significant ADASTRA binding evidence (FDR = 0.030; 15 experiments) exposed a possible model false negative, while KLF6, RAD21, and SP1 abstained after eight-step investigations.KLF6 had conflicting motif and cCRE evidence; RAD21 had a coverage-qualified nonsignificant test (FDR = 0.65); SP1 lacked direct experimental evidence at the locus.Source: §3.1 · p. 5

3.2 Observation-Dependent Branching

• The same planner produced structurally different trajectories because intermediate observations differed, continuing when evidence left a hypothesis unresolved and stopping when it was sufficient.The FOXA1 and KLF6 trajectories took three and eight steps, respectively, under the same code path.Source: §3.2 · p. 6

3.3 Frontier LLM Baseline

• Frontier LLMs confidently conflated 8q24/MYC biology with variant-specific TF binding claims and did not distinguish tissue expression from allele-specific binding at this position.This was an informal observation, not a controlled benchmark.Source: §3.3 · p. 6

3.4 AlphaGenome Atlas Cross-Reference

• AlphaGenome’s variant-level AVI score was 0.570, but its TF-specific predictions were excluded from planner and verifier decisions because the score was identical across TFs and could not drive observation-dependent branching.The queried TF tracks showed substantial biosample dependence.Source: §3.4 · p. 6

3.5 Fixed vs. LLM-Mediated Planning

• The LLM planner reached identical verdicts in all five cases using 9 calls rather than the fixed planner’s 13, declining cCRE queries that could not resolve TF-specific binding or direction.LLM planner sequences were identical across three replicates; the comparison does not establish a general planning advantage.Source: §3.5 · pp. 6–7, Table 2 · p. 7

4.1 Architecture Enforces What Prompting Cannot

• Architectural enforcement, rather than prompting, keeps biological classification deterministic and prevents LLMs from overriding pre-classified evidence.The LLM may select investigation actions or write reports, but cannot classify evidence or assign verdicts.Source: §4.1 · p. 7

4.2 Prediction vs. Interpretation

• ARGUS addresses the downstream question of whether a predicted variant effect is supported, rather than competing with upstream prediction models.Its planner, verifier, and stopping rules are model-agnostic.Source: §4.2 · p. 7

4.3 Falsification as a Design Pattern for Agentic Science

• ARGUS treats predictions as testable hypotheses, using explicit epistemic states and data-structure flags to distinguish computational evidence from experimental claims.States communicate what has been tested and what remains unknown; experimental claims require direct experimental evidence.Source: §4.3 · p. 7

4.4 Limitations

• The evaluation covers five hypotheses at two loci, not predictive accuracy, controlled hallucination reduction, or end-to-end natural-language inference.Initial DVR outputs are precomputed.Source: §4.4 · p. 8
• The planner uses a fixed evidence priority rather than a learned or value-of-information policy, while the falsification battery covers only three evidence sources.The paper identifies adaptive planning and additional sources such as eQTL, ATAC-seq, literature, and independent model predictions as possible extensions.Source: §4.4 · p. 8
• DNABERT blind spots include saturation and limited sensitivity outside core motifs; interpretation also depends on tissue matching, assay aggregation, and allele harmonization.Saturation rescue addresses the most severe blind spot, while systematic characterization of model failures remains ongoing.Source: §4.4 · p. 8
• The local ENCODE cCRE lookup uses the same underlying data as SCREEN but lacks its biosample-specific activity annotations.Source: §4.4 · p. 8

• Four divergent trajectories at rs6983267 show that observation-dependent investigation can be achieved within an auditable framework using real evidence.ARGUS separates deterministic science from agentic reasoning and makes each verdict traceable to a specific evidence observation.Source: §5 · p. 8