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

The 30-second take

• What: ARGUS separates deterministic regulatory-genomics computation from LLM planning: 458 DNABERT TF-binding models sit inside a hypothesis-directed loop with a planner, a verifier that interprets each observation deterministically, and real ADASTRA, JASPAR, and ENCODE queries — demonstrated on rs6983267 with divergent, evidence-constrained verdicts across four TFs.
• Why it matters: Interpreting the noncoding genome is still an elite, error-prone service, and fluent AI makes it worse by fabricating evidence. An agent that must abstain is a step toward cheaper, more trustworthy variant interpretation as a default layer — if the same discipline holds beyond one locus.
• Who should care: Genomic-medicine and regulatory-genomics groups, AI-for-biology teams tired of TF hallucinations, and anyone building tool-using agents that should know when to stop.

What the paper actually did

Over 90% of disease-associated GWAS variants fall in noncoding regulatory regions, yet functional interpretation remains open. Large language models asked to interpret such variants routinely hallucinate transcription-factor binding changes, fabricate experimental support, and over-weight tiny signals. ARGUS (Agentic Regulatory Genomics for an Uncertainty-aware Scientist) strictly separates deterministic biological computation from LLM-mediated reasoning. It wraps 458 DNABERT-based TF-binding models in a hypothesis-directed loop: a planner picks evidence sources from current uncertainty, a verifier deterministically interprets each observation, and intermediate results change the path. On rs6983267 at the 8q24 cancer risk locus, the same planner produces four divergent trajectories for four TFs. FOXA1 is rescued in 3 steps when real ADASTRA allele-specific binding data (15 experiments, FDR = 0.030) reveals a model false negative masked by saturation. KLF6 traverses 8 steps across ADASTRA, JASPAR motif analysis, and ENCODE cCRE annotation before abstaining on mixed indirect evidence. RAD21 abstains in 8 steps after ADASTRA returns a coverage-qualified but nonsignificant allelic test (5 experiments, FDR = 0.65). SP1, which shares FOXA1's saturated retained prediction, abstains because no direct experimental evidence exists at this locus. All observations come from real ADASTRA, JASPAR, and ENCODE cCRE queries; none are simulated. A comparison of fixed-priority and LLM-mediated planning shows the LLM planner reaches identical verdicts with fewer tool calls by declining evidence that cannot resolve the claim under test.

What makes this disruptive

The scarce capability is honest interpretation of noncoding risk variants. Unconstrained LLMs fail in a specifically dangerous way: they invent TF stories. ARGUS's split — deterministic compute, LLM only as a planner, verifier that can force abstention — is a governance pattern as much as a genomics method. The rs6983267 case study is sharp because the same planner both rescues a saturated-model false negative (FOXA1) and refuses three other popular-looking TFs. Scarcity under pressure: expert genomic judgment that today sits in well-funded labs. One locus is not a genome-wide product, but an agent that will say “not enough evidence” is the disruptive social contract.

Why it matters (outside the lab)

Abundance lens: reading the regulatory genome is an elite service, and naive AI threatens to make bad readings abundant. Constrained agents are how cheaper interpretation could become a default without becoming fiction. Near-term, treat ARGUS as a design pattern for tool-using biology agents. Mid-horizon, clinic and regulation still sit in front of any default variant report. No year. The abundance we want is trustworthy calls and principled abstentions, not more confident paragraphs.

Limitations & open questions

The worked example is one famous variant and four TFs. 458 DNABERT models can still be wrong; FOXA1 needed experimental rescue of a saturated false negative. Abstention is only as good as the verifier's rules and the databases queried. Preprint ≠ product; this is not a clinical interpreter. Abundance is not automatic. GWAS-to-function remains open; ARGUS organizes evidence, it does not close biology. Read the PDF for planner prompts, verifier logic, and how far beyond rs6983267 they tested.