ARGUS: evidence-constrained agentic reasoning for noncoding regulatory variant interpretation

Dutta et al. couple 458 DNABERT models with deterministic biological verifiers, preventing LLM hallucinations when prioritizing noncoding single nucleotide variants.

Primary source: arXiv:2610.12281: https://arxiv.org/abs/2610.12281

What’s new: Over 90% of disease-associated single nucleotide polymorphisms identified by genome-wide association studies reside in noncoding regulatory regions rather than protein-coding sequences. Identifying which mutations alter transcription factor (TF) binding or disrupt enhancers remains a major challenge in genomic medicine. While large language models can ingest biomedical literature, models prompted directly on regulatory variants routinely hallucinate TF binding alterations, fabricate experimental support, and assign biological significance to statistically negligible signals. In Unlocking the Regulatory Genome by ARGUS: An Evidence-Constrained Agentic Framework for Interpreting Single Nucleotide Variants (arXiv:2610.12281, 8 Oct 2026), Dutta et al. present ARGUS (Agentic Regulatory Genomics for an Uncertainty-aware Scientist), strictly separating deterministic biological computation from high-level reasoning.

ARGUS wraps 458 specialized DNABERT models of transcription factor binding into a closed-loop investigation workflow. An LLM planner selects evidence sources based on residual uncertainty, while a deterministic verifier interprets database outputs: including real ADASTRA allele-specific binding experiments, JASPAR motif analysis, and ENCODE candidate cis-regulatory elements (cCREs). Evaluating variant rs6983267 at the 8q24 colorectal and prostate cancer risk locus, the system demonstrates calibrated decision-making across four candidate transcription factors. ARGUS rescues FOXA1 binding in three steps when empirical ADASTRA data (15 experiments, false discovery rate = 0.030) overturns an initial model false negative caused by predictor saturation. For KLF6 and RAD21, the system traverses up to eight steps before correctly abstaining due to mixed indirect evidence or non-significant experimental results, avoiding overconfident false classifications.

Why it matters: Genomic foundation models output raw binding scores that frequently conflict with cell-type-specific experimental assays. By using deterministic verification to gate every evidentiary step and enforcing explicit abstention when evidence is equivocal, ARGUS demonstrates how agentic frameworks can assist molecular geneticists without introducing generative hallucinations into clinical pipelines.

Caveats: The framework currently relies on pre-trained DNABERT sequence models and requires access to curated functional genomics databases (ADASTRA, JASPAR, ENCODE). Genomic loci lacking experimental ChIP-seq or chromatin accessibility data will trigger abstentions rather than definitive classifications.