A new AI framework for reading cancer-linked DNA mutations is built to flag uncertainty rather than fabricate an answer.
ARGUS (Agentic Regulatory Genomics for an Uncertainty-aware Scientist) pairs 458 DNABERT-based transcription factor binding models with an LLM planner, while a separate deterministic verifier, not the LLM, decides what each piece of evidence actually means. Researchers ran it on rs6983267, a variant at the 8q24 cancer risk locus, across four transcription factors, and got four different outcomes. FOXA1 was rescued in three steps once real ADASTRA allele-specific binding data (15 experiments, FDR = 0.030) revealed the underlying prediction model had produced a false negative masked by signal saturation. The other three were not so lucky: KLF6 abstained after eight steps of mixed evidence across ADASTRA, JASPAR, and ENCODE cCRE data, RAD21 abstained after an ADASTRA allelic test came back nonsignificant (5 experiments, FDR = 0.65), and SP1 abstained outright because no direct experimental evidence exists at that locus despite sharing FOXA1's saturated prediction.
Over 90% of GWAS-linked disease variants sit in noncoding regulatory DNA, exactly the territory where LLMs asked to interpret binding changes tend to fabricate citations and overstate flimsy statistics. ARGUS's bet is that keeping the hallucination-prone model away from the actual biology, and only letting it choose which test to run next, produces answers that are honestly uncertain more often than confidently wrong. The researchers also found their LLM planner matched a fixed-priority search strategy's verdicts while using fewer tool calls, because it learned to skip evidence that could not resolve the question at hand.
A tool that abstains three times out of four is not flashy, but in a field where confident wrong answers send labs down dead-end experiments, three honest shrugs beat one more hallucinated binding site.