Researchers at Stony Brook University built ARGUS, an AI system that checks a sequence model’s predictions about DNA variants against existing biological evidence before generating a report. In a small test, data from past experiments revealed a FOXA1 binding difference that the underlying model could not resolve; ARGUS abstained on three other proteins at the same variant when the evidence remained uncertain. The authors list the Oct. 8 arXiv preprint as accepted at the NeurIPS 2026 Workshop on Agentic AI for Biological Discovery.
DNA variants are single-letter changes. Some occur in noncoding regulatory DNA, which can help control gene activity. Transcription factors are proteins that bind DNA and help regulate genes. Sequence models can predict whether a variant changes this binding, but a prediction does not show what experiments found.
ARGUS begins with 458 DNABERT-based models that predict binding for selected transcription factors. It compares the reference and alternate versions of a variant. If both predicted binding scores are above 0.95, it marks the result saturated: both scores are near the model’s top range, where their similarity may hide a difference between the DNA versions.
The system then retrieves three kinds of evidence: existing experimental allele-specific binding data from ADASTRA; computational DNA-motif scores from JASPAR; and regulatory-region annotations from ENCODE. A deterministic verifier applies fixed rules to those results. An LLM writes the final report from labeled findings. An optional Claude Sonnet 4.6 planner can choose the next evidence source but cannot change the evidence or the verifier’s verdict.
At rs6983267, a variant in the 8q24 cancer-risk locus, the sequence model predicted FOXA1 binding probabilities of 0.995 for each DNA version. Existing ADASTRA ChIP-seq data, pooled from 15 experiments, favored the reference version (475 reads; FDR-adjusted p=0.030). ARGUS marked this a rescued prediction: the experimental data showed an allelic difference that the saturated model output treated as equal. The team queried existing data; it did not conduct new lab experiments for this demonstration.
ARGUS did not make a binding claim for three other transcription factors at the same variant. ADASTRA had no KLF6 result for that SNP; a JASPAR motif score pointed opposite the model’s prediction, while ENCODE supplied regulatory context only. For RAD21, an ADASTRA test passed the study’s coverage filter but was not significant (five experiments, 104 reads; FDR-adjusted p=0.65). SP1 had no direct experimental result at that SNP, and its motif score could not resolve a model prediction that did not specify which allele should bind more. ARGUS abstained in all three cases. That means the available evidence did not settle those specific claims, not that the proteins cannot bind there.
In a separate comparison across five hypotheses at two variants, a fixed-order planner and the LLM planner reached the same verdicts. The LLM planner used nine evidence-tool calls instead of 13, skipping ENCODE lookups that could not resolve transcription-factor-specific binding. Its results were the same across three runs on frozen evidence snapshots. The authors describe this as a small demonstration of tool selection, not evidence of a general planning advantage.
The evaluation covers five hypotheses at two loci, starts from precomputed model predictions, and is neither an accuracy benchmark nor a controlled test of reduced hallucinations. The paper’s comparison with three general-purpose LLMs was also informal. It is an arXiv version 1 preprint; clinical use and independent replication were not evaluated.
Sources
Full paper and methods; ARGUS code repository; Figure 1 from Dutta et al., arXiv:2610.12281v1, CC BY 4.0.