Both Alleles Scored 0.995. The Experiment Disagreed.

TL;DR for operators

A scientific AI system should not treat a computational prediction as a verified finding simply because an LLM can explain it or retrieve supporting annotations. ARGUS demonstrates an alternative architecture: prediction models generate hypotheses, an agent retrieves relevant observations, deterministic rules determine scientific status, and an LLM communicates conclusions within explicit evidence constraints.

The distinction changes outcomes. At one cancer-associated genetic variant, experimental observations revealed a binding difference missed by a saturated prediction model. Three other hypotheses remained unresolved, despite additional investigation. An optional LLM planner also reduced evidence-tool queries from 13 to 9 across five tested hypotheses without changing their verdicts.

For biomedical AI developers, the immediate lesson concerns the allocation of scientific authority: flexible evidence acquisition can coexist with rule-governed adjudication. The demonstrated benefits concern auditability and limited investigation efficiency, not established clinical accuracy or measured hallucination reduction.

Two predictions of 0.995 concealed a biological difference

A research team investigating a genetic variant must determine whether a change in DNA affects how a regulatory protein binds to it. A computational model can predict binding behavior for the original and altered DNA sequences, but the prediction alone cannot establish what occurs experimentally.

At the cancer-associated variant rs6983267, a sequence-based model assigned the transcription factor FOXA1 a binding probability of 0.995 for both alleles. These nearly identical predictions suggested retained binding without a detectable difference between the two DNA versions.

Yet independent experimental observations indicated significant allele-specific binding: an adjusted significance measure of FDR 0.030, supported by records covering 15 experiments and 475 reads.

This discrepancy is central to Unlocking the Regulatory Genome by ARGUS, by Dutta and colleagues.¹ Their framework classified the FOXA1 hypothesis as rescued because experimental evidence identified an allelic effect that the upstream prediction failed to distinguish.

The issue was model saturation. When both predictions approach the upper probability limit, the model may provide little discrimination between alternatives, even when an experimentally detectable difference exists.

Importantly, rescue did not involve retraining or correcting the underlying model parameters. It changed the evidence-supported interpretation of a particular hypothesis.

The same variant also produced three other unresolved transcription factor hypotheses. That contrast raises a more consequential question than whether an AI system can retrieve biological information: what evidence entitles the system to make a scientific claim?

ARGUS places scientific judgment outside the LLM

ARGUS is not a new genomic foundation model. It is an investigation architecture built around existing prediction models and independent evidence resources.

The upstream system uses 458 DNABERT-based transcription factor models to evaluate reference and alternative DNA sequences. Deterministic rules classify their predictions, including warnings when both binding probabilities exceed 0.95.

Each prediction initializes a hypothesis rather than becoming an accepted scientific conclusion.

The investigation then separates four responsibilities:

1. Prediction: Sequence-based models propose a possible binding effect.
2. Evidence acquisition: A planner selects relevant sources according to the current hypothesis state.
3. Verification: Deterministic rules assess observations, their provenance, and their implications for the hypothesis.
4. Reporting: An LLM explains the resulting evidence and verdict without authority to revise the scientific classification.

The operational evidence sources have different purposes.

ADASTRA supplies experimental allele-specific binding observations, including statistical significance and coverage. JASPAR provides computational motif scores that can support or challenge a predicted binding direction. ENCODE cCRE identifies regulatory-element context, but an overlapping regulatory annotation does not independently prove transcription factor binding.

This evidence hierarchy is enforced by the verifier, not left to the language model’s interpretation.

The system maintains an explicit hypothesis state containing the prediction, accumulated observations, current epistemic status, and investigation history. After each evidence query, the verifier updates that state. The next action can depend on what the previous query revealed.

This makes investigation conditional rather than a mandatory sequence of database lookups.

Possible outcomes include supported, contradicted, rescued, partially supported, and abstained. Direct experimental evidence can justify stronger verdicts; consistent indirect observations have more limited authority. Unresolved or conflicting evidence can terminate in abstention.

The architecture therefore constrains both what evidence can establish and what the reporter may claim.

Four hypotheses reveal why abstention matters

The study’s principal demonstration investigates four transcription factors at the same variant, rs6983267, using the default evidence-priority planner.

Source: Section 3.1, Table 1. FOXA1 terminates after three recorded trajectory steps; the other hypotheses each require eight. These step counts are not evidence-tool-call counts.

These outcomes illustrate three distinct reasons not to issue a definitive scientific interpretation.

For KLF6, the upstream model predicts stronger binding, but JASPAR suggests an effect in the opposite direction. ENCODE supplies enhancer-like regulatory context without settling that conflict. More contextual information does not eliminate disagreement between the computational signals.

For RAD21, the experimental observation has sufficient reported coverage but is statistically nonsignificant. An FDR of 0.65 does not demonstrate an allelic difference. Equally, it does not establish that the difference is absent. ARGUS therefore refrains from treating the observation as confirmation of the upstream non-differential prediction.

For SP1, direct experimental evidence is unavailable at the locus. Its motif information cannot adjudicate the non-directional retained-binding prediction. Regulatory context cannot fill that evidentiary gap.

FOXA1 and SP1 are especially instructive: both begin with saturated retained-binding predictions, yet only FOXA1 has decisive experimental observations.

The system’s different outcomes arise from the evidence available for each claim, not simply from the apparent confidence of the prediction.

Flexible planning removes four queries without changing verdicts

The authors also examine whether scientific investigation can become more selective without transferring biological judgment to an LLM.

The default planner follows a fixed priority: ADASTRA, JASPAR, then ENCODE cCRE, stopping when the hypothesis is resolved or the investigation budget is exhausted.

An optional Claude Sonnet 4.6 planner instead chooses evidence sources according to the hypothesis state. Both policies operate on identical frozen evidence snapshots, share the same deterministic verifier and stopping rules, and have a maximum of three evidence-tool calls per hypothesis.

For four unresolved hypotheses, the LLM planner omits ENCODE cCRE because regulatory context cannot resolve the remaining transcription-factor-specific binding question. FOXA1 requires only the decisive ADASTRA query under either policy.

Three independent LLM-planner runs reproduce the reported action sequences.

This is a small within-hypothesis planning comparison, not evidence of general LLM-planner superiority. It measures evidence-tool usage, not total execution cost, and does not comprehensively account for LLM token consumption or latency.

Still, it identifies a plausible design advantage: allow the agent discretion over which evidence to request, while keeping the rules governing scientific conclusions fixed.

Auditability is demonstrated; clinical reliability is not

For developers of genomic interpretation services, ARGUS suggests a concrete separation between predictive output, experimental verification, unresolved uncertainty, and reportable findings.

A research-facing product could expose the hypothesis trajectory, identify the observation supporting each verdict, distinguish computational agreement from experimental confirmation, and explicitly disclose unanswered questions. These are system-design implications drawn by Cognaptus, not measured downstream product outcomes.

The same principle could apply to other high-consequence workflows where retrieved information varies in evidentiary authority. The affected design decision is whether an agent can promote observations into conclusions or must pass them through a separately governed verification layer.

Several boundaries remain material.

The evaluation includes only five hypotheses across two loci, with four detailed trajectories at one variant. It does not provide a population-scale accuracy benchmark, prospective clinical validation, or controlled measurement of hallucination reduction. The upstream predictions are precomputed rather than evaluated as part of a complete end-to-end workflow.

The operational verifier draws on three evidence sources. Relevant biological information outside those sources, including tissue-specific evidence and other experimental modalities, may alter interpretation. The authors’ separate AlphaGenome Atlas cross-reference supplies additional computational predictions but does not participate in ARGUS verdicts.

Finally, deterministic verification constrains scientific state and reporting permissions; it does not by itself guarantee the factual accuracy of every sentence generated downstream. The paper does not systematically evaluate that remaining reporting risk.

Scientific authority is the central design decision

ARGUS demonstrates an agent that can investigate a prediction, encounter contradictory or inconclusive observations, and preserve uncertainty rather than forcing a definitive answer.

Its strongest result is architectural: scientific classification remains governed by explicit rules while evidence acquisition can adapt to observations. The FOXA1 rescue shows why independent verification matters, and the three abstentions show why retrieving additional information is not equivalent to establishing a finding.

For organizations building scientific AI systems, the question is not merely how much evidence an agent can retrieve. It is whether the system can explain, enforce, and audit the conditions under which that evidence becomes a conclusion.

Cognaptus: Automate the Present, Incubate the Future.