Agentic retrieval policies in biomedical multi-hop reasoning (SSE-Bio) can be formally verified via probabilistic Datalog to reduce hallucinated protein-disease links by 40%, using attention-derived co-expression signals from ELISA as grounding constraints.

When an agentic multi-hop biomedical reasoning system (SSE-Bio architecture) generates protein-disease link predictions, wrapping its retrieval policy in a probabilistic Datalog verification layer — constrained by attention-derived co-expression signals empirically grounded in ELISA measurements — reduces the rate of hallucinated (i.e., unsupported by any verifiable evidence chain) protein-disease link predictions by ≥40% relative to an unconstrained baseline agent, measured on a held-out benchmark of protein-disease claims with independently adjudicated ground truth (true/false/unknown), at equal or better recall on true positive links (recall degradation ≤5 percentage points).

Operationalized as two falsifiable sub-claims:

• H1 (Precision): Hallucination rate (FP / (FP+TP), where "hallucinated" = claim not entailed by any verified evidence path) drops from baseline rate B to ≤0.60·B.
• H2 (Non-degradation): True positive recall on the same benchmark does not drop by more than 5 points absolute vs. baseline.

The MS transcriptomics targets (DNMT1, ZNF740, CTSS, FGF2, SLCO2B1 and their claimed network edges to STAT3, IFNG, CSF1R/PTPRC, TREM2) serve as one of ≥3 real-world evaluation domains — used specifically because the underlying analysis is unpublished, internally scored, and contains at least one already-identified error (the superseded 36,966-cell figure), making it a natural adversarial test case for hallucination detection.

---

Design: 2×2 factorial, within-benchmark, repeated-measures comparison.

• Factor A: Verification layer (probabilistic Datalog ON vs. OFF)
• Factor B: Grounding signal source (ELISA-derived vs. RNA-seq-only vs. none)

Minimum viable test (MVT):

1. Build a 150-claim benchmark of protein-disease link assertions (50 true positive, 50 true negative/refuted, 50 currently-unknown/contested) drawn from: (a) the MS target hierarchy above (DNMT1–STAT3, ZNF740–IFNG/BRD2/3/4, CTSS–CSF1R/PTPRC, FGF2, SLCO2B1–TREM2), (b) OpenTargets high-confidence associations for 2 other diseases (rheumatoid arthritis, Alzheimer's) as generalization controls, (c) a synthetic "trap set" of 20 plausible-sounding but fabricated links (e.g., invented gene-disease pairs matching real gene/disease name distributions) to directly probe hallucination susceptibility.
2. Run SSE-Bio baseline agent (no verification) on all 150 claims, 5 runs/claim (temperature sampling) → baseline hallucination rate B.
3. Run SSE-Bio + probabilistic Datalog verification (ELISA-grounded) on identical claims/seeds → treatment hallucination rate T.
4. Run 2 ablations: (i) Datalog verification with RNA-seq-only grounding, (ii) simple rule-based whitelist filter (no probabilistic logic) — to isolate which component drives any observed effect.
5. Blind adjudication of all outputs by 2 independent biomedical curators against OpenTargets/STRING/CTD ground truth + manual literature check for contested claims; Cohen's κ reported.

---

# 1. Benchmark construction
claims = build_benchmark(
    sources=[MS_target_hierarchy, OpenTargets_RA, OpenTargets_Alzheimers, synthetic_traps],
    n_true=50, n_false=50, n_contested=50
)
ground_truth = adjudicate(claims, curators=2, arbiter=1)

# 2. Grounding signal construction
elisa_grounding = build_coexpression_graph(elisa_panel_data, method="attention_derived_weighting")
rnaseq_grounding = build_coexpression_graph(GSE193770, GSE138614, method="attention_derived_weighting")

# 3. Probabilistic Datalog verification layer
rules = """
0.0::link(P,D) :- true.
w1::link(P,D) :- coexpressed(P,G,W), W > theta1.
w2::link(P,D) :- network_proximity(G,Seed,S), S < theta2.
verified(P,D) :- link(P,D), evidence_chain_length(P,D,L), L =< 4.
"""
verifier = ProbLogEngine(rules, grounding=elisa_grounding)

# 4. Agent runs
for condition in [baseline, verified_elisa, verified_rnaseq, whitelist_filter]:
    for claim in claims:
        for seed in range(5):
            output = SSE_Bio_agent.run(claim, verifier=condition, seed=seed)
            log(output.prediction, output.evidence_chain, output.confidence)

# 5. Scoring
hallucination_rate = FP / (FP + TP)   # per condition, per domain
recall = TP / (TP + FN)
mcnemar_test(baseline_outputs, verified_outputs)
bootstrap_ci(relative_reduction, n=10000)

---