Forensic report: DualGPT-AB — a dual-stage generative optimization framework for therapeutic antibody design (10.1038/s43588-026-00976-0)

Academic fraud report · Geng Detector

Summary

Verdict: Highly suspicious (orange). This Nature Computational Science 2026 paper by Xie et al. presents a generative AI framework for antibody optimisation followed by wet-lab validation (SPR, ELISA, ADCC, CDC). Multiple high-severity concerns were identified. First, a duplicated P-value of 0.4183 appears for two different comparisons in Figure 6k, and the same value (clearly non-significant, >0.05) is cited in the text as evidence for P<0.05 in SK-BR-3 cells, a direct self-contradiction. Second, the SPR Methods state a flow rate of '30 μmol l−1', a concentration unit rather than the required volume/time (μl/min), consistent with LLM-style hallucination. Third, control-group death-rate values in Fig. 6j/l/m show suspiciously tidy decimal patterns. Fourth, the timeline—extensive wet-lab work from a computational group with a ~9-month revision window—raises plausibility questions. Confidence is moderate-high for findings 1–2; findings 3–4 are suggestive but not conclusive. No image manipulation was assessed. Independent raw-data inspection is required for confirmation.

Verdict

Highly suspicious. The paper contains an internally inconsistent P-value that contradicts its own statistical claims, a physically nonsensical SPR flow-rate unit, and unusually tidy control values in cytotoxicity bar charts. Findings are flagged as needing raw-data verification by the journal or an institutional investigation.

Key findings

Statistical self-contradiction (severity: critical) — Figure 6k legend lists P=0.4183 twice for two distinct comparisons in SK-BR-3 ADCC; the text claims CDRH3_68 vs Herceptin is P<0.05 in SK-BR-3, but the legend value (0.4183) is non-significant.
Plausible AI-hallucinated methodology (severity: high) — SPR section states a flow rate of "30 μmol l−1"; μmol l−1 is a concentration unit, not a flow-rate unit (expected: μl/min). Typical Biacore 8K+ flow rates are 10–30 μl/min.
Suspiciously regular control values (severity: moderate) — In Fig. 6j (AU565 ADCC), No Ab = 15.77% and CDRH3_Raw = 16.97% (Δ=1.2); Fig. 6l (AU565 CDC), No Ab = 9.4 and Raw = 9.1 (Δ=0.3); Fig. 6m (SK-BR-3 CDC), No Ab = 6.2 and Raw = 6.1 (Δ=0.1). The decimal-end uniformity is atypical for biological replicates.
Timeline plausibility concerns (severity: moderate) — Authors are affiliated with computer science and AI centres. Nine full-length antibodies were reportedly synthesised and validated via SPR, ELISA, ADCC and CDC within a revision window of <1 year; manuscript revised and accepted between April 2025 and early 2026, while citing a June 2025 preprint (Ref. 49, Passaro et al., Boltz-2).

Evidence highlights

DOI: 10.1038/s43588-026-00976-0
Figure 6j,k legend — duplicated P=0.4183 for "versus CDRH3_Raw" and "CDRH3_68 versus Herceptin" in SK-BR-3 ADCC.
Figure 6k legend — CDRH3_68 vs Herceptin, P=0.4183 (non-significant); Fig. 6j — same comparison in AU565, P=0.0186 (significant).
Methods, SPR analysis — "constant flow rate of 30 μmol l−1" on a Biacore 8K+ (RRID/cat. 29337763, Cytiva).
Methods, ELISA — antibody lot numbers listed, e.g. CDRH3_39 lot KH18SE09503; Mason et al. antibody KH19JL04174.
Reference 49 — Passaro et al., Boltz-2, bioRxiv posted 2025-06-14.
Manuscript timeline — Received 7 April 2025; Accepted 4 March 2026.

Notes

This report is AI-assisted and intended for academic discussion only. The duplicated P-value and the unit error are objective textual facts; their ultimate cause (data fabrication, AI-generated text, or copy-editing artefacts) cannot be determined without the raw SPR sensorgrams, flow cytometry FCS files, and source statistics. Other limitations: bar-chart values were read from the published figure and may differ from underlying data; timeline feasibility is qualitative. No image-manipulation check was performed.