Integrity assessment report: "Interpretable identification of cancer genes across biological networks via transformer-powered graph representation learning" (TREE), Nature Biomedical Engineering, DOI:10.1038/s41551-024-01312-5

Academic fraud report · Geng Detector

Summary

Verdict: Questionable (🟡). This is a purely computational/AI methodology paper, so the image-manipulation heuristics used in traditional Geng reports (Western blot splicing, flow cytometry reuse) are not applicable; scrutiny shifts to logical consistency between figures, captions, and Methods. Four issues were flagged, none rising to data fabrication. (1) Figure 6f's Venn diagram lists "RF" as an independent comparator alongside LAP, DW, LINE, EMOGI, GAT and TREE, whereas the paper consistently describes comparison against 5 SOTA baselines and only mentions random-forest classifiers as a wrapper for LAP/DW/LINE in Methods. (2) P-value significance thresholds are inconsistently defined: Figure 4 caption uses **<0.05, ****<0.001, while Figure 6 caption uses **<0.05, ***<0.01, contradicting itself. (4) Figure 5a extracted text shows a gene list repeated 4×, plausibly a redundant text-layer artifact but unverified without original image. (5) Timeline (Received 27 Jul 2023; Accepted 1 Nov 2024; datasets and TF 2.6.2 all pre-2023) is internally consistent. No fabrication is alleged; recommend Erratum, not formal misconduct referral.

Verdict

Questionable (🟡) — Multiple inconsistencies between figure captions, Methods text, and figure content. Severity is moderate; pattern suggests sloppy copy-paste pipeline rather than deliberate fabrication. An Erratum is recommended over a misconduct referral.

Key findings

Ghost baseline "RF" in Figure 6f: The Venn diagram includes a circle labeled RF alongside LAP, DW, LINE, EMOGI, GAT, and TREE (7 sets total). The paper repeatedly frames comparison against 5 SOTA methods, and Methods only discusses RF as a classifier wrapper for LAP/DW/LINE — not as an independent baseline. Either the figure or the Methods text is wrong.
Inconsistent significance notation: Figure 4 caption defines P < 0.05, P < 0.001; Figure 6 caption defines P < 0.05, *P < 0.01. The meaning of changes between panels, indicating non-uniform legend generation.
Repeated gene-list strings in Figure 5a text extraction: A long gene-name sequence (TP53 NFE2L2 MAF HOXD12 … CRMP1 PPM1B IRF1) appears 4× in the extracted text. Could be an unflattened text-layer artifact or overlapping subplot labels; needs visual inspection of the original PDF.
Timeline sanity check: Received 27 July 2023; Accepted 1 November 2024; datasets (TCGA, STRING v11.0, TarBase v7.0) and TensorFlow v2.6.2 are all consistent with the submission window. No anachronisms detected.

Evidence highlights

Fig 6f text enumerates seven circles: LAP, DW, LINE, EMOGI, GAT, TREE, RF.
Fig 4 caption (as quoted): P < 0.05, P < 0.001.

Fig 6 caption (as quoted): P < 0.05, *P < 0.01.

Fig 5a extracted text contains the gene-list block repeated 4 consecutive times.

DOI: 10.1038/s41551-024-01312-5; published online 9 January 2025; code repo referenced as GitHub: Blair1213/TREE.
Notes

Image-level forensics (Western blot reuse, gel splicing, microscopy cloning) are not applicable to this paper because it contains no wet-lab figures.

The "RF" discrepancy could be an honest mislabeling of the RF-augmented LAP/DW/LINE pipeline as a separate entity, but as written it misleads readers and complicates reproducibility.

The P-value inconsistency, if propagated into supplementary statistics, could affect interpretation of significance in panels beyond Fig 4 and Fig 6 — recommend full audit of every caption containing asterisks.

The Fig 5a repetition finding is tentative**; depends on PDF text-extraction behavior and should be confirmed by inspecting the rendered figure for overlapping text boxes.
No evidence of fabricated datasets or backdated reagents. Confidence in the overall verdict is moderate; the issues are real but individually minor and cumulatively indicative of insufficient editorial polish rather than fraud.