[HyTruongSon]

GraphPathology separates two questions: whether graph context improves a matched local visual predictor, and whether it adds value beyond the same observed molecular anchors. This journal revision preserves the original arXiv v1 and all earlier empirical findings while completing the focused ResNet18/Phikon comparison.

Matched GCN-assisted reconstruction improves correlation, MSE and MAE over local-assisted prediction in all 22 Bassiouni patients with both frozen backbones. Median paired relative MSE changes are −17.84% for ResNet18 and −1.70% for Phikon. These gains do not establish superiority to molecular interpolation: both GCN variants worsen the three Bassiouni cohort-median endpoints versus identical-anchor IDW. Image-only comparisons, GraphSAGE, mean/shuffled controls and individual Wu effects retain mixed and negative findings. Common-valid-gene sensitivity leaves the primary GCN correlation gains unchanged.

The manuscript leads with the controlled 22-patient placement result, replaces the opening with the expanded Introduction and separate Related Work, and uses a 178-word abstract with generic GNN wording. It removes only the requested conceptual information-boundary schematic, retaining its asset outside the active package and its scientific content in Methods. All earlier nonlocal and latent-coordinate negative results remain. The source includes a compact paired table/figure and a separate fitting-resource table.

Research provenance:

• Committed exploratory protocol before new fitting; source/validation-guided common two-rate refinement before any Phikon development/external outcome.
• 72 candidate fits: 18 reused, 54 new; 36 validation-selected heads, all planned seeds retained. Frozen Phikon features cover 74,020 locations in 29 patients on local CPU.
• 936 selected patient–model bundles: 468 reused and 468 new. Full local checks verify all 72 states and 936 gene arrays; new scoring performs 66,744 exact hidden-label corruption checks and 7,956 image/anchor-independence checks.
• The original three primary image-head binaries remain missing. New heads are distinct experiments.

Validation:

• 142 software tests pass; original publication regression passes 29 display targets and 71 numerical checks.
• 8,887 protected prior scientific files remain unchanged; historical archive audit checks 74,253 files with zero changed/missing files; all 10 protected table bodies are unchanged.
• PDFs: 94 combined / 42 article / 52 supplementary pages, 40 figures, 24 tables and 40 cited references. No final TeX diagnostics; continuous combined pagination and independent S numbering.
• All pages were rendered/reviewed, with native inspection of changed material and final SI flow. Standalone documents match combined content. Clean source extraction builds without BibTeX/data/checkpoints; standalone display regeneration reproduces the table files and PNG exactly.
• A fresh public checkout passes integrity checks and explicitly counts private states/arrays as absent rather than claiming neural replay.

Reports and provenance: experiments/journal_revision/backbone_graph_comparison/COMPARISON_REPORT.md, VALIDATION_REPORT.md, candidate/selected manifests, paired summaries and manuscript/journal_revision/transformer_egnn/CLAIM_EVIDENCE_AUDIT.md. Reproduction commands: docs/MATCHED_BACKBONE_COMPARISON.md.

The full editable LaTeX ZIP and full numerical/head-state bundles are prepared for author-local review, with exact checksums in their manifests; they are not public release assets or DOI deposits. Upstream encoders and raw tissue/count matrices are excluded. Sole-human authorship, author direction/oversight versus AI assistance, and the Editorial Board disclosure are preserved.

Keep this PR draft and unmerged for final author review. No journal/arXiv submission, public release or paid computing service was initiated.

Final minor editorial polish

Commit 897fde3 makes only the requested final copyedits: natural Phikon model prose, corrected Figure 6/S6 historical scope, an original-architecture qualifier in Figure 1, and one prospective predefined-panel acquisition sentence. Full experimental details, the byte-identical 178-word abstract, numerical findings, protocols, model weights, disclosures and arXiv v1 remain unchanged.

The only edited manuscript TeX sources are article_content.tex, supplementary_content.tex and figures/Figure 1.tex under manuscript/journal_revision/transformer_egnn/. Documentation and the three compiled PDFs/build report are refreshed. Exact scope/checks: docs/FINAL_MINOR_POLISH_REPORT.md.

Validation: 142 tests; 29 display targets/71 numerical checks; 8,887 protected scientific files and 74,253 historical archive files preserved; 72 states and 936 full gene arrays checked. The 94/42/52-page documents compile without TeX diagnostics; affected pages were rendered and visually inspected. The refreshed author-local source ZIP builds cleanly without BibTeX or research inputs. No new experiments or restructuring.

Keep this PR open, draft and unmerged for author review.

Author-approved disclosure and merge authorization

The author approved replacing the final Methods disclosure verbatim and authorized merging after validation. The replacement is in article_content.tex; the existing integrity check and attribution test now require its exact wording and end-of-Methods placement. Scientific content outside that paragraph and the abstract are unchanged. All 142 local tests and existing scientific-integrity checks pass. PDFs and a clean standalone source ZIP are refreshed; affected rendering was inspected. See docs/AI_DISCLOSURE_UPDATE.md. The earlier draft/unmerged instructions above are superseded by this author authorization. Merge after the latest GitHub checks pass.