Which stiffness-associated expression signals survive replicate sensitivity in a synthetic TME dataset?

Reanalyze the public processed human expression matrix from GSE107063 to identify reproducible signals associated with soft gel, intermediate gel and glass culture. Deliver a benchmark that helps a synthetic tumour-microenvironment project decide which transcript-level changes warrant validation when selecting culture stiffness. A complete result showing instability or insufficient evidence is eligible; novel biomarkers are not required.

The listed project aims to improve cancer-therapeutic discovery, translational relevance and response prediction with synthetic tumour microenvironments. A relevant practical uncertainty is whether observed expression shifts are robust across replicates or dominated by one sample. The fixed dataset contains human MDA-MB-453 cells cultured on 5 kPa gel, 30 kPa gel or glass, four deposited replicates per condition. Use only these twelve human samples, not the accompanying mouse experiment or raw CEL files.

This is a robustness benchmark on real processed measurements. It cannot establish clinical drug response, identify causal resistance mechanisms, validate a new tumour model, or measure direct drug sensitivity. The paper's drug-response experiments provide context, not labels for training a predictive model.

• The exact compressed human processed input, a provenance.json recording source URL, SHA-256 and retrieved date, and samples.csv mapping all twelve matrix columns to GEO sample accession and culture condition. Record source descriptions of replication and preprocessing; distinguish known facts from unresolved batch or pairing information.
• analysis.py or analysis.R plus README.md: an executable pipeline using the deposited numeric values without silently re-normalizing or log-transforming them. Inspect and explain their scale from metadata and distribution; if log-base information is not adequately documented, retain “deposited processed expression units” and avoid unsupported fold-change interpretation. Remove only wholly empty formatting columns/lines automatically. Report missing values, duplicated IDs and any unusable rows with explicit rules and counts. No outcome-driven removal of samples.
• qc.svg or qc.png: sample distributions, sample correlations and a PCA display using features complete across all samples. Center each feature for PCA, without variance scaling, and show condition and sample IDs. Identify possible anomalous samples without asserting batch correction is justified by absent metadata.
• contrasts.csv.gz: one row per input feature and each of soft-minus-glass, intermediate-minus-glass and soft-minus-intermediate. Report group means, mean difference, two-sided Welch t-test p value and Benjamini-Hochberg q value calculated separately over finite tests within each contrast. Use sample variances with denominator n−1. Where a test is undefined, retain its effect if computable and mark the p/q as missing with a reason; do not fabricate significance. This prescribed Welch analysis is a transparent robustness baseline, not a claim to recreate the paper's limma analysis.
• stability.csv.gz: for every feature/contrast, remove each of the eight relevant samples in turn and recompute its mean difference and Welch test, giving eight leave-one-sample-out estimates. Report minimum/maximum difference, sign preservation fraction relative to the full-data difference, and the largest absolute change from the full-data difference. Zero and undefined cases need explicit handling. Also report how the full-data leading 100 features by ascending p value change rank in each omission fit; break p-value ties by feature ID. Do not treat these dependent refits as independent replication or probabilities.
• decision.md: compare the three contrasts, identify stable and fragile signals with exact feature IDs, and state whether the intermediate condition appears to follow or differ from the soft-versus-glass pattern. Specify any rule used to prioritize features before presenting the shortlist, justify it and include sensitivity to the rule. Translate the findings into a concrete proposed validation comparison for selecting a synthetic culture condition, with what additional measurement would change that decision. Gene-symbol annotation is optional, but if used must cite the platform mapping and preserve ambiguous/multiple mappings. No forced positive shortlist is required. Include sources and a clear distinction between expression association and drug efficacy.

All artifact bytes must be included. Compressed numeric tables are acceptable and must be readable by the submitted code. The README must give environment versions and one command regenerating numeric results and plots from the included input.

Fixed public inputs, as accessible on 17 September 2026:

1. GEO GSE107063, *Influence of substrate stiffness on chemotherapeutic response in breast cancer*: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE107063 . Human processed file GSE107063_All_Expression_Human.txt.gz: https://www.ncbi.nlm.nih.gov/geo/download/?acc=GSE107063&file=GSE107063_All_Expression_Human.txt.gz&format=file . Verified SHA-256: 7fa520db62472c4cca42eb79a9da50f347d76bf785bbcb35414364801336785a (3,472,007 compressed bytes). Use this byte version for the numerical benchmark. Samples GSM2860476 through GSM2860487; sample metadata exemplified by https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSM2860476 . The series page links all samples.
2. Medina et al. (2019), *Identification of a Mechanogenetic Link between Substrate Stiffness and Chemotherapeutic Response in Breast Cancer*, DOI 10.1016/j.biomaterials.2019.02.018, PMC6474249: https://pmc.ncbi.nlm.nih.gov/articles/PMC6474249/ . Methods 2.7–2.8 and the human-expression results supply experimental context.
3. Optional gene annotation only: platform GPL15207, https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GPL15207 . No annotation is required to complete the feature-level benchmark.

Project context: https://openlabs-git-codex-openlabs-elgora-adapter-bio-xyz.vercel.app/projects/da8d4ec9-aa84-45b5-9d6b-aa3f8afaecf9 . Independent contribution; no project-owner endorsement is claimed. No private data, laboratory access, raw-array reprocessing or GPU is required.

Only submissions satisfying all criteria are eligible. Prefer fewer material analysis errors, then the most defensible connection between robustness results and the proposed culture-validation decision, then stronger reproducibility and handling of data ambiguities. A larger hit list earns no preference. Remaining ties go to earlier on-chain submission timestamp, then lower numeric submission ID. No winner is required if none qualifies.