Mechanism-informed prediction of cancer treatment response: data, code and reviewer assessments for a systematic review of 484 studies

This deposit contains the data, code and reviewer records behind a systematic review of computational models that predict cancer treatment response and represent tumour mechanism within the model: "Mechanism-Informed Artificial Intelligence for Cancer Therapy Selection: A Systematic Review of Proliferative Signalling, Apoptosis Resistance, and the Preclinical–Clinical Translation Gap" (PROSPERO CRD420261507128). What changed in this version Version 2.0.0 reports the final reconciled analysis of 484 studies: 203 with treatment outcomes observed in patients and 281 preclinical or in-silico. It is the 487 studies of the earlier deposit less three records (442, 495 and 2472), excluded at the final primary-text reconciliation because they are feature-only models: tumour biology enters them as selected inputs, not as mechanism represented inside the model. Record 127 (ModCell) is placed in the preclinical arm and its disputed coupling is not counted in the primary analysis; record 851 carries an Expression of Concern and is retained. Seven sensitivity analyses, among them the 487-study set, the disputed coupling reinstated and the omission of record 851, are reported beside the primary analysis. The risk-of-bias audit is now recorded from the two reviewers' own entries. Two reviewers each assessed separate records: 175 records, 64 evaluation forms and 1,280 signalling answers, with the author of every entry recorded. The 600 answers that an AI comparison arm pre-filled are labelled as such, and drafts written by an AI assistant on 20 October 2026 are kept in a separate file labelled as AI drafts and are used in no count. The human-assessment tables of version 1.2.0 are superseded: they drew on sheets of the lead assessor's workbook that repeat those AI drafts. Every dataset, supplementary workbook and figure of the final analysis is rebuilt by numbered scripts from the data in the deposit, and a verifier recounts the numbers the manuscript reports. The Supplementary Methods, the legends of the supplementary tables and figures, and the PRISMA 2020 checklist for the submitted manuscript (with the PRISMA 2020 for Abstracts checklist) are added. The archive of the earlier deposit (search, screening, adjudication and extraction) is carried forward complete, as it stood on 17 October 2026. What is here current_analysis: the 484-study dataset and the 487-study sensitivity dataset, with a data dictionary; the exclusions, correction log, primary-text claim checks and sensitivity results; the reviewers' scope classifications, evaluation forms and signalling answers, with their workbooks as received; Supplementary Tables 2, 4 to 9, 11, 12 and 14 as workbooks with live formulas, and an index to Tables 1 to 18; the eleven figures (SVG and PNG) with the numbers each one plots; the Supplementary Methods and supplementary legends (PDF) and the PRISMA 2020 checklist; and the scripts and verifier. historical_archive: the five database search strategies (run on 26 and 27 August 2026), the pipeline scripts with the screening rule set, the screening decision on all 68,464 deduplicated records, the adjudication of the 2,483 prioritised records, the independent duplicate screening, backward citation searching (run between 14 and 19 September 2026), the 487-study extraction, the written specifications against which the later screening, extraction and risk-of-bias passes were run, and the remaining supplementary tables. The search strategies, the screening rule set and these specifications are the protocol documents that the manuscript points to; the README lists them. It describes the 487-study analysis as it then stood; the README says which of its files are superseded. What is deliberately not here The raw database export files, and the intermediate files built directly from them, are not deposited, because the databases' licences do not permit redistribution; everything derived from them is deposited as DOI, PubMed identifier, year and title. Publishers' full texts are not redistributed, and the manuscript is not part of this deposit. The headline result, so that it can be checked directly Of the 484 studies, proliferation is an explicit model variable in 269 and cell death in 186; the two are coupled within one model in 122. In 110 of those the death term is a generic kill rate, and in only 1 (record 395, a preclinical model) is it named apoptotic machinery; counting the disputed ModCell configuration gives 123 and 2. Neither establishes longitudinally updated coupling in treated patients. Of 27 retained clinical prediction records, 24 contain a high-risk evaluation, 2 have unclear evaluations only and 1 has no evaluation form; none is rated low. This is a finding within the retrieved and prioritised literature, not proof that no such model exists. The numbers are reproduced by running current_analysis/scripts/38_verify_package.py.