2026-10-14T07:18:38Z https://www.scidb.cn/oai

10.57760/sciencedb.45991 2026-09-10T14:06:28Z

2026-09-10 Revised Analysis Code and Aggregate Results for Machine-Learning Prediction and Structure–Retention Interpretation of Flavonoid RP-HPLC RI-like Behavior doi:10.57760/sciencedb.45991 en This dataset record provides analysis code, record-level source identifiers, model configurations, provenance audits, and aggregate results for flavonoid RI-like retention prediction and structure–retention analysis within a defined RP-C18–water/acetonitrile–formic acid chromatographic chemistry. The internal study cohort is reconstructed directly from RepoRT at commit 0a8feafb77c09c278850cc6628b18b4ef1210419 using a declared manifest of 1,969 record identifiers. Removal of 19 redundant records according to the specified rules yields a primary analysis dataset containing 1,950 records, 871 source-preserving canonical structures, and nine method groups. Representative sources span 25 RepoRT systems, whereas the complete set of 2,005 source associations spans 31 systems. PredRet, MetaboLights, and contributor names are upstream provenance annotations recorded within RepoRT, rather than additional databases independently downloaded and merged for this study. The restriction concerns chromatographic chemistry; gradient programs, column dimensions, flow rates, temperatures, and instrument platforms may still differ.The package includes mappings among study-record identifiers, structure identifiers, method groups, and RepoRT sources; integrity checks for 200 source files; and executable reconstruction and analysis entry points. After obtaining the specified source files, users can generate the complete internal modeling dataset in a separate local directory and rerun training-fold RI-like normalization, molecule-level and record-level evaluations, shared-compound and cross-method RI-like consistency analyses, leave-one-method-group-out diagnostics, glycoside-like flag ablation, SHAP sensitivity analysis, and source and structural-identity sensitivity analyses. Fixed model parameters are used in this reanalysis, without nested cross-validation. The best internal model score is therefore not an unbiased performance estimate independent of candidate-model comparison. The public reconstruction workflow does not automatically rediscover the historical 3,327 candidate records or reproduce undocumented early manual screening.The package does not directly redistribute third-party record-level retention times, molecular structures, or original publications. External retention-order validation requires a separately obtained input containing 53 literature records for pure standards in the specified format. The package provides the input template, computation code, and aggregate results, but not these external record-level data. Access, attribution, and redistribution of the underlying data remain subject to the applicable source terms. RI-like is an empirical relative measure defined by within-method-group retention-time ranges, rather than a stable retention index calibrated with external reference standards. Model applicability is limited to retention prediction and structure–retention analysis within the chromatographic methods represented in the training data. External validation assesses retention-order agreement only and does not establish absolute retention-time prediction capability for entirely unseen chromatographic methods. Flavonoids; reversed-phase high-performance liquid chromatography; RI-like; machine learning; SHAP; structure–retention relationship; external retention-order validation Wang Min PUBLIC https://creativecommons.org/licenses/by/4.0/ dataset Science Data Bank