ChromaGRT: Curated RepoRT Dataset and Validation Splits for Chromatographic-Condition-Aware Retention Time Prediction

This dataset accompanies the study “ChromaGRT: a chromatographic-condition-aware graph transformer for retention time prediction in unseen reversed-phase LC methods”. It contains the curated data used to train and evaluate ChromaGRT, a graph-transformer model designed to predict absolute liquid-chromatography retention times from molecular structure and chromatographic-method information. The dataset was derived from RepoRT and curated to provide a method-centred benchmark for retention-time prediction in reversed-phase liquid chromatography. After preprocessing and quality-control procedures, the final modelling dataset contains 37,627 retention-time records corresponding to 13,357 unique standardized molecular structures measured under 151 distinct chromatographic conditions. The chromatographic metadata include information describing mobile- and stationary-phase composition, physical and operating parameters, stationary-phase selectivity, and gradient programs. Molecular information includes standardized molecular identifiers together with auxiliary descriptors such as monoisotopic mass and logP. The resource also provides the predefined validation partitions used in the study, enabling reproducible evaluation under three complementary generalization scenarios: random molecule–condition splits, Bemis–Murcko scaffold splits, and chromatographic-condition splits in which complete LC methods are excluded from training. These partitions support benchmarking of interpolation within known chromatographic conditions, structural generalization to unseen molecular scaffolds, and method-level generalization to previously unseen chromatographic conditions. The dataset is intended to support the development and reproducible evaluation of machine-learning methods for chromatographic retention-time prediction, particularly models that explicitly incorporate experimental LC conditions.