Curated Dataset of Leukemia Cell Lines with Experimental GI50 Values

The dataset was constructed by integrating compound bioactivity data from two primary sources: the dataset reported by Aljarf et al. and the ChEMBL database. To ensure relevance to leukemia-specific drug response, the data were filtered to retain only three human leukemia cell lines: K-562, CCRF-CEM, and HL-60(TB).

Following data integration, quality control procedures, including the removal of duplicate entries and inconsistent records, were applied to obtain a clean and reliable dataset. The resulting dataset comprises 69,270 compound–activity pairs, distributed across the selected cell lines as follows: 28,602 for K-562, 27,850 for CCRF-CEM, and 12,818 for HL-60(TB).

Each compound is represented using a SMILES string and is associated with a transformed bioactivity value (pGI50), obtained by applying a logarithmic transformation to the original GI50 measurements. In addition, a binary activity label is provided for each compound, where compounds are classified as active or inactive based on a predefined pGI50 threshold, enabling both regression and classification tasks.

To mitigate class imbalance and ensure suitability for machine learning tasks, each subset corresponding to a given cell line was balanced as per the activity labels.