NCATS-sol

Dataset: Download it here.

Dataset description: 2,453 compounds and binary labels indicating whether they have low solubility.

Dataset preprocessing

• Download the original dataset from here, which contains 2,532 records;
• Drop one compound (2 rows) with inconsistent outcomes;

PUBCHEM_CID	PUBCHEM_EXT_DATASOURCE_SMILES	PUBCHEM_ACTIVITY_OUTCOME	Phenotype	Analysis Comment
661788	CC(=O)NC1=CC=C(C=C1)OCC2=C(C=CC(=C2)C3NC4=CC=CC=C4C(=O)N3CC5=CC=CO5)OC	Active	Moderate/High	class = 0
661788	CC(=O)NC1=CC=C(C=C1)OCC2=C(C=CC(=C2)C3NC4=CC=CC=C4C(=O)N3CC5=CC=CO5)OC	Inactive	Low	class = 1

• Drop one duplicated row;

PUBCHEM_CID	PUBCHEM_EXT_DATASOURCE_SMILES	PUBCHEM_ACTIVITY_OUTCOME	Phenotype	Analysis Comment
135422895	CN=C1CN=C(C2=C(N1)C=CC(=C2)Cl)C3=CC=CN3	Active	Moderate/High	class = 0
135422895	CN=C1CN=C(C2=C(N1)C=CC(=C2)Cl)C3=CC=CN3	Active	Moderate/High	class = 0

• Drop 76 compounds (rows) with an inconclusive outcome;
• Generate a new column low_solubility based on the Analysis Comment column: Low phenotype is mapped to positive class, Moderate/High negative class;
• Use RDKit to transform the SMILES to their canonical forms;

Reference

1. H. Sun, P. Shah, K. Nguyen, K. R. Yu, E. Kerns, M. Kabir, Y. Wang, and X. Xu, Predictive models of aqueous solubility of organic compounds built on a large dataset of high integrity, Bioorganic & Medicinal Chemistry 27, 3110 (2019).