Theoretical and Computational Chemistry

MASSA Algorithm: automated rational sampling of training and test subsets for QSAR modelling



The use of computer-aided drug design has become an essential part of drug development. In this context, QSAR models capable of predicting biological activities, toxicity, and pharmacokinetic properties were widely used to search bioactive molecules in chemical databases for lead compounds. The preparation of dataset used to build these models has a strong influence on the quality of the generated models, and sampling these data requires that the original dataset be divided into training (used for model training) and test sets (used to further evaluate the predictability of the model). This sampling can be done randomly or rationally, but the rational or purposeful division is superior to the random method. In this paper, we present MASSA Algorithm ("Molecular dAta Set SAmpling Algorithm"), an open-source, user-friendly Python tool that can be used to perform automatic sampling of datasets of molecules into training and test sets by exploring the biological, physicochemical, and structural spaces of molecules using Principal Component Analysis, Hierarchical Clustering Analysis, and K-modes clustering. This proposed algorithm is very useful when the variables for QSAR model generation is not available or to construct multiple QSAR models with the same training and test sets. When compared to random sampling, the presented algorithm produces models with lower variability and higher values across multiple replicates for validation metrics. These results were obtained even when the descriptors used in the QSAR/QSPR were different from those used in the separation of training and test sets, indicating that this tool can be used to build models for more than one QSAR/QSPR technique. Finally, the tool also generates useful graphical representations of the distribution that can provide insights into the data.


Thumbnail image of manuscript_12aug.pdf

Supplementary material

Thumbnail image of
Supplementary Material
Similarity maps for the seven datasets and their training-test distributions from MASSA, random, and referential (from the original study) approaches.