Machine Learning-based Analysis of Electronic Properties as Predictors of Anticholinesterase Activity in Chalcone Derivatives

In this study, we investigated the correlation between the electronic properties of anticholinesterase compounds and their biological activity. While the methodology of such correlation is well-established and has been effectively utilized in previous studies, we employed a more sophisticated approach: machine learning. Initially, we focused on a set of 22 molecules sharing a common chalcone skeleton and categorized them into two groups based on their IC50 indices: high activity and low activity. Utilizing the open-source software Orca, we conducted calculations to determine the geometries and electronic structures of these molecules. Over a hundred parameters were collected from these calculations, serving as the foundation for the features used in machine learning. These parameters included the Mulliken and Lowdin electronic populations of each atom within the skeleton, molecular orbital energies, and Mayer's free valences. Through our analysis, we developed numerous models and identified several successful candidates for effectively distinguishing between the two groups. Notably, the most informative descriptor for this separation relied solely on electronic populations and orbital energies. By understanding which computationally calculated properties are most relevant to specific biological activities, we can significantly enhance the efficiency of drug development processes, saving both time and resources.