TropMol: a cloud-based web tool for virtual screening and early-stage prediction of acetylcholinesterase inhibitors using machine learning

Abstract

Alzheimer's disease (AD) is the most common type of dementia, accounting for at least two-thirds of dementia cases in people aged 65 and older. Numerous approaches have been studied for the treatment of this disease, including the cholinergic hypothesis. Acetylcholinesterase (AChE) is the most promising target studied within the cholinergic hypothesis for the treatment of AD. Therefore, it is necessary to develop predictive models for the identification of AChE inhibitors. Thus, general drug design models can assist chemical synthesis groups and biochemical testing laboratories by enabling virtual screening and drug design. In this work, the objective is to build a generic molecular screening prediction model for public, online and free use based on pIC₅₀, using a random forest model (RF). For this, a dataset with approximately 16 000 compounds and 134 classes of descriptors was used, resulting in more than 2 000 000 calculated descriptors. Other algorithms were studied, such as gradient boosting, XGBoost, LightGBM, and RF with descriptors from principal component analysis (PCA), but none demonstrated significantly superior results compared to the RF model. The final model studied obtained an R² = 0.76 with a 15% test set and obtained an R² = 0.73 with a 30% test set, with rigorous Y-scrambling confirming the absence of chance correlation. External validation performed on an independent test set comprising 10% of the data yielded an R² of 0.77 and an RMSE of 0.67, statistically confirming that the model retains high predictive accuracy for novel chemical scaffolds and is free from overfitting. It is suggested that compounds containing oxime groups (RR'C = NOH) and those with high structural branching (higher Balaban index) tend to be less potent AChE inhibitors (negative correlation). In addition, some descriptors indicate that electronic charge distribution, molecular surface area, and hydrophobicity play important roles in correlating with the inhibitory activity (pIC₅₀) of the compounds. The presence of linear alkane chains also seems relevant to activity (positive correlation and greater importance). The data and models are available at the following link: (https://colab.research.google.com/drive/1gMcuXAsrqTIBMNnsCEWG9xfkK7aaZAbn?usp=sharing).