Targeting neurodegeneration: three machine learning methods for G9a inhibitors discovery using PubChem and scikit-learn

In light of the increasing interest in G9a’s role in neuroscience, three machine learning (ML) models, that are time efficient and cost effective, were developed to support researchers in this area. The models are based on data provided by PubChem and performed by algorithms interpreted by the scikit-learn Python-based ML library. The first ML model aimed to predict the efficacy magnitude of active G9a inhibitors. The ML models were trained with 3112 and tested with 778 samples. The Gradient Boosting Regressor perform the best, achieving 17.81% means relative error, 21.48% mean absolute error, 27.39% root mean squared error and 0.02 coefficient of determination (R²) error. The goal of the second ML model, called a CID_SID ML model, utilised PubChem identifiers to predict the G9a inhibition of a small biomolecule that has been primarily designed for different purposes. The ML models were trained with 58,552 samples and tested with 14,000. The most suitable classifier for this case study was the Extreme Gradient Boosting Classifier, which obtained 79.7% accuracy, 83.2% precision,67.7% recall, 74.7% F1-score and 78.4% ROC. Up to date, this methodology has been used in seven studies, achieving a mean accuracy of 82.75%, precision of 90.71%, Recall of 73.01%, F1-score of 80.79% and ROC of 80.63% across all case studies. The third ML model utilised IUPAC names. It was based on the Random Forest Classifier algorithm, trained with 19,455 samples and tested with 14,100. The probability of this prediction was 68.2% accuracy. Its feature importance list was reordered by the relative proportion of active cases in which they participate. Thus, “iodide” was identified as the one with the highest relative proportion of the active cases to all cases where this fragment participated. In addition, ‘iodo’ was identified as the most desirable fragment, and “phenylcarbamate” as the least desirable based on their participation only in active or inactive cases, respectively. The computational approach has been initially developed and demonstrated using a case study on Tyrosyl-DNA phosphodiesterase 1(TDP 1) inhibition.