Caracterização de técnicas de aprendizado de máquina para classificação de reatividade entre compostos e proteínas

Abstract

This work aimed to investigate the binary prediction of reactivity between chemical compounds and proteins, using the Drug-Target Interaction (DTI) approach through different machine learning models. The methodology involved segmenting cataloged reactivity indices into groups of reactive and non-reactive compounds, constructing databases for each target compound, and extracting protein feature vectors using various patterns, such as APAAC, CTDD, CTRiad, DDE, Geary, K-SEP PSSM, PFAM, QSO, and SPMAP. The data were collected from the ChEMBL database, and the protein representation was obtained from the UniProt database. The binary classification was performed by separating the reactivity values into reactive and non-reactive, converting them to 1 and 0, respectively. For the analysis, classification models were applied, including Random Forest, KNN, SVM, MLP, XGBoost, and Dummy Classifier, with hyperparameter optimization using the Grid Search algorithm. The results showed that, although no model consistently outperformed the others, some model-feature pairs presented promising results. XGBoost with the CTriad vector achieved an average accuracy of 72.78%, while DDE with Random Forest reached 71.08%. The MLP model, on the other hand, exhibited below-average performance, possibly due to the scarcity of data in the dataset. A detailed analysis of the results revealed that XGBoost and Random Forest excelled in terms of performance, especially in datasets with a smaller number of samples, such as the ChEMBL104 compound. KNN and K-SEP PSSM also showed above-average results, while SVM, using the Geary vector, demonstrated one of the best performances across many compounds, although it faced limitations in some cases. These findings suggest that the choice of model and feature vector is crucial for the effectiveness of predicting reactivity between compounds and proteins. The research highlights the importance of further investigation into the interactions between models and features, as well as the need for a larger volume of data to improve the generalization and consistency of more complex models.