Avaliação de modelos de redes neurais artificiais na classificação de similaridade estrutural molecular

Abstract

Molecular similarity is a central concept in cheminformatics, with relevant applications in the evaluation of new drugs, especially orphan drugs. Traditional methods for this task rely on fingerprints combined with logistic regression models, which, despite their good performance, have inherent limitations related to structural information loss and classifier linearity. This work investigates the performance of deep learning models applied directly to molecular similarity classification from 2D and 3D representations, through three convolutional architectures inspired by LeNet-5 and adapted for the simultaneous processing of pairs of 2D images and 3D representations. The results indicate that the proposed models did not outperform the logistic regression baseline reported in the literature, achieving ROCAUC values between 0.5627 and 0.6356, compared to values between 0.845 and 0.988 for the reference models. The reduced training data volume (200 molecular pairs) is identified as the main limiting factor. The study contributes to the field by transparently reporting the challenges of the approach and identifying directions for future research, such as dataset expansion, class balancing, and the exploration of transfer learning architectures.