Prediction of molecular properties represented by SMILES via semi-supervised learning

• Grant number: 23/06444-0
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: August 01, 2023
• End date: December 31, 2023
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Marcos Gonçalves Quiles
• Grantee: Piero Andreeto Liduvino Ribeiro
• Host Institution: Instituto de Ciência e Tecnologia (ICT). Universidade Federal de São Paulo (UNIFESP). Campus São José dos Campos. São José dos Campos , SP, Brazil

Abstract

Understanding the properties of a material is essential to evaluate its application. The study of properties can be carried out in various ways, from theoretical investigation to experimental evaluation of the material under analysis. However, the time and cost required to evaluate a particular material can be high, restricting the discovery of new materials to a small portion of the chemical universe. In recent years, machine learning models have been widely used to assist in this task. Specifically, based on previously known data, machine learning models are trained and subsequently used to evaluate new materials. In this scenario, we proposed in (1) a machine learning model based on Multilayer Perceptron (MLP) networks to predict various properties of molecules contained in the public repository QM9. The results obtained demonstrated that the approach is promising and can provide accurate prediction results, even in the absence of geometric information about the molecules. However, the study considered only one dataset and used only supervised techniques, which limits the application of the model to scenarios where the training set is 100% labeled. In order to evaluate the generality of the model proposed in (1) and expand its use to scenarios where the training set may contain unlabeled examples, we will incorporate semi-supervised routines into the model and evaluate its prediction ability with other datasets available in the literature.