Deep learning studies on molecular properties prediction

Abstract

The use of deep learning in fundamental sciences has increased significantly. This suggests that it may replace some traditional methods in computational chemistry, biology, and related fields. In Quantum Mechanical (QM) methods, calculating the total energy of molecules requires considerable computational effort. One must solve the Hartree-Fock (or Kohn-Sham in the case of Density Functional Theory (DFT)) equations; the time required depends on the number of nuclei, electrons, and a number of basis functions, scaling exponentially. However, Machine Learning (ML) methods might be used instead of QM methods, which have less computational time and reasonable accuracy. Machine learning has been proven to be computationally cheaper than methods such as DFT. Recently, many studies utilized deep learning techniques for prediction in QM functional theory energies, drug discovery, material properties, etc. To perform predictions of molecular properties, we should start with some raw data such as Cartesian coordinates of atoms and the charges. However, machine learning models cannot be directly fed with these features. The predictor property should be invariant toward rotation, translation, and permutation. This thesis investigates Deep Neural Networks (DNN) to predict Hartree-Fock energies utilizing an interactive computational chemistry repository platform called Retrievium, which consists of chemical structures of different sizes computed by Gaussian software. Different models have been investigated in this study, such as a matrix-based descriptor model, which utilizes both Convolutional Neural Network (CNN) and Multilayer Perceptrons (MLP) networks; ReMLP-NET, which employs MLP network, and a Transformer-Potential utilizing a Transformer Encoder architecture. These methods are all compared to each other, and we have discussed their pros and cons. They also have been compared with one of the well-known models, ANI, and the results show that our models perform appropriately in the employed datasets.