Search

In this paper, a machine learning method is proposed to extract molecular features as floating-point numbers in a high dimensional space from the language-like description Simplified Molecular Input Line Entry Specification (SMILES). Principle component analysis showed that this method can successfully classify alkanes and alcohols and also the chain lengths of the molecular by their location in a three-dimensional feature space. A neural network model is build using the location of a compound in this high dimensional space as input to predict the “sigma-profile”, the charge distribution of the molecule near a perfect infinite conductor, which is calculated by quantum mechanics. The sigma-profile can be used in the COSMOSAC model for predicting thermodynamic properties such as activity coefficient. Preliminary results showed that an accurate neural work model with generalization ability can be developed.

Computer-aided molecular design requires the ability to predict different molecular properties of interesting from using molecular structure. Traditional quantitative structural property relations were developed by extracting molecular features for predicting various properties. Hence domains of molecular features are different for predictions of different properties. In this work, the concept of a universal translator was used to develop a universal digital chemical space by translating and projecting the chemical representation SMILES to a high-dimensional space that can be collapsed into different molecular fingerprints. We demonstrated different kinds of pure component properties, such as electrical and thermodynamic properties can be predicted by a simple input of molecular structure, SMILES. This method eliminates the need to manually extract different molecular features for predicting different properties. The ability of model to predict sigma profiles also pave the way of prediction phase equilibria of mixtures using molecular structure only.