An Architecture of Deep Learning in QSPR Modeling for the Prediction of Critical Properties Using Molecular Signatures

Generally, product design and chemical processes largely depend on the physical properties of the materials. Therefore, understanding the relationship between the physical properties and molecular structures of materials is highly desirable for effective optimization and design of high-performance materials. To this end, several mathematical prediction models which are mostly based on semi-empirical quantitative structure-property relationships (QSPR) have been developed. However, these methods cannot be used to automatically characterize molecular structures with physical properties.

For instance, group contribution methods have been majorly used to predict the physical properties with molecular substructures while the topological indices methods are majorly used to characterize the molecular structures in graph theory. On the other hand, the application of atom signatures together with multiple linear regressions are not preferred owing to the complex encoding and mapping involved.

Interestingly, the present advancement in technology especially in computing and machine learning have attracted interest of researchers. Presently, deep learning of the molecular structures and correlating them with the physical properties have been identified as a promising solution. In particular, artificial neural networks have paved the way for the characterization of the various molecular descriptors with regard to the physical properties. However, numerous challenges such as limited computing abilities, frequent changes in the product design and properties have compromised the quality and feasibility of the modeling approach.

Recently, research has shown the feasibility of deep learning in artificial intelligence for automatically capturing the relevant data from a variety of bigger data. This will enable effective formulation of the models through deep neural networks. However, this will be realized by developing more efficient information carriers like images and graphs for representing the molecular structures during modeling.

To this note, the research group of Dr. Weifeng Shen from Chongqing University in collaboration with Dr. Jingzheng Ren from The Hong Kong Polytechnic University and Dr. Mario Eden from Auburn University develop a deep learning approach to automatically predict the physical properties of the molecular structures. They intended to automatically capture the valuable molecular features of the semi-empirical quantitative structure-property relationships modeling for a wider range of substances. Their research work is currently published in the journal, AIChE Journal.

Briefly, the developed architecture entailed: deep neural network based on the tree-structured long short-term memory network and back-propagation neural network which was also used to depict the molecular structures and correlated properties respectively. Next, data preparations were completed by molecular encoding strategies in conjunction with the embedded algorithms for the generation of vector representations. Lastly, in terms of accuracy, the learned deep neural networks models were compared to the classical group contribution methods.

The authors observed that the developed deep neural network approach was capable of automatically capturing relevant molecular features for the semi-empirical quantitative structure-property relationships modeling from the provided text-type descriptors. For instance, this did not require any form of frequency counts or calculation of the molecular descriptors.

In summary, the research team successfully developed a deep learning architecture for prediction of critical properties based on the molecular signatures. To actualize their study, they build, trained and tested semi-empirical quantitative structure-property relationships models based on the deep neural network approach using critical properties of several compounds. In general, it has been proved the learned deep neural network model is more accurate with the ability to cover a wide range of molecular structures. Altogether, the study will advance the use of deep learning and artificial intelligence for proper understanding of the relationship between the molecular structures and physical properties.