Deep learning applications for the design of novel molecules : an active learning deep learning platform for de novo generation of UV absorbing molecules

The primary outcome of this project is an Active Learning Deep Learning AI Platform to accelerate the development of novel biobased UV-absorbing small molecules, applicable to other types of molecules as well. It leverages deep learning-based generative and prediction models to predict and optimize compound properties iteratively. It offers a data-driven, efficient approach to designing new molecules from biobased (or other) starting molecules with some UV activity. The focus will be on biobased UV-absorbing molecules due to a distinct need for novel biobased UV compounds. Additionally, the system demonstrates broad applicability across various industries.

The most challenging task in the machine learning solution process is to gather and prepare relevant datasets. In this regard, several domain specific properties were explored, and in light of the absence of any proprietary dataset, publicly available datasets were explored for the relevant properties. A dataset consisting of UV absorption data of several thousand small molecules was identified as a useful dataset to build a machine learning model to predict UV absorption properties for a given molecule. Several Deep learning-based ML algorithms were examined, and testing yielded promising prediction results, which could be further optimized and transferred into building generative models to generate novel molecules. In silico and lab experiments were conducted to verify the feasibility and practical applicability of the proposed methods. And further established a framework for using functional property prediction and generative processes in the domain-specific area.

We have proposed an RNN network for predicting UV absorption max with high accuracy and have verified its performance through wet-lab and in silico tests. The process by which we developed the optimized RNN-based property prediction architecture enables us to generalize the model to other property datasets. We further study conditional Variational AutoEncoders as effective generative models for generating optimized molecules from a given molecule, and their effectiveness is confirmed by screening the resulting molecules against the starting molecule.