A team of researchers from the Skolkovo Institute of Science and Technology, the University of Vienna, and Sirius University of Science and Technology has published a study in the Journal of Computer-Aided Molecular Design presenting guidelines for enhancing drug discovery processes by utilizing multitask learning techniques.
Effect of data enrichment on multitask learning. Credit: Ekaterina Sosnina/Skoltech
In universities, students often take related courses, such as physics and mathematics, which leads to a better understanding of both subjects. Similarly, learning a new language is easier for those who have already had experience with languages, particularly cognate ones. The same principle applies to machine learning, where a neural network can better comprehend multiple “subjects” if it learns them simultaneously. Considering that neural networks are one of the best approaches for predicting the biological properties of new chemical compounds, the question arises: How can we assist a neural network in simultaneously learning and predicting the properties of chemical compounds in relation to multiple biological targets?
Researchers analyzed three datasets for this purpose: a dataset with information on the antiviral activity of molecules and two datasets with information on the impact of molecules on various proteins in our body. The datasets varied in the completeness of information on each protein or virus. During the study, scientists discovered that adding data to the dataset is an effective method for improving prediction accuracy. Furthermore, they demonstrated that the more informative the original dataset is, the more noticeable this improvement. As a result of this work, the research team prepared a set of recommendations for using data enrichment technology to improve the quality and stability of predictions, as well as methods for objectively evaluating the improvement achieved.
“Multitask learning is widely used in many scientific fields. Unsurprisingly, it is increasingly being applied to develop new drugs. However, the potential of this approach has not yet been fully explored, presenting us with numerous unresolved tasks,” the lead author of the study, Skoltech PhD candidate Ekaterina Sosnina, notes. “We were inspired by the possibility of using multitask learning to develop new drug candidates and searched for ways to improve this approach. Following our recommendations, researchers in drug discovery will enhance the predictive accuracy of their models and accelerate the identification of novel drug candidates.”
Contact information:
Skoltech Communications
+7 (495) 280 14 81
