The role of different sampling methods in improving biological activity prediction using deep belief network

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning in Drug Discovery

Background:

• Advancements in chemistry yield vast molecular data, complicating quantitative structure-activity relationship (QSAR) studies.
• Accurate parameter initialization is crucial for statistical modeling, particularly for deep neural networks (DNNs).
• Random parameter selection in DNNs often results in suboptimal performance for biological activity prediction.

Purpose of the Study:

• To investigate the application of deep belief networks (DBN) for initializing DNNs in QSAR.
• To evaluate the impact of different sampling approaches within DBN on DNN performance.
• To enhance the prediction of biological activity for a large dataset of molecules.

Main Methods:

• Implemented deep belief networks (DBN), comprising stacked restricted Boltzmann machines, to initialize deep neural networks (DNNs).
• Developed and applied three distinct sampling approaches to address the log likelihood gradient computation in DBN.
• Utilized the DBN-initialized DNNs to predict biological activity for fifteen Kaggle targets comprising over 70,000 molecules.

Main Results:

• DBN-based initialization significantly improved the performance of DNN models for biological activity prediction.
• The models demonstrated superior predictive power compared to DNNs initialized with random parameters.
• Different DBN sampling strategies showed varying impacts on the final prediction accuracy.

Conclusions:

• Deep belief networks offer a robust method for parameter initialization in DNNs for QSAR.
• This approach enhances the accuracy of predicting molecular biological activity.
• The findings suggest DBNs are a valuable tool for improving machine learning models in medicinal chemistry.