BioActNet: A Machine Learning Approach to Biological Activity Prediction Using Molecular Fingerprints

Abstract

Phytochemical-based drug development offers several advantages, including the discovery of novel therapeutic agents derived from natural sources, often with fewer side effects compared to synthetic drugs. These compounds, found in a variety of plants, have evolved to serve protective functions which makes them great candidates for pharmacological applications. However, the traditional knowledge-based approach to phytochemical drug development has significant limitations. It relies heavily on well-documented plants, which restricts the scope of exploration to already known phytochemicals, thereby potentially overlooking a vast array of unexplored natural compounds with therapeutic potential. To overcome this limitation, we propose a novel approach that leverages advanced computational techniques to predict biological activity from input phytochemicals. The proposed system consists of two modules: the phytochemical preprocessing module and the bioactivity prediction module. The preprocessing module takes the molecular structures of phytochemicals and converts them into molecular fingerprint representations to be fed into the subsequent machine learning-based biological activity prediction network. This prediction network then takes these molecular fingerprints as input and outputs the probability of various biological activities. To enhance the accuracy of the system, a vector shift technique is introduced, which can be easily applied to the prediction module without altering its network architecture. Comprehensive experiments demonstrated that the proposed system achieved state-of-the-art performance, with an accuracy of 90.80% on a public phytochemical dataset.