Transfer inhibitory potency prediction to binary classification: A model only needs a small training set.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Machine learning in drug discovery

Background:

• Selecting promising compounds from large libraries for experimental screening is a significant bottleneck in drug discovery.
• Accurate prediction of compound activity, such as inhibition constant (Ki) and half maximal inhibitory concentration (IC50), is crucial.

Purpose of the Study:

• To develop a machine learning model for predicting compound efficacy (Ki, IC50) using a small training dataset.
• To transform the prediction task into a binary classification problem focused on compound ranking.
• To establish an efficient in silico approach for prioritizing chemicals in drug discovery pipelines.

Main Methods:

• A data augmentation strategy was designed to maximize information from the training set.
• A novel reward function for deep reinforcement learning was formulated to balance feature selection and prediction accuracy.
• A particle swarm optimized support vector machine was employed for binary classification.
• A soft voting mechanism was utilized to reconcile classification outputs.

Main Results:

• The model demonstrated high and reliable accuracy in predicting compound properties.
• The model effectively ranks compounds based on selected molecular descriptors.
• The developed approach proved capable of prioritizing chemicals for experimental studies.

Conclusions:

• The proposed machine learning model offers a potential ligand-based in silico strategy for efficient chemical prioritization.
• This method reduces the laboriousness of compound selection in early-stage drug discovery.
• The model's ability to perform accurate ranking with limited data makes it a valuable tool for experimental studies.