Convolutional Neural Network Model Based on 2D Fingerprint for Bioactivity Prediction.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Predicting molecular behavior and properties is crucial for various scientific processes.
• Current computational methods like machine learning have limitations in accuracy and error rates for predicting physicochemical properties.

Purpose of the Study:

• To develop a novel deep learning technique for accurate prediction of chemical compound bioactivity.
• To introduce and evaluate a new molecular matrix representation called Mol2mat.

Main Methods:

• Developed a deep learning convolutional neural network (CNN) model.
• Utilized a novel Mol2mat molecular matrix representation derived from 2D-fingerprint descriptors.
• Experimented with standard datasets (MDDR, Sutherland) and analyzed combinations of eight fingerprints, focusing on the five best descriptors.

Main Results:

• The proposed CNN method combined with Mol2mat representation achieved a high performance of 98% AUC.
• A specific combination of three fingerprints (ECFP4, EPFP4, and ECFC4) yielded the best results.
• Outperformed state-of-the-art machine learning algorithms including NaiveB, LSVM, and RBFN.

Conclusions:

• The novel deep learning approach with Mol2mat representation offers a significant advancement in predicting chemical bioactivity.
• This method demonstrates superior accuracy and reduced error rates compared to traditional machine learning algorithms.
• Highlights the potential of deep learning and novel molecular representations in cheminformatics and drug discovery.