Predicting Potent Compounds Using a Conditional Variational Autoencoder Based upon a New Structure-Potency...

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Quantitative structure-activity relationship (QSAR) models are crucial for predicting bioactive compound potency.
• Traditional QSAR models often rely on linear or nonlinear approaches, with machine learning powering nonlinear methods.
• A need exists for unified representations that capture both structural features and potency information.

Purpose of the Study:

• To introduce a novel molecular fingerprint, the structure-potency fingerprint (SPFP), for enhanced potency prediction.
• To develop and apply a conditional variational autoencoder (CVAE) model utilizing SPFPs.
• To evaluate the performance of the SPFP-CVAE approach against established methods like support vector regression (SVR).

Main Methods:

• Designed a novel structure-potency fingerprint (SPFP) encoding both structural features and potency values.
• Developed a conditional variational autoencoder (CVAE) model trained on SPFPs.
• Inputted only the structural module of test compounds to predict their potency module via the CVAE.

Main Results:

• The SPFP-CVAE approach achieved accuracy comparable to state-of-the-art SVR for predicting compound potency across different activity classes.
• Highly potent compounds were predicted with accuracy similar to both SVR and deep neural networks.
• The unified SPFP representation effectively enabled accurate potency predictions.

Conclusions:

• The novel SPFP-CVAE method offers a powerful and accurate approach for predicting bioactive compound potency.
• This unified representation and CVAE model advance the field of QSAR and machine learning in drug discovery.
• The SPFP-CVAE demonstrates competitive performance, particularly for predicting highly potent compounds.