A Deep-Learning Proteomic-Scale Approach for Drug Design.

Area of Science:

• Computational chemistry
• Pharmacology
• Machine learning

Background:

• Novel therapeutic discovery is increasingly reliant on computational methods.
• The Computational Analysis of Novel Drug Opportunities (CANDO) platform uses a holistic approach to assess drug-protein interactions for efficacy and safety.
• Existing methods require optimization for generating novel drug candidates.

Purpose of the Study:

• To develop and validate a machine learning model for designing novel, effective, and safe drug-like compounds.
• To reduce the dimensionality of drug-proteome interaction signatures using deep learning.
• To generate novel compounds by recreating target interaction signatures.

Main Methods:

• Utilized a deep learning-based autoencoder for dimensionality reduction of CANDO-computed drug-proteome interaction signatures.
• Employed a reduced conditional variational autoencoder to generate novel drug-like compounds based on target "objective" signatures.
• Designed compounds to mimic interaction signatures of twenty approved and experimental drugs.

Main Results:

• 16/20 designed compounds showed significantly higher behavioral similarity to objective drugs compared to controls (p ≤ 0.05).
• 20/20 designed compounds demonstrated improved behavioral similarity against random controls.
• Redesigned compounds from rational drug design objectives outperformed those from natural sources (p ≤ 0.05), indicating learned design principles.
• Designed compounds exhibited structural diversity and synthetic feasibility.

Conclusions:

• This study demonstrates a significant advancement in automating holistic therapeutic design using machine learning.
• The approach enables the rapid generation of novel, effective, and safe drug leads for diverse indications.
• Successfully generated enhanced drug designs for non-small cell lung cancer and anti-aging applications.