Generative machine learning for de novo drug discovery: A systematic review

Area of Science:

• Computational Chemistry
• Artificial Intelligence
• Drug Discovery

Background:

• Machine learning (ML) and generative models are transforming de novo drug discovery.
• AI algorithms can efficiently auto-generate novel drug-like molecules.
• Various model frameworks and input formats enhance generative molecular design.

Purpose of the Study:

• To systematically review ML models for de novo drug design over the last five years.
• To identify challenges and solutions in computational molecule design.
• To analyze molecular encoding methods used in generative AI for drug discovery.

Main Methods:

• Systematic literature review of experimental articles and reviews (PubMed, ScienceDirect, etc.).
• Identified 87 studies via database search and 12 via citation searching.
• Analyzed prominent ML algorithms, challenges, and molecular encoding techniques.

Main Results:

• Six prominent ML algorithms identified: LSTM-RNNs, VAEs, GANs, AAEs, evolutionary algorithms, GRU-RNNs.
• Eight central challenges highlighted: homogeneity, synthesizability, data limitations, interpretability, multi-property optimization, incomparability, size, and evaluation.
• Molecules encoded as strings, 2D graphs, or 3D graphs.

Conclusions:

• ML has significantly advanced de novo drug design efficiency.
• Addressing challenges in synthesizability, interpretability, and optimization is crucial for future progress.
• Continued evolution of ML approaches and encoding methods will shape future drug discovery.