Machine Learning for De Novo Molecular Generation: A Comprehensive Review

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence and machine learning
• Drug discovery and medicinal chemistry

Background:

• Deep generative models offer powerful computational tools for *de novo* molecular design.
• They enable efficient exploration of chemical spaces beyond traditional experimental methods.
• Machine learning is revolutionizing drug discovery by generating novel molecular structures.

Purpose of the Study:

• To provide a comprehensive survey of machine learning-driven molecular generation.
• To systematically organize the field based on molecular representations, model architectures, and evaluation frameworks.
• To analyze state-of-the-art generative models, their mechanisms, strengths, limitations, and failure modes.

Main Methods:

• Review and systematic organization of generative models: Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), Recurrent Neural Networks (RNNs), Transformers, Diffusion Models, Normalizing Flows, and Hybrid Architectures.
• Analysis of algorithmic failure modes and practical deployment challenges.
• Discussion of applications in distribution learning, goal-directed generation, and specific therapeutic domains like CNS drug discovery.

Main Results:

• Detailed taxonomy and analysis of various deep generative model architectures for molecular design.
• Critical evaluation of the gap between computational benchmarks and practical medicinal chemistry, including synthetic feasibility and experimental validation.
• Identification of persistent challenges and promising future opportunities, such as physics-informed models and large language models.

Conclusions:

• Deep generative models are crucial for *de novo* molecular design, but face challenges in practical deployment.
• Addressing limitations in synthetic feasibility, experimental validation, and multiparameter optimization is key for next-generation drug discovery.
• Future directions include integrating physics, leveraging large language models, and utilizing autonomous laboratories for enhanced drug design.