[Paper Review] Deep Denerative Models for Drug Design and Response

Karina Zadorozhny, Lada Nuzhna|arXiv (Cornell University)|Sep 14, 2021

Computational Drug Discovery Methods40 references4 citations

TL;DR

This review synthesizes deep generative models (DGMs) for drug design and response prediction, evaluating their ability to generate chemically valid, biologically relevant molecules. It highlights state-of-the-art architectures like VAEs, GANs, and diffusion models, emphasizes the need for better molecular representations, and identifies data quality, benchmarking, and reproducibility as key challenges.

ABSTRACT

Designing new chemical compounds with desired pharmaceutical properties is a challenging task and takes years of development and testing. Still, a majority of new drugs fail to prove efficient. Recent success of deep generative modeling holds promises of generation and optimization of new molecules. In this review paper, we provide an overview of the current generative models, and describe necessary biological and chemical terminology, including molecular representations needed to understand the field of drug design and drug response. We present commonly used chemical and biological databases, and tools for generative modeling. Finally, we summarize the current state of generative modeling for drug design and drug response prediction, highlighting the state-of-art approaches and limitations the field is currently facing.

Motivation & Objective

• To provide a comprehensive overview of deep generative models applied to drug design and drug response prediction.
• To clarify essential chemical and biological terminology, including molecular representations and key databases.
• To evaluate current state-of-the-art generative approaches for small molecules, gene therapy, and protein sequences.
• To identify critical limitations such as data quality, lack of standard benchmarks, and reproducibility issues in preclinical data.
• To advocate for improved molecular representations and rigorous evaluation standards to enhance real-world applicability in drug discovery.

Proposed method

• Surveying and comparing major deep generative architectures: Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), Autoencoders (AEs), and Adversarial Autoencoders (AAEs).
• Analyzing molecular representations such as SMILES strings, molecular fingerprints, and graph-based encodings, with emphasis on their limitations in capturing chirality and 3D structure.
• Evaluating generative models trained on high-throughput screening (HTS) data and biological databases to predict drug response and optimize for efficacy and safety.
• Assessing model performance using standardized benchmarks like GuacaMol and MOSES to ensure reproducibility and comparability.
• Applying generative models to generate novel compounds, such as DDR1 kinase inhibitors, and comparing them to known FDA-approved drugs.
• Highlighting the importance of incorporating structural and stereochemical information (e.g., chirality) into latent space representations to improve chemical validity and biological relevance.

Experimental results

Research questions

• RQ1 How do deep generative models improve the efficiency and accuracy of de novo drug design compared to traditional QSAR or docking methods?
• RQ2 What are the key limitations of current molecular representations (e.g., SMILES, fingerprints) in capturing biologically relevant properties like chirality and 3D conformation?
• RQ3 To what extent do generative models trained on real-world biological data produce novel, synthetically feasible, and biologically active compounds?
• RQ4 How can standardized benchmarks and evaluation protocols improve the reproducibility and reliability of generative models in drug discovery?
• RQ5 What role does data quality and reproducibility of training data play in the success of deep generative models for drug response prediction?

Key findings

• Generative models such as VAEs and GANs can produce novel molecular structures with desired pharmacological properties, including DDR1 kinase inhibitors that resemble FDA-approved drugs.
• Despite progress, 75–97% of novel drugs fail in clinical trials due to lack of efficacy or safety, underscoring the need for better prediction of in vivo response.
• Current models often fail to capture chirality, leading to generated molecules that may not match the biological activity of their enantiomers, a critical limitation in drug design.
• The field suffers from a lack of standardized benchmarks; proposed tools like GuacaMol and MOSES are emerging but not yet universally adopted.
• Some generated compounds are nearly identical to existing patented drugs (e.g., ponatinib), raising concerns about intellectual property and data leakage from training sets.
• Reproducibility issues in biological data—such as non-reproducible oncology drug targets—undermine model performance, even with state-of-the-art architectures.