Molecular generative model based on conditional variational autoencoder for de novo molecular design (1806.05805v1)

Abstract: We propose a molecular generative model based on the conditional variational autoencoder for de novo molecular design. It is specialized to control multiple molecular properties simultaneously by imposing them on a latent space. As a proof of concept, we demonstrate that it can be used to generate drug-like molecules with five target properties. We were also able to adjust a single property without changing the others and to manipulate it beyond the range of the dataset.

Molecular Generative Model Based on Conditional Variational Autoencoder for De Novo Molecular Design

The paper presents an innovative approach to molecular design utilizing a Conditional Variational Autoencoder (CVAE) framework for generating molecules with predefined properties. The authors highlight the limitations of purely experimental techniques in traversing the vast chemical space, estimated to encompass between $10^{23}$ to $10^{60}$ drug-like molecules. Within this context, computational methods, particularly those leveraging deep learning, offer promising solutions to streamline the discovery of novel compounds with target features.

Methodology

The authors propose a CVAE model that effectively generates drug-like molecules by controlling multiple molecular properties simultaneously in a latent space encoded with these attributes. Unlike traditional VAEs, the CVAE incorporates condition vectors into its objective function, ensuring that molecule generation adheres to predefined criteria. The adjustment of individual properties while maintaining the rest of the chemical profile is a notable development, reflecting the model's potential for precision-driven molecular design.

Results

Through experimentation, molecules were generated with up to five distinct properties: molecular weight (MW), partition coefficient (LogP), number of hydrogen bond donors (HBD), number of hydrogen bond acceptors (HBA), and topological polar surface area (TPSA). The model demonstrated the ability to precisely adjust one property without affecting others and extend property values beyond those present in the training set. This capability underscores the CVAE's flexibility and applicability in molecular design, providing tools to explore regions of chemical space that are typically inaccessible by empirical methods.

Numerous examples validate the feasibility of using CVAE for effective molecular engineering. For instance, generating molecules with specific properties of drugs like Aspirin and Tamiflu was accomplished successfully. A significant demonstration involved manipulating the LogP value while maintaining other molecular features constant, showcasing the model's refined control over complex molecular interdependencies.

Challenges and Future Direction

While the CVAE method provides significant advantages, it presents challenges such as a low success rate in generating specific molecules, attributed to the discrete nature of molecular representations like SMILES and inherent correlations between targeted properties. The paper explores various latent vector sampling strategies, finding sampling around known molecules yields better outcomes.

The authors suggest avenues for further enhancement, including adopting more sophisticated molecular representations that encompass 3D structural information, potentially elevating the efficacy and accuracy of molecule generation. The integration of approaches like graph-based encodings and reinforcement learning could further refine the model's capabilities and enhance the diversity and validity rate of generated molecules.

Implications

This paper contributes a significant computational tool for de novo design in drug discovery and materials science. The ability to control complex molecular properties simultaneously opens new possibilities in tailoring compounds for specific therapeutic, structural, or functional needs. Future research may pivot towards refining the model through advanced representations and optimization techniques, promising broader applicability in real-world molecular engineering challenges. The CVAE's framework establishes a robust foundation for generating diverse, functional, and novel molecules, aligning with ongoing advancements in artificial intelligence and computational chemistry.