Novel Vector Embeddings by Molecular Attributes helps de-novo Molecular Generation

The increasing need for molecules with tailored properties in fields such as drug discovery and chemical engineering has driven the development of advanced computational methods for molecular design. Machine learning (ML) based approaches focused on de-novo molecular generation have recently gained significant attention. This paper presents a new transformer-based vector embedding generator that can be combined with Generative Adversarial Networks (GANs) to generate new molecules with desired properties. The embedding generator introduces a new molecular descriptor instead of SMILES strings, integrating Morgan fingerprints with global molecular attributes, enabling the transformer to capture local functional groups and broader molecular characteristics. The transformer model achieves a reconversion accuracy of 94% while translating the molecular descriptors back to SMILES strings, reinforcing the validity of the new descriptor for generative tasks. The model is validated for generating novel odorant molecules using a labeled dataset of odorant and non-odorant compounds. The GAN generates diverse molecules, maintaining high novelty, uniqueness, and validity. This work highlights the potential of using the proposed advanced molecular descriptors with transformers and generative AI techniques to accelerate the discovery of molecules with tailored properties, offering a powerful tool for a wide range of applications in molecular design.