Enhancing <i>De Novo</i> Drug Design across\nMultiple Therapeutic Targets with CVAE Generative Models

Drug discovery is a costly and time-consuming process,\nnecessitating\ninnovative strategies to enhance efficiency across different stages,\nfrom initial hit identification to final market approval. Recent advancement\nin deep learning (DL), particularly in <i>de novo</i> drug\ndesign, show promise. Generative models, a subclass of DL algorithms,\nhave significantly accelerated the <i>de novo</i> drug design\nprocess by exploring vast areas of chemical space. Here, we introduce\na Conditional Variational Autoencoder (CVAE) generative model tailored\nfor <i>de novo</i> molecular design tasks, utilizing both\nSMILES and SELFIES as molecular representations. Our computational\nframework successfully generates molecules with specific property\nprofiles validated though metrics such as uniqueness, validity, novelty,\nquantitative estimate of drug-likeness (QED), and synthetic accessibility\n(SA). We evaluated our model’s efficacy in generating novel\nmolecules capable of binding to three therapeutic molecular targets:\nCDK2, PPARγ, and DPP-IV. Comparing with state-of-the-art frameworks\ndemonstrated our model’s ability to achieve higher structural\ndiversity while maintaining the molecular properties ranges observed\nin the training set molecules. This proposed model stands as a valuable\nresource for advancing <i>de novo</i> molecular design capabilities.