Pharmacophore-constrained de novo drug design with diffusion bridge

Computer-aided drug design (CADD) plays a crucial role in the modern drug discovery procedure. However, conventional CADD approaches such as virtual screening are undertaken to search for the candidates with optimal molecular properties in a vast chemistry library. Although accelerated by the high-throughput technology, this process can still be time-consuming and costly [1] since the relationship between chemical structures and the molecular property of interest is obscure. De novo design, on the other hand, models the chemical space of molecular structures and properties and seeks for the optimal candidates in a directed manner [2] instead of enumerating every possibility, thus facilitating the drug discovery process. Moreover, the flourishing of deep generative models in various domains such as large language models and image synthesis has endowed us an opportunity of applying deep learning to improving de novo drug design algorithms. Generative models including variational autoencoder (VAE) [3], generative adversarial networks (GAN) [4] and denoising diffusion probabilistic models (DDPM) [5], have been successfully adapted for molecular design. Initially, researchers tend to represent drugs with linear notations such as Simplified Molecular-Input Line-Entry System (SMILES) [6] due to its simplicity. Then long-short term memory (LSTM) networks were readily applied to encoding the SMILES notations, and VAE and GAN algorithms were utilized for generation [7, 8, 9, 10]. Such methods, however, suffered from low chemistry validity of generated molecules since the structural information is neglected in SMILES notations.