Structured State-Space Sequence Models for De Novo Drug Design

De novo drug design has witnessed remarkable progress since the advent of generative deep learning. Despite current progress, novel methods are needed to efficiently chart the vast chemical space in search of structurally diverse and bioactive molecules. Here we introduce a recently proposed generative deep learning approach, termed Structured State-Space Sequence Model (S4), into de novo molecular design for the first time. S4 has outperformed existing approaches in several application domains, but its potential in the molecular sciences is currently untapped. In this work, we systematically benchmark S4 with two state-of-the-art approaches – Long-Short Term Memory networks (LSTM) and Generative Pretrained Transformers (GPT) – on an array of tasks relevant to drug discovery. Our results show that S4 has a superior performance in 67% of the metrics analyzed and bears an unprecedented promise to design novel and structurally diverse bioactive molecules. Thanks to its capability to learn from the entire molecular sequence at once and its efficiency in generating promising molecular candidates, S4 offers the potential to become state-of-the-art in de novo drug design, and beyond