Abstract
The discovery of superior molecular solutions through computational methods is critical for innovative technologies and their role in addressing pressing resources, health, and environmental issues. Despite its short timespan, the synergetic application of deep learning to inverse molecular design has outpaced decades of theoretical efforts, bearing promise to transform current molecular design paradigms. Herein, we provide an overview of the element of computational inverse molecular design and offer our views on current limitations and outstanding challenges. In our perspective, three main directions are identified for each element and analyzed in terms of their merits and relevant novel deep learning developments. For the molecular representations element, Graph Neural Networks (GNNs), grids, and knowledge graphs (KGs) are discussed for enhancing the expressivity, complexity, descriptivity of relevant molecular information, respectively. Second, chemical text mining, accelerated quantum chemical calculations, and transfer learning are explored to augment the size and the accuracy of current property data and predictive models. Last, emerging trends in design methods including generative modeling, reinforcement learning (RL), and active learning (AL) are examined for optimizing not only computational costs, but also experimental and simulation efforts. The presented discussions are aimed at catalyzing progress and interdisciplinary collaborations toward general-purpose inverse design frameworks.