An Interactive Bidirectional De Novo Protein Design System

Abstract

We present a lightweight, interactive server that combines two generative models, a structure-aware graph neural network (GNN) for de novo sequence design and an ESMFold-based structure predictor, into a single bidirectional design loop. The tool requires no software installation beyond a web browser, and it visualizes and quantifies both directions of the sequence-structure mapping between predicted sequences and structures. Through the same interface, an included user-accessible Kabsch alignment tool superimposes a predicted structure onto the target backbone. It reports the resulting root mean square deviation directly in the interface. Rather than a new model, this paper contributes the tool itself: an accessible, flexible, and openly sourced codebase that lowers the barrier to using and analyzing a model. This interactive server can also be used to validate design hypotheses interactively and reused by other researchers with their own checkpoints. While slightly more challenging, the codebase includes a section for users to add their own model structures while still using the interactive server code. This paper describes the system's architecture, interface, and intended workflow, and discusses its role as an open-source tool. All data and models used in this paper are derived from our previous companion papers [3, 4].