Preprint pair presents structure-aware graph neural network system for de novo protein design
Two related bioRxiv preprints describe a graph neural network framework and an interactive web server that combine sequence design with ESMFold-based structure prediction in a bidirectional design loop — neither has yet been peer-reviewed.
Two companion preprints posted to bioRxiv describe closely related work on computational de novo protein design using graph neural networks (GNNs) and ESMFold-based structure prediction.
The first preprint presents a GNN model trained to predict amino acid sequences from protein backbone coordinates, using a composite loss function that heavily weights structural similarity rather than sequence recovery alone. By integrating differentiable structure prediction via ESMFold, the model can directly optimise 3D conformational accuracy — an approach the authors argue better suits applications where structural fidelity to a target backbone is the primary goal.
The second preprint describes an interactive web server that wraps a structure-aware GNN for sequence design and an ESMFold-based structure predictor into a single bidirectional design loop accessible through a web browser, without requiring local software installation. The server includes a Kabsch alignment tool for superimposing predicted structures onto target backbones and visualises both directions of the sequence–structure mapping.
Both preprints are from bioRxiv and have not been peer-reviewed. The work is relevant to structural biologists, computational biologists, and protein engineers, and intersects with the active field of AI-assisted protein design. As preprints, the findings and methods should be treated as preliminary until they have undergone peer review and independent replication.
Sources
Read the original reporting — these are the public sources this summary draws from.
-
Primary sourcePreprint bioRxiv (Cold Spring Harbor Laboratory) · 2026-10-28An Interactive Bidirectional De Novo Protein Design System
-
Preprint bioRxiv (Cold Spring Harbor Laboratory) · 2026-10-28De novo Protein Generating Model via Structure-Aware Graph Neural Networks