In this work, we introduce AutoFragDiff, a fragment-based autoregressive
diffusion model for generating 3D molecular structures conditioned on target
protein structures. We employ geometric vector perceptrons to predict atom
types and spatial coordinates of new molecular fragments conditioned on
molecular scaffolds and protein pockets. Our approach improves the local
geometry of the resulting 3D molecules while maintaining high predicted binding
affinity to protein targets. The model can also perform scaffold extension from
user-provided starting molecular scaffold.

Introduction to AutoFragDiff: Generating 3D Molecular Structures

In the field of drug discovery, the ability to generate accurate and diverse 3D molecular structures is crucial for identifying potential drug candidates. Researchers have long been exploring various generative models to achieve this goal. In this context, AutoFragDiff presents a new approach that leverages fragment-based autoregressive diffusion modeling.

AutoFragDiff addresses the challenge of generating 3D molecular structures conditioned on target protein structures. This means that the model takes into account the specific protein targets and their binding affinity when generating the molecular structures. This is a noteworthy development, as it enhances the predictive accuracy and specificity of the generated structures.

The Role of Geometric Vector Perceptrons

One of the key elements of AutoFragDiff is its use of geometric vector perceptrons. These perceptrons are employed to predict the types of atoms and their spatial coordinates in newly generated molecular fragments. By incorporating geometric information, AutoFragDiff goes beyond simple predictive models and introduces a multi-disciplinary approach that embraces concepts from both molecular chemistry and geometric modeling.

The incorporation of geometric vector perceptrons offers several advantages. Firstly, it improves the local geometry of the resulting 3D molecules, enhancing their overall structural integrity. This is crucial for accurately representing the molecular interactions between the generated structures and the target protein. Furthermore, the use of geometric vector perceptrons enables precise control over the arrangement and positioning of atoms within the generated fragments.

Improving Binding Affinity while Maintaining Local Geometry

A unique aspect of AutoFragDiff is its ability to balance two crucial considerations: improving local geometry and maintaining high predicted binding affinity to protein targets. Generating 3D molecular structures that exhibit both accurate geometry and strong binding affinity is a challenging task, often necessitating trade-offs between the two. However, AutoFragDiff aims to strike a balance by leveraging fragment-based autoregressive diffusion modeling.

By conditioning the generative model on molecular scaffolds and protein pockets, AutoFragDiff can focus on improving the local geometry of the generated structures while still preserving their binding affinity. This is a promising development, as it opens up avenues for more precise drug design and optimization.

Scaffold Extension for Molecular Design

In addition to generating 3D molecular structures from scratch, AutoFragDiff also offers scaffold extension capabilities. This means that given a user-provided starting molecular scaffold, the model can further expand and refine it. Scaffold extension is a valuable feature for medicinal chemists and drug discovery researchers as it allows them to iteratively design and optimize molecules based on a known starting point.

By combining fragment-based autoregressive diffusion modeling with scaffold extension, AutoFragDiff empowers researchers to generate highly tailored molecular structures that exhibit desired binding affinity to specific protein targets. This seamless integration of generative models with molecular design tools holds great potential in accelerating the drug discovery process and facilitating more effective treatments.


AutoFragDiff presents a novel approach for generating 3D molecular structures conditioned on target protein structures. By incorporating fragment-based autoregressive diffusion modeling and employing geometric vector perceptrons, this multi-disciplinary model improves both the local geometry and the binding affinity of the generated structures. Additionally, the scaffold extension capability enhances its utility in molecular design tasks. AutoFragDiff showcases the potential of integrating concepts from molecular chemistry, machine learning, and geometric modeling to advance drug discovery efforts.

Read the original article