PANDORA: Diffusion-based Protein Conformation Generation

ICLR 2026 Conference Submission3871 Authors

Keywords: Protein conformation, diffusion, conformation generation

Abstract:

The field of protein design has garnered significant attention in AI for Science (AI4Science). While most existing studies focus on generating native protein structures for applications such as binding target discovery and drug design, this paper tackles a different problem. We aim to generate both native and non-native protein conformations, i.e., 3D structures along the folding pathway. This task holds significant potential for advancing key applications, including folding pathway prediction, drug discovery, and the study of protein misfolding diseases.
To address this challenge, we introduce $\underline{Pandora}$ (i.e., $\underline{P}$rotein Conform$\underline{a}$tio$\underline{n}$ $\underline{D}$iffusi$\underline{o}$n Model for Gene$\underline{ra}$tion), a novel diffusion-based framework designed to generate diverse, physically and chemically plausible protein backbone structures. By leveraging a diffusion architecture, Pandora captures a broad spectrum of folding patterns while adhering to biophysical constraints. Extensive experiments across multiple protein folding pathway datasets demonstrate the effectiveness and generalizability of our approach in producing realistic and biologically meaningful conformations. The implementation of Pandora is publicly available at [https://anonymous.4open.science/r/Pandora-71AD](https://anonymous.4open.science/r/Pandora-71AD).

Primary Area: generative models

Submission Number: 3871