AlphaFolding: 4D Diffusion for Dynamic Protein Structure Prediction with Reference and Motion Guidance

TL;DR

This work tackles dynamic protein structure prediction by introducing AlphaFolding, a 4D diffusion model that generates multi-time-step protein trajectories while modeling both backbone and side-chain atoms. The approach combines a unified SE(3) diffusion process with a reference network to preserve structural consistency and a motion-alignment module to enforce temporal coherence across frames. It demonstrates strong performance on benchmark datasets up to 256 amino acids across 32 time steps, outperforming state-of-the-art static diffusion methods on long-horizon trajectory generation and showing robust generalization to unseen proteins. By leveraging MD data and explicit temporal guidance, the method provides a scalable, efficient framework for simulating realistic protein dynamics with potential impact on drug design and experimental planning.

Abstract

Protein structure prediction is pivotal for understanding the structure-function relationship of proteins, advancing biological research, and facilitating pharmaceutical development and experimental design. While deep learning methods and the expanded availability of experimental 3D protein structures have accelerated structure prediction, the dynamic nature of protein structures has received limited attention. This study introduces an innovative 4D diffusion model incorporating molecular dynamics (MD) simulation data to learn dynamic protein structures. Our approach is distinguished by the following components: (1) a unified diffusion model capable of generating dynamic protein structures, including both the backbone and side chains, utilizing atomic grouping and side-chain dihedral angle predictions; (2) a reference network that enhances structural consistency by integrating the latent embeddings of the initial 3D protein structures; and (3) a motion alignment module aimed at improving temporal structural coherence across multiple time steps. To our knowledge, this is the first diffusion-based model aimed at predicting protein trajectories across multiple time steps simultaneously. Validation on benchmark datasets demonstrates that our model exhibits high accuracy in predicting dynamic 3D structures of proteins containing up to 256 amino acids over 32 time steps, effectively capturing both local flexibility in stable states and significant conformational changes. URL: https://fudan-generative-vision.github.io/AlphaFolding/#/

Figures (13)

• Figure 1: 4D dynamic protein prediction. Given an initial 3D structure for reference, our model predicts dynamic proteins at the following 32 time steps simultaneously. We present the predicted 3D protein structures at the intermediate time steps for illustration.
• Figure 2: The overview of our proposed approach. The diffusion-based generative model that takes the reference structure and corresponding residue sequence as input and produces a sequence of denoised 3D protein structures as output. We use the 3D structure embedder and GeoFormer for embedding the 3D protein structures and residue sequences, respectively. The Invariant Point Attention (IPA) updates node features by integrating information from the explicit frames of residues. The Reference Network and Motion Alignment module are based on the reference 3D protein structure to capture a sequence of 3D protein dynamics. The entire generative model is formulated as a score-based diffusion model, with node and edge feature embedding updated through the EdgeUpdate and BackboneUpdate modules.
• Figure 3: Structure of spatial module and motion alignment. The spatial module encodes the structural characteristics of the reference 3D protein structure from reference network to preserve its features during dynamic structure generation. The motion alignment is comprised of stacked temporal transformer layers and used to generate protein dynamics. The depth of the color indicates different time steps.
• Figure 4: Distribution Analysis. Sample distribution over first two TIC components for different proteins. The darker the points, the higher their frequency of occurrence. The blue curve represents the kernel density distribution estimated from the MD data.
• Figure 5: Qualitative Result. Our model prediction (blue) and the MD simulation results (red). In the first line, texts on the left refer the to protein's PDB ID and the corresponding chain, and the time on the right represents the time it takes for the reference structure to transition to this predicted structure.