Generative Modeling of Full-Atom Protein Conformations using Latent Diffusion on Graph Embeddings

Aditya Sengar, Ali Hariri, Daniel Probst, Patrick Barth, Pierre Vandergheynst · 2025

Read Generative Modeling of Full-Atom Protein Conformations using Latent Diffusion on Graph Embeddings by Aditya Sengar, Ali Hariri, Daniel Probst, Patrick Barth, Pierre Vandergheynst (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-17064.txt.

Introduction

Aditya Sengar and co-authors generate full-atom protein shapes with a diffusion model that works in a compressed space in a 2025 paper. Flexible receptors like the dopamine D2 receptor shift among many poses, and most generators drop atoms or miss that variety. Their pipeline first squeezes simulation frames into small graph embeddings.

Three pooling styles for the squeeze are compared, and a diffusion model learns to dream new embeddings. A decoder rebuilds every atom, side chains included, from the dreams. Sequential and residue-wise pooling matched the reference ensemble closely, atom for atom. The method offers drug hunters realistic receptor crowds to test against.

Catalog record

CREATORAditya Sengar, Ali Hariri, Daniel Probst, Patrick Barth, Pierre Vandergheynst
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE123 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-17064

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Generative Modeling of Full-Atom Protein Conformations using Latent Diffusion on Graph Embeddings Aditya Sengar, Ali Hariri, Daniel Probst, Patrick Barth, Pierre Vandergheynst Abstract Generating diverse, all-atom conformational ensembles of dynamic proteins such as G-protein-coupled receptors (GPCRs) is critical for understanding their function, yet most generative models simplify atomic detail or ignore conformational diversity altogether. We present latent diffusion for full protein generation (LD-FPG), a framework that constructs complete all-atom protein structures, including every side-chain heavy atom, directly from molecular dynamics (MD) trajectories. LD-FPG employs a Chebyshev graph neural network (ChebNet) to obtain low-dimensional latent embeddings of protein conformations, which are processed using three pooling strategies: blind, sequential and residue-based. A diffusion model trained on these latent representations generates new samples that a decoder, optionally regularized by dihedral-angle losses, maps back to Cartesian coordinates. Using D2R-MD, a 2 ​ μ ​ s 2\,\mu\text{s} MD trajectory (12 000 frames) of the human dopamine D 2 2 receptor in a membrane environment, the sequential and residue-based pooling strategy reproduces the reference ensemble with high structural fidelity (all-atom lDDT ∼ 0.7 \sim 0.7 ; C ​ α C\alpha -lDDT ∼ 0.8 \sim 0.8 ) and recovers backbone and side-chain dihedral-angle distributions with a Jensen–Shannon divergence < 0.03 <0.03 compared to the MD data. LD-FPG thereby offers a practical route to system-specific, all-atom ensemble generation for large proteins, providing a promising tool for structure-based therapeutic design on complex, dynamic targets. The D2R-MD dataset and our implementation are freely available to facilitate further research. 1 Introduction Proteins function as dynamic molecular machines whose biological activities critically depend on transitioning between distinct conformational states [ 1 , 2 ] . Landmark artificial intelligence methods like AlphaFold2 [ 3 ] and others [ 4 , 5 , 6 , 7 ] have advanced structure prediction but predominantly predict single static conformations, limiting their utility for systems with conformational heterogeneity. Accurate modeling of an ensemble of accessible conformations is essential to elucidate protein function and guide therapeutic design [ 8 , 9 , 10 , 11 ] . Crucially, these ensembles must explicitly represent all atomic details, particularly side chains, whose subtle conformational rearrangements often govern molecular recognition and catalytic mechanisms [ 12 ] . Despite rapid advancements, existing generative models frequently fall short in capturing the detailed dynamics of side-chain movements specific to particular proteins [ 13 ] . Many powerful methodologies have focused either on de novo backbone designs [ 14 , 15 , 16 ] or on generating static all-atom structures [ 17 , 18 ] , but neither produce comprehensive conformational ensembles. Ensemble-generating approaches—via perturbations of static predictors [ 19 , 20 ] , flow-matched variants [ 21 ] , or general MD-trained generators [ 22 , 23 ] —typically operate at the backbone or coarse-grained level. Consequently, they fail to capture the intricate all-atom rearrangements, particularly involving side chains, that are critical for function. Although promising latent space [ 24 , 25 , 26 ] and physics-informed models [ 27 , 28 ] have emerged, their capability to generate high-resolution all-atom ensembles from MD data reflecting functional transitions remains unproven, highlighting a significant unmet need for specialized generative frameworks. G protein-coupled receptors (GPCRs) provide a compelling example of dynamic systems wherein precise all-atom modeling is indispensable [ 29 ] . This large family of transmembrane receptors, comprising over 800 human members, is responsible for mediating most known transmembrane signal transduction [ 30 , 31 ] and is targeted by approximately 50% of all marketed drugs [ 32 , 33 ] . GPCR signaling involves conformational transitions among multiple states, frequently induced by ligand binding [ 34 ] , and occurs through intricate allosteric mechanisms where specific side-chain interactions are crucial [ 35 , 36 ] . Capturing such dynamic, atomically detailed landscapes is vital to understanding receptor signaling specificity [ 37 ] , biased agonism [ 38 ] , and the design of drugs targeting unique allosteric sites [ 39 , 40 , 41 ] , yet studying these events computationally remains formidable [ 42 , 43 , 44 ] . Current predictive methods for GPCR dynamics [ 45 ] do not generate the comprehensive all-atom conformational landscapes essential for mechanistic understanding. To address this critical requirement, we introduce Latent Diffusion for Full Protein Generation ( LD-FPG ), a generative framework designed to learn and generate diverse, all-atom conformational ensembles from existing MD simulation data of a target protein , explicitly including side-chain details. Rather than simulating new trajectories, LD-FPG leverages extensive MD datasets. Our approach employs a Chebyshev Spectral Graph Convolutional Network (ChebNet) [ 46 ] to encode all-atom MD snapshots into a compact latent representation. A Denoising Diffusion Probabilistic Model (DDPM) [ 47 ] is then trained to explore this learned latent manifold, and latent representations are decoded back into full all-atom Cartesian coordinates. We demonstrate our framework on extensive MD simulations of the human dopamine D2 receptor (D2R), systematically evaluating distinct decoder pooling strategies. Our primary contributions in this paper are: 1. To the best of our knowledge, we present the first latent diffusion modeling framework specifically tailored to generate complete, all-atom protein conformational ensembles , capturing both backbone and side-chain dynamics directly from MD simulations. 2. We introduce and critically evaluate a novel graph-based autoencoder architecture utilizing ChebNet combined with distinct decoder pooling strategies, offering insights into dynamic protein ensemble generation. 3. Using the D2R system, we demonstrate our method’s ability to generate high-fidelity ensembles, highlight the advantages of residue-based pooling, and assess the impact of auxiliary dihedral loss terms on generative accuracy. Our approach provides a computationally efficient tool for exploring complex dynamics in switchable proteins, supporting both fundamental mechanistic studies and drug discovery applications. The remainder of this paper is structured as follows: Section 2 reviews related work; Section 3 details our proposed methodology; Section 4 presents the experimental setup and results; and Section 5 summarizes our findings and outlines future directions. 2 Related Work Generative Models for Protein Design and Static Structure Prediction. Deep generative models have made significant strides in protein science, employing techniques such as diffusion [ 47 ] , flow-matching [ 48 , 49 ] , and learned latent spaces [ 24 , 50 ] to tackle complex structural tasks. For de novo backbone design, various methods, often inspired by the success of predictors like AlphaFold2 [ 3 ] and diffusion-based approaches like RFdiffusion [ 14 ] have emerged. Notable examples include FoldingDiff [ 15 ] , alongside flow-matching models such as FoldFlow2 [ 49 ] . Latent space strategies have also been pivotal in this domain; for instance, LatentDiff [ 24 ] generates novel protein backbones using an equivariant diffusion model within a condensed latent space, while Ig-VAE [ 50 ] utilizes a variational autoencoder for class-specific backbone generation (e.g., for immunoglobulins). While powerful for creating new folds (e.g., Proteina [ 16 ] ) or specific components, these methods typically target soluble proteins, often generate only backbone coordinates, and are not primarily designed for sampling multiple conformations of a specific, existing protein. The generation of complete static all-atom structures has also seen considerable progress. Models like Protpardelle [ 17 ] and Chroma [ 18 ] can produce full static structures from sequence information.

Read the complete text