Deep Generative Modeling of Protein Conformations: A Comprehensive Review

Abstract

Proteins are dynamic macromolecules whose functions are intricately linked to their structural flexibility. Recent breakthroughs in deep learning have enabled accurate prediction of static protein structures. However, understanding protein function is more complex. It often requires access to a diverse ensemble of conformations. Traditional sampling techniques exist to help with this. These include molecular dynamics and Monte Carlo simulations. These techniques can explore conformational landscapes. However, they have limitations as they are often limited by high computational cost and suffer from slow convergence. In response, deep generative models (DGMs) have emerged as a powerful alternative for efficient and scalable protein conformation sampling. Leveraging architectures such as variational autoencoders, normalizing flows, generative adversarial networks, and diffusion models, DGMs can learn complex, high-dimensional distributions over protein conformations directly from data. This survey on generative models for protein conformation sampling provides a comprehensive overview of recent advances in this emerging field. We categorize existing models based on generative architecture, structural representation, and target tasks. We also discuss key datasets, evaluation metrics, limitations, and opportunities for integrating physics-based knowledge with data-driven models. By bridging machine learning and structural biology, DGMs are poised to transform our ability to model, design, and understand dynamic protein behavior.

1. Introduction

Proteins are fundamental biomolecules that carry out essential functions in living organisms, from catalyzing metabolic reactions to transmitting signals and forming structural components. The function of a protein is intimately tied to its three-dimensional (3D) structure, which is determined by the sequence of amino acids through a complex folding process. Consequently, predicting a protein’s structure from its amino acid sequence, a problem known as protein structure prediction, has been a central challenge in computational biology for decades.

Recent breakthroughs, such as AlphaFold2 [1], have dramatically advanced static structure prediction, achieving near-experimental accuracy and reshaping the landscape of structural biology. These models offer valuable insights for drug discovery [2,3], enzyme engineering [4], protein-protein interaction prediction [5,6], construction of a comprehensive protein database [7], along with experimental validation of protein-protein complexes [8], and study of biologically relevant states [9]. However, they provide only a single, static snapshot of a protein’s structure. In reality, proteins are inherently dynamic and often adopt multiple conformations to perform their biological functions.

This is a significant milestone that is likely to support a variety of structure-function studies by computational and molecular biologists alike, but it is still a first-order approximation to protein structure modeling. To fully understand and model protein function and characterize the various molecular interactions in which a protein molecule participates in a cell, including interactions of protein targets with therapeutic, small-molecule leads, we need to account for the intrinsic structural plasticity of protein molecules. We have many examples of proteins that switch between different structures to regulate interactions with different molecular partners in the cell. Nowhere is this structural plasticity more evident than in the spike protein of the SARS-CoV-2 virus, where open-to-close structural changes are key to the ability of the protein to evade our immune system just long enough to then be able to position itself and bind with the ACE2 receptors in the lungs [10].

Another relevant example is the Calmodulin protein [11,12], which undergoes significant conformational changes upon calcium binding. This is evident upon the experimental observation of the transition of its motifs from “closed” to “open” during the binding process with other molecular partners. This transition facilitates Calmodulin to play a crucial role in regulating numerous cellular processes and maintaining overall cellular health. These examples further strengthen the argument that finding the conformations that a protein adopts to regulate interactions with molecular partners in the cell is an important open problem in molecular biology.

Therefore, obtaining a broad view of the protein conformation space is the main objective of the computational research proposed here. This problem, however, poses an outstanding challenge and has been approached by computational biologists and computer scientists alike with a variety of computational methods over the decades. Most prominent among these are those based on optimization, which can be broadly characterized into three groups: those based on numerical optimization, stochastic optimization, and hybrid methods.

In numerical optimization methods, a physics-based objective function is defined and optimized through numerical simulation to navigate the conformational space and produce physically relevant conformations. These are known as molecular dynamics (MD) in the computational biology literature. One of the prominent ways researchers augment potential energy functions during the simulation process is through the selection of collective variables [13,14,15,16,17]. CVs represent a set of Cartesian coordinates through a collection of basis functions that can act as a lower-dimensional representation of a complex system. This system of representation subsequently reduces the complexity of the system, leading to its inherent advantage compared to other approximations of the molecular dynamics algorithm. Other approaches include umbrella sampling [18], isolating low frequency vibrations [19], and Temperature-accelerated molecular dynamics (TAMD) [20], adaptive sampling [21] among many others. Despite the vast array of research conducted in this field, it suffers from problems such as inability to reach biologically relevant timescales [22] along with high computational cost [23] and force field limitations [24].

Another powerful yet less computationally expensive approach based on stochastic optimization is Evolutionary Algorithms (EA). EA is a biologically inspired algorithm based on works such as [25,26] that mimics human evolution, and has been successfully applied to the protein structure prediction domain. Its stochastic nature is conducive to exploring a vast conformational space and identifying energetically favorable structures [27,28,29,30,31,32,33]. Despite its promises, this approach is susceptible to problems such as slow convergence [34]. Finding the global optimum also becomes challenging with the increasing randomness of the energy optimization landscape [35].

A recent line of work involving deep learning techniques for conformation sampling revolves around adopting a single structure predictor, AlphaFold2 [1], or its evolutions like Alphafold3 [36] or EsmFold [37], for the task. Although AlphaFold is primarily designed for the prediction of structure from sequence, its adaptations and integration with generative models open new possibilities for sampling and exploring conformational diversity [38]. One notable work built on AlphaFold2 is MSA subsampling [39], used as a benchmark in a variety of literature involving DGMs in this survey.

Deep generative models (DGMs) offer a fundamentally different approach to sampling protein conformational space. Traditional physics-based simulations, such as molecular dynamics (MD), are limited by small time steps and high computational cost, making it difficult to access the full range of biologically relevant protein motions, especially for large systems or intrinsically disordered regions (IDRs).

In contrast, DGMs learn a parametric model of the equilibrium distribution of protein conformations directly from data, enabling rapid generation of diverse, independent structural samples. This allows scalable exploration of conformational landscapes that are otherwise prohibitively expensive to access with conventional simulations.

In this review, we focus on deep generative models with explicit probabilistic mechanisms, including variational autoencoders (VAEs), generative adversarial networks (GANs), normalizing flows, and diffusion models. This review offers a scope broader than that of Barethiya et al. [40], which focuses on benchmarking specific VAE and diffusion models, while this work provides a general overview of this nascent field. Our work encompasses a more diverse range of protein types compared to Erdhos et al. [41] and addresses a fundamentally different objective than previous studies [42,43], which focus specifically on deep learning models for protein design rather than the broader analytical framework we present here. Figure 1 presents a review of key studies on the integration of Multiple Sequence Alignment, Sequence, and Structural Data in Protein Conformational Analysis and Trajectory Sampling.

The paper is arranged as follows: in the Section 3, we discuss the basics of the generative models surveyed; Section 4 discusses the relevant papers and their applications in the context of protein conformation modeling; Section 5 discusses popular datasets used for conformation modeling; Section 6 discusses different metrics used to compare generated conformations with the reference dataset followed by the Section 7 to discuss the challenges and opportunities that lie ahead in the broader context of modeling protein structural plasticity.

2. Problem Definition

In this review, we explore deep generative models that address two closely related but distinct challenges in structural biology: protein conformation sampling and protein trajectory generation.

2.1. Protein Conformation Sampling

Protein conformation sampling refers to the process of generating structural configurations that reflect the underlying equilibrium distribution of a protein. Ideally, a sampling method produces conformations $\mathbf{x}$ such that:

$$p_S\left(\mathbf{x}\right)\approx p_{\mathrm{eq}}\left(\mathbf{x}\right)$$

(1)

where $p_S\left(\mathbf{x}\right)$ is the sampling distribution induced by the method, and $p_{\mathrm{eq}}\left(\mathbf{x}\right)$ is the true equilibrium distribution. Traditional approaches such as molecular dynamics (MD) and Monte Carlo simulations are commonly used for this task.

The goal is to generate a representative ensemble of conformations:

$$\{{\mathbf{x}}_1,{\mathbf{x}}_2,\dots ,{\mathbf{x}}_N\}$$

(2)

This captures the diversity and relative probabilities of protein structures under physiological conditions. Such ensembles are crucial for studying conformational variability, functional motions, and interaction propensities.

The equilibrium distribution that governs these conformations is given by Boltzmann statistics:

$$p_{\mathrm{eq}}\left(\mathbf{x}\right)={\displaystyle \frac{1}{Z}}{e}^{-\beta E\left(\mathbf{x}\right)}$$

(3)

where $E\left(\mathbf{x}\right)$ the potential energy of conformation $\mathbf{x}$, $\beta =1/\left(k_{B}T\right)$ is the inverse thermal energy, and Z is the partition function ensuring normalization. This formulation implies that lower-energy conformations are exponentially more probable than higher-energy ones.

2.2. Protein Trajectory Generation

While conformation sampling captures thermodynamic diversity, protein trajectory generation focuses on the time-resolved evolution of protein structure. These trajectories are typically generated via MD simulations, which numerically integrate Newton’s equations of motion under a specified force field and thermodynamic ensemble [44,45].

For a protein consisting of N atoms, a trajectory is defined as:

$$\mathbf{X}\left(t\right)=\left[{\mathbf{x}}_1\left(t\right),{\mathbf{x}}_2\left(t\right),\dots ,{\mathbf{x}}_N\left(t\right)\right]\in {\mathbb{R}}^{3N}$$

(4)

where ${\mathbf{x}}_i\left(t\right)\in {\mathbb{R}}^3$ represents the position vector of the atom i at time t, and $\mathbf{X}\left(t\right)$ denotes the full protein conformation at that time.

Together, conformation sampling and trajectory generation provide complementary insights: the former captures the statistical ensemble of accessible structures, while the latter reveals the dynamical transitions between them.

3. Basics of Deep Generative Modeling

This section provides an overview of the four primary classes of deep generative models that have been applied to protein conformation sampling. Each architecture offers distinct advantages and trade-offs in terms of training stability, computational efficiency, and the ability to model complex conformational distributions. An overview of the architectures of the models discussed in this literature can be found in Figure 2. We begin with an examination of generative adversarial networks.

3.1. Generative Adversarial Networks

Generative Adversarial Networks (GANs) are generative models composed of two neural networks: a generator G and a discriminator D trained in a competitive learning process. The generator G learns to map random noise $z\sim p_z\left(z\right)$ to synthetic samples $G\left(z\right)$, while the discriminator D attempts to distinguish between real samples $x\sim p_{\mathrm{data}}\left(x\right)$ and generated ones. The objective is formulated as:

$$\underset{G}{min}\underset{D}{max}{\mathbb{E}}_{x\sim p_{\mathrm{data}}}[logD\left(x\right)]+{\mathbb{E}}_{z\sim p_z}[log(1-D\left(G\left(z\right)\right))]$$

(5)

Here, ${\mathbb{E}}_{x\sim p_{\mathrm{data}}}[logD\left(x\right)]$ represents the discriminator’s confidence in correctly identifying real samples, while ${\mathbb{E}}_{z\sim p_z}[log(1-D\left(G\left(z\right)\right))]$ penalizing it for mistakenly identifying generated samples as real.

In protein conformation sampling, GANs offer a powerful alternative to molecular dynamics by directly learning the distribution of conformational states from simulation data. Once trained, they can generate statistically independent structures that capture both local flexibility and large-scale conformational transitions, providing ensembles at a fraction of the computational cost. Conditional GANs, such as idpGAN, extend this capability to sequence-dependent sampling, enabling generalization to unseen proteins and enhancing transferability across systems [46]. Beyond ensemble generation, GANs also provide a framework for approximating protein trajectories by sampling continuous states from the learned distribution, thereby reconstructing dynamic pathways without the need for long MD simulations. This highlights their potential to accelerate exploration of conformational landscapes and to generate physically consistent trajectories that improve our understanding of biomolecular dynamics [47].

3.2. Variational Autoencoders

Variational Autoencoders (VAEs) [48,49] learn a probabilistic mapping between data and a latent space through variational inference. It consists of two neural networks: an encoder $q_{\varphi }(z\mid x)$, which maps an input x to a distribution over latent variables z, and a decoder $p_{\theta }(x\mid z)$, which reconstructs data from latent samples. The training objective maximizes a variational lower bound (ELBO) on the data log-likelihood:

$$\mathcal{L}(\theta ,\varphi ;x)={\mathbb{E}}_{z\sim q_{\varphi }(z\mid x)}[log{p}_{\theta }(x\mid z)]-D_{\mathrm{KL}}(q_{\varphi }(z\mid x)\parallel p\left(z\right))$$

(6)

Here, the first term encourages accurate reconstruction of the input, while the second term regularizes the latent space by minimizing the Kullback–Leibler divergence between the approximate posterior $q_{\varphi }(z\mid x)$ and a prior $p\left(z\right)$, typically a standard Gaussian.

For application, VAEs have been shown to effectively capture the complex, high-dimensional conformational landscapes of both intrinsically disordered and ordered proteins [50]. By learning a smooth and structured latent representation, VAEs can generate novel conformations that extend beyond those sampled in short MD trajectories, thereby enriching conformational ensembles with states that would otherwise require prohibitively long simulations to observe. Moreover, points interpolated within the latent space correspond to physically plausible intermediate structures, enabling reconstruction of continuous transition pathways and offering a trajectory-like view of protein dynamics. Recent advances, such as latent space assisted adaptive sampling (LAST) [51], demonstrate that VAE-generated conformations can serve as efficient starting points for iterative MD simulations, accelerating the discovery of transition pathways and expanding conformational coverage. Thus, VAEs provide a powerful framework for augmenting molecular dynamics, enhancing conformational ensemble generation, and approximating protein trajectories in a computationally efficient manner.

3.3. Flow

Flow-based models, or normalizing flows [52,53], are generative models that learn an invertible transformation $f_{\theta }$ to map a simple base distribution $p_z\left(z\right)$ (such as a Gaussian) into a complex data distribution $p_x\left(x\right)$, such that $f_{\theta }\left(z\right)=x$ and $f_{\theta }^{-1}\left(x\right)=z$. This framework enables exact likelihood estimation through the change-of-variables formula:

$$log{p}_x\left(x\right)=log{p}_z\left(f_{\theta }^{-1}\left(x\right)\right)+log\left|det\left({\displaystyle \frac{\partial f_{\theta }^{-1}\left(x\right)}{\partial x}}\right)\right|$$

(7)

Here, the determinant of the Jacobian matrix ${\displaystyle \frac{\partial f_{\theta }^{-1}\left(x\right)}{\partial x}}$ measures how the transformation $f_{\theta }^{-1}$ locally scales or distorts volume around a point x. This term is essential for adjusting the density under the transformation, allowing the model to compute the exact likelihood x based on how much space it occupies relative to the base distribution.

In protein conformation sampling, flow models learn invertible mappings from a latent space to complex 3D structures. However, the requirement for invertibility and a tractable Jacobian determinant limits architectural flexibility. Flow Matching [54] addresses this by learning a time-dependent velocity field $v_{\theta }(x,t)$ that transports samples from a base distribution $x\left(0\right)$ to a target structure $x\left(1\right)$ using linear interpolation:

$$x\left(t\right)=(1-t)x\left(0\right)+tx\left(1\right),\mathcal{L}\left(\theta \right)={\mathbb{E}}_{x\left(0\right),x\left(1\right),t}\left[{∥v_{\theta }(x\left(t\right),t)-(x\left(1\right)-x\left(0\right))∥}^2\right]$$

(8)

This formulation $x\left(t\right)$ lies along a straight path between the start and target conformations. The model is trained to match the constant velocity $x\left(1\right)-x\left(0\right)$ at each point along this path. This enables smooth transitions between structures without requiring an invertible transformation, making Flow Matching suitable for modeling the continuous and high-dimensional landscape of protein conformations.

In practice, flow-based generative models have shown strong potential for protein conformation sampling by directly learning distributions over structural ensembles rather than relying solely on costly molecular simulations. Recent advances, such as AlphaFlow and P2DFlow [55,56], demonstrate that flow matching can be integrated with structural priors (e.g., AlphaFold/ESMFold predictions or MD-derived ensembles) to generate physically plausible protein conformations. By leveraging SE(3)-equivariant architectures and energy-informed priors, these models can capture domain motions, intermediate states, and conformational transitions that underlie protein function. Importantly, flow-based approaches enable trajectory-like sampling of conformations, providing a data-driven alternative to molecular dynamics for exploring equilibrium distributions and transition pathways. This positions flow matching not only as a powerful tool for ensemble generation but also as a promising framework for efficient protein trajectory modeling, with implications for understanding allosteric regulation, stability, and dynamics in diverse biological contexts.

3.4. Diffusion

Diffusion-based generative models  operate by learning to reverse a stochastic process that progressively adds noise to data, transforming it into a tractable prior distribution. In the forward process, a data sample $x_0$ is gradually perturbed through a sequence of latent variables $x_1,\dots ,x_T$ using Gaussian noise, typically following a formulation that preserves variance:

$$q(x_t\mid x_{t-1})=\mathcal{N}(x_t;\sqrt{1-{\beta }_t}{x}_{t-1},{\beta }_{t}I)$$

(9)

The model is trained to approximate the reverse transitions:

$$p_{\theta }(x_{t-1}\mid x_t)=\mathcal{N}(x_{t-1};{\mu }_{\theta }(x_t,t),{\Sigma }_{\theta }(x_t,t))$$

(10)

Here, $x_t$ represents the noisy version of the data at time step t, and ${\beta }_t$ is a predefined noise schedule that controls the variance of the added noise. In the forward process, noise is added step-by-step, causing $x_t$ to gradual loss of information from the original data. The reverse process aims to reconstruct the data by predicting the mean ${\mu }_{\theta }(x_t,t)$ and variance ${\Sigma }_{\theta }(x_t,t)$ of the denoised sample at each step, where $\theta$ represents the parameters of a neural network.

In protein conformation sampling, diffusion models provide a scalable framework for exploring the high-dimensional energy landscape beyond molecular dynamics. By denoising random noise into physically plausible structures, they generate diverse conformations approximating the equilibrium ensemble. Advances with SE(3)-equivariant architectures and physics-guided terms enable sampling that respects structural symmetries and energy constraints [57,58,59]. This not only reveals hidden binding pockets and transient states crucial for allosteric regulation and ligand recognition, but also offers a surrogate for trajectory generation, where denoising implicitly traces pathways between metastable states. Thus, diffusion models combine equilibrium fidelity with kinetic relevance, making them a powerful tool for studying protein dynamics and guiding drug discovery.

4. Taxonomy

Having established the theoretical foundations of deep generative models, we now examine their practical applications to protein conformation sampling and trajectory generation. This section is organized by model architecture, discussing the specific implementations, datasets, and performance characteristics of each approach. We begin with GAN-based methods, which, while less prevalent in this domain, have shown promise in specific applications.

4.1. GAN

4.1.1. Application of Generative Adversarial Networks

The application of GAN for protein conformation sampling has been limited but diverse. Pang et al. [60] propose DeepPath, an active learning (AL) framework to model protein transition pathways that generates low-energy transitions using Active Learning, eliminating the need for a large pre-existing training datasets and incorporating physical feasibility using all-atom MM force fields- empirical potential energy functions that describe interatomic interactions through bonded (bonds, angles, dihedrals) and non-bonded (electrostatic, van der Waals) terms- as an oracle. Despite the lack of comparison with other deep learning based methods, this approach opens up an interesting research avenue for flow and diffusion-based generative models.

To address the challenge of modeling intrinsically disordered proteins (IDPs) [61], Janson et al. developed idpGAN [46], a Transformer-based conditional GAN trained on coarse-grained MD simulation data to rapidly generate diverse, physically realistic conformations. This method utilized multiple discriminators to ensure structural validity and later retrain idpGAN on atomistic simulation data to show that this approach is also extensible to higher-resolution conformational ensemble generation.

A unique application of Generative Adversarial Networks (GANs) has been employed by Liu et al. [62] to tackle the problem of conformation clustering, referred to as AAE-GS, which demonstrates effective clustering performance on a small protein, Trp-cage (PDB ID: 2JOF). Furthermore, the authors provide a model trajectory with a sufficiently long simulation time to address the absence of standardized datasets for assessing clustering quality.

Bouvier et al. [63] develop a generative adversarial network (GAN) to generate stable oligosaccharide conformations on the fly. The architecture includes a generator that maps latent space points to candidate structures and a discriminator that distinguishes these from true energy minima. Subsequently, the generator gets better at ‘fooling’ the discriminator by creating ever more realistic conformations in a zero-sum game.

4.1.2. Discussion

While GAN-based approaches are not yet the dominant paradigm in protein conformation modeling, they offer unique advantages, particularly in scenarios where traditional simulation methods struggle, such as modeling intrinsically disordered protein (IDP) conformations.

However, the limited adoption of GANs in this domain is understandable in retrospect. These models are known to suffer from fundamental challenges, including training instability, mode collapse, and convergence difficulties [64,65,66], which can hinder their reliability and generalization in sensitive biological applications.

4.2. VAE

Compared to generative adversarial networks (GANs), variational autoencoders (VAEs) [49] offer several advantages that make them particularly appealing for modeling protein conformational landscapes. First, VAEs are built on a probabilistic framework that explicitly models a continuous and smooth latent space, where each point corresponds to a plausible protein conformation. This latent space is trained to follow a prior distribution (commonly Gaussian) via the Kullback-Leibler (KL) divergence term in the objective. At the same time, a reconstruction loss ensures that sampled latent vectors decode into realistic conformations.

This structured latent space allows VAEs to act as a low-dimensional proxy for the protein energy landscape, where nearby points correspond to structurally similar states with gradual transitions between them. As a result, VAEs are especially suited to capture the continuum of conformations relevant for flexible proteins or proteins undergoing allosteric changes. Furthermore, the latent space can be systematically explored or interpolated to generate novel, previously unsampled conformations, including rare intermediate states that may be difficult to access via molecular dynamics simulations alone.

4.2.1. Protein Conformation Sampling

VAE has been used for sampling the protein conformation space by leveraging the trained latent space. The principle is to convert high-dimensional protein structural data into a continuous, low-dimensional representation using the VAE Encoder, followed by a search in this space guided by a structure quality metric, and then use the decoder to obtain a backbone-only or full-atom protein structure.

Mansoor et al. [67] leveraged RosettaFold [68] to retrieve the full atom structure to generate an ensemble of undruggable K-Ras protein [69,70]. Tian et al. [71] and Xiao et al. [72] show that the latent space in the VAE can be used to generate unsampled protein conformations when trained on relatively short molecular dynamics simulation data, followed by works from [73,74,75,76,77]. Among these, Degiacomi et al. [75] test the VAE for a complex protein-protein docking scenario (using HIV-1 hexameric capsomer) to account for broad hinge motions taking place upon binding.

Ruzmetov et al. [73] proposes the Internal Coordinate Net (ICoN), which subsequently generates novel synthetic conformations through learned latent space. To facilitate full atom generation, the authors utilize bond-angle-torsion (BAT)-based vector representation for comprehensive sampling of intrinsically disordered protein A$\beta$42’s conformational landscape, validated through metrics like number of conformations within energy cutoff and number of synthetic conformations.

Another notable work by Zhu et al. [50], utilizes short MD simulations to enable enhanced sampling of IDP conformations by introducing a probabilistic latent space that performs better than capturing the conformational landscape than standard autoencoders when measured in terms of metrics such as Pearson Correlation Coefficient (PCC) and Root Mean Squared Distance (RMSD) (details in Section 6).

4.2.2. Protein Trajectory

Molecular dynamics (MD) simulations are a widely used tool for studying protein conformations and dynamics at atomic resolution [44,45,78]. Despite their accuracy, conventional MD simulations often suffer from limited sampling efficiency due to the tendency to become trapped in local energy minima, making it difficult to explore the full conformational landscape [79,80,81].

Recent approaches have sought to overcome this challenge by integrating MD with deep learning techniques. Last [51] demonstrated large conformational changes in a shorter time compared to traditional MD and Stochastic Dynamics Simulation (SDS) [82] methods. After initial training with conformations obtained via short MD simulation, similar to [50], further seeds are selected from the learned latent space, similar to Gaussian accelerated MD simulation [83].

Similarly, Moritsugu et al. [84] introduced VAE-driven Multiscale enhanced sampling (MSES) for modeling the transition between open and closed states of RNA binding protein, again using short MD simulation(100 ns long) as VAE training data. The authors utilize “Motion Tree” [85] as a structural ensemble to model the dynamics of different regions of a given protein structure to model multiple conformations.

4.2.3. Other Applications

Kleiman et al. [86] propose VAMPNet to assign probabilistic memberships to metastable states using softmax outputs, allowing each conformation to belong to multiple states with varying confidence. Bozkurt et al. [87] aim to provide interpretable embeddings, recognizing that the latent space can be multimodal Gaussian. Albu et al. [88] show that the VAE latent space can capture structural information characterizing a certain protein superfamily and thus, can characterize an unseen protein from the same superfamily, using the Structural Alphabet [89] and Angles representation.

4.2.4. Discussion

VAE-based approaches discussed above and summarized in

Table 1. Compressed summary of variational autoencoders for protein conformation sampling. Columns: atomistic granularity (Gran.), Multiple Sequence Alignment conditioning (MSA), equivariance (Equiv.), and evaluation metrics (Metric). Dataset column is placed last.

Paper	Input	Output	Gran.	MSA	Equiv.	Metric	Dataset
Bandyopadhyay et al. [91]	Pairwise C$\alpha$ distances	Latent space repr.	Residue-level ($C\alpha$–$C\alpha$) representation	No	No	VAMP-2 score, Chapman–Kolmogorov (CK) test	200 MD traj (30–100 ns)
Gupta et al. [77]	Heavy-atom Cartesian coords	Reconstruct coords	Heavy-atom level	No	No	RMSD, Experimental data	Q15, A$\beta$40, ChiZ (1–3.5 $\mu$s)
Xiao et al. [72]	Normalize heavy-atom coords	Reconstruct + 3D confs	Heavy-atom level	No	No	Correlation coefficient, RMSD	Calmodulin, $\beta$-lactamase, Ubiquitin MD
Degiacomi et al. [75]	Heavy-atom Cartesian coords	Reconstruct coords	Backbone, C$\beta$ atoms	No	No	RMSD	MurD open/closed MD
Ward et al. [90]	Heavy-atom Cartesian coords	Reconstruct coords	Heavy-atom level	No	No	RMSD, ROC–AUC, Correlation	TEM $\beta$-lactamase, HIV-1 capsomer MD
Tian et al. [51]	Normalize heavy-atom coords	2D latent embeddings	Heavy-atom level	No	No	RMSD, Sampling efficiency	100–200 ps MD (NVT/NPT)
Jin et al. [76]	Selected atom coords (snapshots)	New conformation coords	Heavy-atom level	No	No	RMSD	L-Ala, Calmodulin (20 ns MD)
Mansoor et al. [67]	2D RoseTTAFold templates	2D → 3D structures	Backbone, C$\beta$ atoms	Yes	No	RMSD, CCE	K Ras (10 ns MD)
Ruzmetov et al. [73]	BAT vector representation	All-atom conformation	All-atom	No	Yes	RMSD, Rg	Amyloid-$\beta$ (1 $\mu$s MD)

excel at sampling unseen conformations using extensive MD simulation as training data along with modeling complex protein-protein interaction in some cases, e.g., [90]. Although this is an exciting application, in general, the VAE architecture needs extensive hyperparameter tuning for balancing encoder-decoder performance. Moreover, a VAE trained on one protein system may not generalize well to others, and users must still determine suitable force field parameters for the specific system under study, essentially adding another variable into the mix to generate computationally expensive simulation data.

4.3. Flow

Flow-based models [48,52,53] present an emerging approach in the field of protein conformation sampling. Based on recent developments such as flow matching that combines optimal transport, neural Ordinary Differential Equations, and normalizing flows [54,92,93], flow models offer a principled way to model complex, high-dimensional distributions over protein structures while being conditioned on coevolutionary data and ensuring equivariance, as detailed in the sections below. Related

Table 2. Comparative analysis of flow-based protein conformation generative models. Columns include atomistic granularity (Gran.), MSA conditioning (MSA), equivariance (Equiv.), evaluation metrics, and the dataset used.

Paper	Input	Output	Gran.	MSA	Equiv.	Metric	Dataset
Mahmoud et al. [94]	Coarse-grained protein structures	Full atom protein conformations	All-atom	No	Yes	Dihedral distributions, TICA	MD of Bromodomain and Chignolin
Jing et al. [56]	Protein sequence	Protein backbone conformation	Backbone	Yes	No	lDDT-C$\alpha$	PDB [95], ATLAS [96]
Li et al. [97]	Protein sequence + MSA	Protein backbone conformation	All-atom	No	No	RMSD, RMSF, DCCM correlation	ATLAS [96]
Wolf et al. [98]	Equilibrium backbone structure	Protein backbone conformation	Backbone	No	Yes	RMSD, RMSF, PCA W2	ATLAS [96]
Jin et al. [55]	Protein sequence + Approx. energy	Protein backbone conformation	Backbone	No	Yes	RMSF, RMWD, Rg, PWD	ATLAS [96]

provides a comparative analysis of prominent flow-based deep learning models used for protein conformation sampling.

4.3.1. Alphaflow and Its Variants

Alphaflow and ESMFlow [56] are generative models that fine-tune AlphaFold 2 [1] and ESMFold [37] under a flow-matching [54] framework to sample diverse and accurate protein conformational ensembles from sequence input. The authors modify AlphaFold and ESMFold into denoising models within a custom flow matching framework trained on PDB structures and further fine-tuned on an all-atom MD ensemble dataset ATLAS [96], to model conformational variability. Compared to MSA subsampling, AlphaFLOW and ESMFLOW produce more diverse and precise structures on PDB test sets. Li et al. [97], who introduced Alphaflow-lit, further improved this framework by freezing the AlphaFold 2 [1] embedder and evoformer modules during the training steps, achieving a speedup of 47 times and still performing on par with Alphaflow [56] in terms of attributes such as protein dynamics and local arrangement.

4.3.2. Integration of Equivariance and Full Atom Modeling

Although AlphaFlow and its variants achieve state-of-the-art performance in protein ensemble generation, they do not account for rotational invariance or equivariance. Jin et al. [55] address this limitation with P2DFlow, which innovatively incorporates both the protein sequence and an approximate energy as inputs. This approximate energy is derived by projecting molecular dynamics (MD) simulation results onto a low-dimensional manifold. P2DFlow then leverages ESM-2 [99] to obtain sequence embeddings and employs ESMFold with structural perturbations to produce a noisy structure as a prior. A SE(3)-equivariant architecture integrating the AlphaFold IPA (Invariant Point Attention) module with the E(n)-equivariant EGNN [100] (Equivariant Graph Neural Network) model predicts the vector field for the flow process.

This approach is shown to outperform existing generative models such as AlphaFlow [56] and STR2STR [58] across multiple benchmarks, with the caveat that we obtain generated backbone conformations only, not full atom proteins.

Another flow-based framework focused solely on the generation of backbone conformation was introduced by Wolf et al. [98], as BBFlow. A given “equilibrium” backbone atom structure is used to condition an SE(3) flow matching model for the generation of a set of protein backbone conformations. One important aspect of this work is that the conditioning of the generation is not on the sequence info, such as MSA, but on the equilibrium structure; the authors eliminate the need for evolutionary information and a pre-trained folding model weights (e.g., like Alphaflow [56]). Thus, the work is analogous to a molecular dynamics simulation where an initial protein backbone structure is required. Subsequent benchmarking shows that BBFlow performed better than AlphaFlow(with no templates) for de novo proteins, yielding lower median RMSF and pairwise RMSD values.

Expanding beyond modeling backbone-only atoms, a recent contribution by Mahmoud et al. [94] presents a deep learning framework that combines coarse-grained SIRAH [101] protein representations with normalizing flows. This framework captures multimodal conformational distributions and generates low-energy full-atom protein ensembles that align closely with molecular dynamics (MD) simulations.

4.3.3. Discussion

Arguably, the inclusion of equivariance into the generative model itself is a matter of debate, as AlphaFold 3 [36] omits computationally expensive [102] equivariant layers in its structure generation module. Instead, the model relies on extensive data augmentation to learn approximate equivariance behavior from the training data, a departure from the approach adopted by AlphaFold2 [1], highlighting an alternative strategy for incorporating symmetry awareness in deep protein structure prediction models [103,104,105].

Flow-based generative models present a promising alternative to traditional learning approaches for modeling protein conformations, as they offer exact likelihood estimation and stable training behavior. Despite these advantages, current models are limited in their ability to generate full-atom protein structures that include both backbone and side chains. The studies examined in this context define the generative process on coordinate frames constructed using only backbone atoms, such as $C\alpha$, N, $C\beta$, and O. This constraint largely stems from the models’ focus on improving upon AlphaFlow [56], which itself produces denoised $C\beta$ structures rather than full atom models.

In addition, there are currently no flow-based generative models that explicitly model conformational ensembles of intrinsically disordered proteins, suggesting a valuable opportunity for future research.

4.4. Diffusion

Diffusion-based generative models have rapidly emerged as a powerful paradigm in various areas of protein engineering, including protein design [106,107,108], protein-ligand docking [109,110,111], and protein structure generation [112,113,114]. Their core strength lies in modeling the conformational landscapes of proteins and peptides by learning to reverse a progressive noise corruption process. This iterative denoising procedure enables the generation of diverse, accurate, and physically plausible molecular structures.

Table 3. A comparison of diffusion models for protein conformation sampling showing an extensive enforcement of equivariance into the generative model while ignoring the use of evolutionary MSA data.

Paper	Input	Output	Gran.	MSA	Equiv.	Metric	Dataset
Fan et al. [115]	Protein sequence and MSA	3D protein conformation	All-atom	Yes	Yes	MAT-R, MAT-P, TM-score	All PDB structures before 30 April 2022
Liu et al. [116]	Protein sequence	MD trajectory	All-atom	No	Yes	RMSD, RMSF, MAE	Dynamic PDB
Liu et al. [117]	Protein sequence and Experimental data	Protein conformational ensemble	Backbone	No	Yes	Rg, SASA	Ground truth generated via MD
Lu et al. [118]	Protein sequence	Protein backbone conformations	Backbone	No	Yes	JS-PwD, JS-TIC, JS-Rg	PDB
Wang et al. [59]	Protein sequence	Backbone conformation ensemble	Backbone	No	Yes	RMSD, RMSF, Val-C$\alpha$	PDB
Lu et al. [58]	A starting structure of a target Protein sequence	All-atom protein conformations	All-atom	No	Yes	Val-Clash, RMSD, TICA, Rg	PDB
Zheng et al. [119]	CG protein representation	Protein conformation	All-atom	No	Yes	RMSD, TICA	MD data and force field
Lewis et al. [57]	Protein Sequence	Protein backbone conformations	Backbone	Yes	Yes	RMSD, Free-energy agreement	PDB+ MD+ NMR data

presents a comparative overview of leading diffusion models in protein conformational sampling.

4.4.1. Perturb and Anneal Based Diffusion

Simulated annealing emulates high temperature (early steps), allowing larger jumps, while at low temperature (later steps), it fine-tunes toward high-probability regions. In terms of protein conformation sampling, early noisy samples can explore global conformational space, while later steps focus on fine adjustments.

Str2Str [58], shows that effective conformational sampling can be achieved using only static crystal structures in a zero-shot manner, without MD-based training. It employs an SE(3)-equivariant diffusion framework, using a perturb-and-anneal approach for the forward and reverse stages, respectively. The generative process is guided by score-based annealing, drawing inspiration from the equivariant Invariant Point Attention (IPA) module of AlphaFold2 [1]. Notably, this approach remains agnostic to the underlying energy landscape, allowing flexible modeling without relying on predefined potential functions. Later, full atom conformations are sampled using faspr [120] protein side chain packing tool.

DiG, a subsequent work by Zheng et al. [119] leverages a perturb and anneal diffusion principle to obtain the equilibrium distribution of the protein conditioned on a general descriptor. In a novel contribution, the diffusion process is further supervised by an energy function such as DFT (Density Functional Theory) [121] beyond a standard data-based training. Unlike str2str, the model produces a coarse-grained representation of the protein molecule, comprising individual residue orientations and $C\alpha$ atom coordinates.

BioEmu [57] extends the perturb-and-anneal diffusion framework by training on a diverse dataset that includes static protein structures from the Protein Data Bank (PDB) [95], molecular dynamics (MD) trajectories, and experimental data to learn coarse-grained representations, similar to DiG. Like AlphaFold2, it incorporates sequence information to derive single and pairwise representations, which are then used as inputs to the diffusion model. The resulting model effectively captures equilibrium distributions and demonstrates conformational diversity, such as the open and closed states of the ADK protein. In addition, it supports the prediction of protein stability metrics, compared to the MEGAScale dataset derived from Tsuboyama et al. [122] and outperforms AlphaFlow [56] in conformation generation tasks.

4.4.2. Other Approaches

DeepConformer [123] builds on the protein design model RFDiffusion [124]. The authors argue that protein energy landscape and conformational dynamics can be learned from experimental structures in PDB and coevolution data, rather than the oft-used simulation data. The resulting model demonstrated dynamic properties (measures in terms of ResID vs RMSF) similar to those of MD-sampled structures using KaiB fold switching mechanism and catalytic conformation changes in the Adenylate Kinase protein.

Maddipatla et al. [125] proposed a diffusion-based framework trained on experimental data like NMR (Nuclear Magnetic Resonance) [126] instead, which generates conformational ensembles by treating models such as AlphaFold3 [36] as structural priors and performing posterior inference conditioned on experimental data, and shown to violate fewer constraints when modeling NMR ensembles than AlphaFold3 [36].

In a related method that also leverages training on experimental data, ExEnDiff [117] builds on the pre-trained protein ensemble sampler Str2str [58], generating protein conformation ensembles that more closely approximate the true Boltzmann distribution than those produced by Str2str alone. The advantage of this approach is that the model can have proportionally improved performance with respect to the underlying protein ensemble sampler.

A persistent challenge in protein conformation sampling lies in balancing the high computational cost of collecting molecular dynamics (MD) data informed by the underlying force field with the scalability of energy-agnostic generative models. Str2str-FT and Str2str-NE, proposed by Lu et al. [118], address this by introducing a few-shot learning framework that fuses neural network based sampling with energy landscape awareness.

Their two-step strategy first generates a diverse set of plausible conformations using a neural network, which are further complemented through short MD simulations. This hybrid approach demonstrates improved performance in terms of Jensen-Shannon (JS) divergence and conformation validity, outperforming methods such as idpGAN, the original Str2str, and MSA subsampling [39,46,58].

As a complementary approach to modeling the protein energy landscape, ConfDiff, a diffusion-based method to generate protein conformations, was introduced by Wang et al. (2024) [59]. The model is guided by the input sequence using classifier-free guidance operating on the SE(3) group.

Additionally, it incorporates physics-based rewards derived from energy and force information, helping the model generate more realistic structures. Unlike models like Dig [119], this approach avoids reliance on computationally expensive MD training data and introduces intermediate force supervision to balance quality and diversity compared to prior work like Str2str [58], leading to diverse samples more true to the underlying Boltzmann distribution.

4.4.3. Modeling Instrinsically Disordered Proteins

As discussed in the Section 4.3, IDPs pose a unique challenge as they exist as an ensemble rather than a static structure. Inspired by Str2str [58], IDP-Fold [127] captures the conformational diversity of IDPs using score-based diffusion. IDP-Fold achieves more detailed structural features when compared to idpGAN [46]. Additionally, this approach bypasses the need for Multiple Sequence Alignments (MSAs) or experimental data, supporting the idea that IDPs do not require or benefit much from MSAs [128,129].

Taneja et al. [130] employ a 2-stage approach to model the conformational ensemble of IDPs. First, the pairwise distance vector is generated given the sequence s and the inverted charge mutant $s^{\prime }$. This is followed by pre-trained DDPMs (Denoising Diffusion Probabilistic Models) [131] to map the vector to $C\alpha$ coordinates.

Improving upon idpGAN [46], Janson et al. proposed idpSAM [132], a latent diffusion model combining an SE(3)-invariant autoencoder and a conditional DDPM [131], enabling transferable and generalizable ensemble generation from limited simulation data. In 2 stages of training, a Euclidean-invariant AE is trained to encode protein conformations represented by $C\alpha$ coordinates, followed by a DDPM [131] to decode $C\alpha$ coordinates for subsequent IDP conformations. The resulting model improves upon idpGAN [46] on metrics such as KL divergence and RMSE.

4.4.4. Discussion

Diffusion models employ a fundamentally different approach from auto-encoder and flow-based methodologies. Rather than relying on indirect representations, these models utilize score functions to characterize the energy landscape governing protein folding dynamics directly. This energy-informed framework provides substantial advantages for conformation sampling by incorporating thermodynamically relevant principles. The key strength lies in their ability to generate complete, full-atom protein structures. While typically requiring integration with specialized side-chain optimization algorithms, diffusion models enable a direct sequence-to-conformation modeling pipeline, offering comprehensive structural prediction with both completeness and physical relevance.

5. Dataset

The success of deep generative models for protein conformation sampling depends critically on the quality and diversity of training datasets. This section reviews the primary datasets used in the literature, ranging from experimental structures to molecular dynamics simulations. Each dataset presents unique characteristics in terms of size, structural diversity, and temporal resolution that influence model performance and applicability.

5.1. Protein Data Bank

The Protein Data Bank (PDB) derived from [95] serves as a comprehensive dataset for training diffusion models with experimentally determined protein structures, primarily obtained via X-ray crystallography, nuclear magnetic resonance (NMR), and cryo-electron microscopy (cryo-EM). Computational models frequently utilize curated subsets of these structures, filtered based on experimental resolution quality, sequence length criteria to ensure computational feasibility, and completeness of structural information (e.g., resolution better than 5 Å, sequence length typically between 10 and 512 residues). To prevent data leakage and improve model generalization, training protocols often remove structures sharing high sequence similarity between training, validation, and testing datasets. Additionally, clustering strategies based on sequence similarity (e.g., 30% sequence identity) may be applied to manage redundancy and maintain data diversity [58,59,119].

5.2. AlphaFold Protein Structure Database

The AlphaFold Protein Structure Database (AlphaFold DB) is an extensive repository of protein structures predicted by the AlphaFold 2 artificial intelligence system developed by Google DeepMind [7]. In preparing dataset for computational modeling applications, the AlphaFold DB typically undergoes systematic preprocessing steps, such as sequence clustering based on identity thresholds (e.g., approximately 80% sequence identity), subsequent structural clustering to capture diverse conformational states, and stringent filtering criteria based on structural confidence metrics (e.g., pLDDT scores). This rigorous curation ensures the dataset maintains significant structural variability, thereby providing a valuable resource for training advanced generative models designed to accurately represent protein conformational diversity [57].

5.3. Atlas of Protein Molecular Dynamics

The ATLAS dataset is a large-scale resource comprising all-atom, explicit solvent molecular dynamics (MD) simulations of 1390 structurally diverse, non-membrane proteins [96]. These proteins were selected to represent the full range of Evolutionary Classification of protein Domains (ECOD) structural classes. For each protein, the dataset provides three independent 100 ns MD trajectories, sampled at high temporal resolution (e.g., 10,001 frames per trajectory). In typical data preprocessing pipelines, the trajectories are subsampled at regular intervals (e.g., every 100 frames) to obtain manageable subsets for training, such as 300 frames per protein. Representative conformations can also be selected based on structural or energetic criteria, including approximate energy estimates [55,56,57,97,98].

5.4. Bovine Pancreatic Trypsin Inhibitor

The Bovine Pancreatic Trypsin Inhibitor (BPTI) [133] is a widely used benchmark dataset for evaluating generative models of protein conformation due to its compact size and well-characterized dynamic properties. It consists of 58 residues and exhibits transitions among multiple conformational states. Molecular dynamics simulations of BPTI are commonly sampled at high temporal resolution (e.g., 250 ps intervals), and analysis typically focuses on C$\alpha$ atom coordinates. These settings facilitate the evaluation of a model’s capacity to reproduce conformational diversity and dynamics [59,78].

5.5. MEGAScale

The MEGAscale dataset [122] is a large-scale experimental resource for protein stability, comprising approximately one million measurements of folding free energies obtained via cDNA display proteolysis. A subset of this dataset, including over 750,000 experimental measurements, has been used to train generative models for protein property prediction [57]. As the dataset does not include structural information, complementary molecular dynamics (MD) simulations spanning 25 ms across 271 wild-type proteins and over 22,000 mutants have been employed to model folding-unfolding dynamics. These simulations enable the alignment of generated conformational ensembles with experimentally determined stability landscapes.

5.6. CATH

The CATH (Class-Architecture-Topology-Homologous superfamily) dataset derived from [134] provides a hierarchical classification of protein domains based on their structural and evolutionary relationships, supporting large-scale annotations of protein function and structure. The latest release (v4.3) significantly expands both sequence and structural coverage, with over 500,000 fully classified structural domains and 151 million predicted sequence domains assigned to 5481 superfamilies. In generative modeling applications, curated subsets of CATH domains—typically consisting of 50–200 residues—are used to capture functionally and structurally diverse protein fragments [57]. These domains are simulated using adaptive sampling protocols and high-resolution molecular dynamics, resulting in datasets with cumulative simulation times exceeding 40 ms across more than 1000 systems. Selected test domains reach trajectory lengths of up to 100 μs each, enabling accurate modeling of long-timescale conformational dynamics.

6. Metrics

Please refer to

Table 4. Comprehensive evaluation metrics for assessing protein conformation generation models across five key dimensions.

Metric Category	Score Name	Definition and References
Validity	Val-Clash	Validation metric for bond lengths/angles; [118]
	Steric-clash	Fraction of structures free from Cα clashes [58]
	Ramachandran Plot Score	Measures fraction of $(\varphi ,\psi )$ backbone angles within allowed regions [55,116]
	Contact-map Frequency	Frequency of residue pairs maintained within contact threshold across conformations [116]
	Sanity-check Pass Rate	Fraction of conformations passing comprehensive validation (steric, bonding, and dihedral angle checks) [55]
Precision & Accuracy	TM-score	Template-modeling score (0–1 scale) quantifying global fold similarity and structural alignment quality [115,119]
	RMSD	Root-mean-square deviation measuring atomic position accuracy relative to reference structures [135,136]
	Pearson Correlation	Correlation coefficient between predicted and reference per-residue measurements [56,137]
Diversity	MAT-P/MAT-R	Matching precision/recall: average RMSD to nearest reference/generated structure [115]
	COV-P/COV-R	Coverage precision/recall: percentage of structures covered within distance threshold [138]
	lDDT-Cα	Average local structural dissimilarity between pairs of sampled conformations [56]
Distributional Similarity	JS Divergence (PwD, Rg, TIC)	Jensen-Shannon divergence measuring similarity between generated and MD reference distributions [97,137]
	JS-PwD	Fidelity score derived from Jensen-Shannon divergence of pairwise distance distributions [137]
	Sampler Score	Composite metric combining global (JS divergence) and local (RMSF, secondary structure) similarity [117]
	RMWD	Root-mean Wasserstein distance across atom-wise positional probability distributions [56]
	Coverage/k-recall	Fraction of MD reference structures covered and average distance to k-nearest generated samples [59]
Structural Dynamics	RMSF	Root-mean-square fluctuation quantifying per-residue mobility across conformational ensembles [97,98,116]
	Radius of Gyration (Rg)	Time-dependent measure of overall protein compactness and conformational changes [116]
	Weak/Transient Contacts	Frequency of non-covalent contact formation relative to reference structure [55]

. Details of the individual metrics can be found in the Appendix A Section.

7. Future Work

7.1. Challenges

7.1.1. Lack of Interpretability

The lack of interpretability in deep learning models for protein conformation sampling is a real challenge because it makes it hard to understand how or why the model generates certain structures. These models might produce realistic-looking conformations, but we often don’t know what parts of the input, e.g., sequence features or structural patterns, are influencing the output. That means we cannot easily tell if the model is learning meaningful biology or just picking up on patterns in the data. This makes it harder to trust the results, especially when there’s little experimental data to compare with, making the model a black box.

7.1.2. Simulation of Full-Atom Protein Trajectory

Sampling protein conformations, and especially full structural trajectories, presents a unique challenge because it involves capturing how atomic coordinates change over time. This makes it a complex, multidimensional problem. Although some models have been proposed as data-driven alternatives to molecular dynamics simulations [135,139], accurately tracking the progression of structural changes, including both backbone and side chain atoms, remains a significant and open challenge.

7.1.3. Integration of Other Biomolecules

As noted by Lewis et al. [57], deep generative models driven by data can complement insights from molecular dynamics (MD) simulations by efficiently exploring the conformational landscape of proteins. However, integrating other biomolecules, such as DNA, RNA, and ions, into these models remains a significant challenge to account for the structural changes upon binding. With the advent of AlphaFold 3 [36], which enables static modeling of such complexes, there is now a potential path forward to capture the conformational changes induced by protein interactions with these cellular partners.

7.2. Opportunities (See Box 1)

7.2.1. Aligning with Physical Feedback

Physical alignment methods specifically designed for generative models, such as Lu et al. [140], demonstrate a promising path forward. The conformation sampling models discussed in this survey lack concrete mechanisms for aligning the generative process through explicit energy calculations of individual protein structures. This novel approach, termed Energy Based Alignment (EBA), can be employed to align pre-trained generative models toward specific structural classes without requiring computationally expensive retraining procedures.

Box 1. What is Missing?

• Intrinsically Disordered Protein specific Flow models.
• Harmonized benchmarks across datasets.
• Interpretable Conformational Sampling focused Deep Learning models [141].
• Integration of DNA, RNA, and ions as a part of large scale conformational landscape models aided by Alphafold 3 [36].

7.2.2. Textual Representation

Schwing et al. [142] explore the textual representation of protein conformations from MD trajectories using vectors of the Ramachandran basin, as opposed to all-atom or coarse-grained representations. This abstraction suggests a future direction in which text-to-conformation models could emerge, combining advances in the fields of natural language processing and protein structure modeling.

7.2.3. Structural Language Models for Conformational Plasticity

Recently, protein language models have gained significant attention for a variety of applications, including the prediction of protein structure, functional annotation, and prediction of intrinsically disordered regions (IDR). In line with these developments, Lu et al. [143] introduce Structural Language Models (SLMs) aimed at modeling protein structural plasticity, capturing the inherent flexibility of proteins under different functional and environmental conditions. This approach could inform the development of future protein language models that translate sequences into structures to act as a priori for full-fledged protein conformation modeling.

7.2.4. Full-Atom End-to-End Structure Generation

One of the central challenges in protein conformation modeling is the end-to-end generation of full-atom protein structures, including both backbone and side chain atoms. Alphafolding [135] addresses this in a limited setting by incorporating 4D dynamics prediction up to a 256 amino acid long protein chain to learn from molecular dynamics data. For each residue, it incorporates the side chain torsion angle loss represented as points on the unit circle is inspired by AlphaFold [1].

7.2.5. Generative Modeling of Peptide Dynamics

Jing et al. [139] propose a generative model designed to replicate the molecular dynamics of peptides. This opens promising opportunities for scaling generative models toward full-scale emulation of molecular dynamics for protein systems.

8. Conclusions

Deep generative models are reshaping protein structure modeling by moving beyond static representations to capture the dynamic conformational ensembles that underlie biological function. Unlike traditional approaches that often focus on a single best structure, these models aim to represent the full landscape of protein conformations, offering a richer view of structural variability and dynamics. In this review, we highlight recent advances in applying GANs, VAEs, normalizing flows, and diffusion models to protein conformation sampling, showcasing how they provide scalable and efficient alternatives to physics-based simulations such as molecular dynamics.

Each class of generative model brings unique strengths to the table. GANs offer fast generation, but face challenges with stability and mode collapse. VAEs provide interpretable and continuous latent spaces, which are valuable for adaptive sampling and ensemble learning. Flow-based models combine efficiency with tractable likelihood estimation and are naturally compatible with physical constraints. Diffusion models, meanwhile, have emerged as a particularly powerful paradigm due to their flexibility, equivariance support, and capacity to model complex, multimodal distributions, including those relevant for intrinsically disordered proteins and coarse-grained simulations.