Protein molecules are inherently dynamic and modulate their interactions with different molecular partners by accessing different tertiary structures under physiological conditions. Elucidating such structures remains challenging. Current momentum in deep learning and the powerful performance of generative adversarial networks (GANs) in complex domains, such as computer vision, inspires us to investigate GANs on their ability to generate physically-realistic protein tertiary structures. The analysis presented here shows that several GAN models fail to capture complex, distal structural patterns present in protein tertiary structures. The study additionally reveals that mechanisms touted as effective in stabilizing the training of a GAN model are not all effective, and that performance based on loss alone may be orthogonal to performance based on the quality of generated datasets. A novel contribution in this study is the demonstration that Wasserstein GAN strikes a good balance and manages to capture both local and distal patterns, thus presenting a first step towards more powerful deep generative models for exploring a possibly very diverse set of structures supporting diverse activities of a protein molecule in the cell.

The protein modeling community has long been interested in dimensionality reduction of structure data. Motivated by rapid progress in neural network research, we investigate autoencoders of various architectures on reducing the dimensionality of protein structure data generated by template-free protein structure prediction methods. We show that autoencoders that model nonlinear relationships among variables outperform linear dimensionality reduction. We evaluate various architectures and propose a better-performing one. We further show that the learned, low-dimensional latent representations capture inherent information useful for structure prediction. Given the ease with which open-source neural network libraries, such as Keras, which we employ here, allow constructing, training, and evaluating neural networks, we believe that autoencoders will gain in popularity in the structure biology community and open up further avenues of research.

Rapid growth in molecular structure data is renewing interest in featurizing structure. Featurizations that retain information on biological activity are particularly sought for protein molecules, where decades of research have shown that indeed structure encodes function. Research on featurization of protein structure is active, but here we assess the promise of autoencoders. Motivated by rapid progress in neural network research, we investigate and evaluate autoencoders on yielding linear and nonlinear featurizations of protein tertiary structures. An additional reason we focus on autoencoders as the engine to obtain featurizations is the versatility of their architectures and the ease with which changes to architecture yield linear versus nonlinear features. While open-source neural network libraries, such as Keras, which we employ here, greatly facilitate constructing, training, and evaluating autoencoder architectures and conducting model search, autoencoders have not yet gained popularity in the structure biology community. Here we demonstrate their utility in a practical context. Employing autoencoder-based featurizations, we address the classic problem of decoy selection in protein structure prediction. Utilizing off-the-shelf supervised learning methods, we demonstrate that the featurizations are indeed meaningful and allow detecting active tertiary structures, thus opening the way for further avenues of research.

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.

AlphaFold3 represents a major advance in protein structure prediction, yet its performance on intrinsically disordered proteins remains uncharacterized. We present the first systematic evaluation of AF3 on disordered systems, revealing a striking dichotomy. For monomers, AF3's pLDDT scores reliably predict disorder (MCC: 0.693), matching AlphaFold2 and rivaling dedicated predictors. This consistency across fundamentally different architectures confirms that disorder prediction emerges from training data, not model design. For multimers, the picture grows complex. Despite comparable aggregate performance (mean DockQ: 0.563 vs 0.571), AF3 and AF2 achieve these results through fundamentally different mechanisms. Conventional structural features explain 58% of AF2's variance but only 42% of AF3's. Users cannot predict when AF3 will succeed or fail from interface properties alone. On disorder-to-order transitions (MFIB benchmark), both models perform equally well, successfully predicting final folded states. Yet seed variance analysis reveals AF3's failures are deterministic: the model converges to identical structures across independent runs, whether correct or incorrect, indicating rigid structural priors override available information. Our findings establish AF3 as reliable for the prediction of monomer disorder but unpredictable for multimers. Architectural innovation alone cannot overcome training data bias. Progress demands disorder-enriched datasets and ensemble sampling, not merely novel architectures.

Extensive research on deep neural networks shows that complex deep learning models have considerably improved our ability to predict the native structure of a protein amino acid sequence. With the release of AlphaFold2, an entirely data-driven approach to machine learning, we can now predict the native tertiary structure of a given protein sequence with great precision. On the other hand, research into deep learning frameworks that can take into account protein structural plasticity is in its early stages. Obtaining a multi-structure view of a protein molecule remains an outstanding challenge in computational structural biology. In this paper, we make two key contributions. We first propose a novel end-to-end generative model framework, a new formulation under Equivariant Graph Neural Networks (EGNN) based encoding and Graph Variational Autoencoder (GVAE), advancing our ability to generate realistic tertiary structures. Most existing models rely on 2D convolution, with protein structures represented by contact maps or distance matrices. In contrast, our presented model learns over both 3D coordinates of protein structure and sequence directly. The second contribution of this paper is control of tertiary structure realism. We show that through the loss function, we can control properties of generated tertiary structures. We suggest different terms in the loss function and analyze how those terms allow us to recover realistic patterns, such as backbone, short-range, and long-range contacts that we find in tertiary structures. Additionally, we conduct a careful analysis along several metrics that measure the physical realism of generated tertiary structures and show that EGNN encoding-based GVAE are effective models for generating physically-realistic structures.