14 September 2024 in Materials Today
28 June 2023 in Food Chemistry
24 April 2025 in IEEE Transactions on Computational Biology and Bioinformatics
B-factor is a measure of ray attenuation or scattering caused by atomic thermal motion during X-ray diffraction of protein crystal structure. B-factor reflects the vibration of atoms; hence, it is the most common experimental descriptor of protein flexibility and has been extensively applied in the studies of protein dynamics, screening of bioactive small molecules, and protein engineering. The prediction of B-factor profiles has considerable significance for analyzing the dynamic properties of unknown proteins. Deep learning technology has developed rapidly in recent years and has been widely implemented in many research fields, especially structural biology. In this paper, a deep neural network model based on bidirectional long short-term memory (biLSTM) network is proposed to predict the B-factor profile of a protein by combining its sequence-based features and structure-based features. Based on a large dataset of high-resolution proteins, our method predicts the B-factor profiles with an average Pearson correlation coefficient (PCC) of 0.71, and 85% of the B-factor profiles have a PCC greater than 0.6, which indicates a strong correlation between predicted and experimental values. In addition, our method remarkably outperforms the existing methods on four test datasets with different protein sizes.
15 December 2025 in ChemRxiv
As a crucial parameter obtained through NMR spectroscopy for elucidating the dynamic behavior of proteins in solution, backbone N-H order parameters indicate the flexibility of atoms or groups on the picosecond to nanosecond timescale, thereby facilitating a comprehensive understanding of protein function and mechanisms. In NMR experiments, obtaining order parameters involves measuring relaxation parameters and steady-state NOE, followed by conducting Lipari-Szabo model-free analysis, which is a complex and challenging process. In this paper, we employ the self-training strategy, a semi-supervised learning (SSL) approach, to improve the capability of machine learning algorithms in predicting backbone N-H order parameters through iterative training. Throughout the iterative process, we gradually enrich the training set by selecting unlabeled data with high-quality pseudo labels, allowing the student model to progressively enhance its generalization performance under the guidance of the teacher model. The final student model is evaluated on a test set comprising 10 proteins, achieving an average Pearson correlation coefficient of 0.854 between the predicted and experimental order parameters, representing a 4.5% improvement over the original teacher model. It demonstrates that self-training effectively leverages unlabeled structural ensemble data to improve protein fast-dynamics prediction.
15 December 2025 in ChemRxiv
Protein dynamics are pivotal to biological function, and elucidating these dynamic properties is essential for understanding their behavior in cellular processes. Nuclear magnetic resonance (NMR) spectroscopy quantifies residue-specific motional freedoms through order parameters (S²), thereby providing critical insights into local structural flexibility and conformational dynamics. Nevertheless, accurate prediction of NMR order parameters remains a critical challenge in structural biology. Traditional approaches often rely on experimental structures or suffer from limited accuracy when using sequence-based methods. To address this challenge, we introduce SOPPCL, a novel deep learning framework that significantly improves sequence-based prediction of order parameters by integrating advanced protein sequence representations and contrastive learning. Our method leverages the ESM-2 protein language model to capture high-dimensional semantic features from amino acid sequences and employs HHblits-derived HMM profiles to encode evolutionary information. To further enhance feature discriminability, SOPPCL introduces a contrastive learning module that aligns and optimizes the fused representations of ESM-2 and HMM features by maximizing mutual information between positive pairs while minimizing similarity among negative pairs. Subsequently, the refined features are fed into a regression network to predict order parameter values. Evaluation on a benchmark dataset containing 10 proteins shows that SOPPCL achieves a Pearson correlation coefficient of 0.845 and a root mean square error of 0.132, surpassing sequence-only baselines such as DynaMine (PCC=0.464, RMSE=0.178) by 82% and 26% relative improvements, respectively. By eliminating reliance on structural data, SOPPCL provides a powerful computational approach for investigating protein dynamics directly from sequence information, with potential applications in NMR-assisted drug design and functional annotation. Altogether, our study highlights the synergistic benefits of integrating protein language models, evolutionary information, and contrastive learning for advancing the prediction of biomolecular properties.