Yongna Yuan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Yongna Yuan has a total of 79 co-authors. Here are the top co-authors by number of shared publications:
- Ruisheng Zhang
Lanzhou University
14 shared publications Ruisheng Zhang
School of Information Science & Engineering, Lanzhou University, South Tianshui Road, Lanzhou, Gansu, 730000, China
8 shared publicationsWei Su
Lanzhou University
8 shared publications- Gaili Li
School of Mathematics and Computer Science, Northwest Minzu University, Chengguan, Lanzhou, China
6 shared publications - Wei Su
Lanzhou University
6 shared publications Zhenyu Liu
Lanzhou University
5 shared publications
Latest Publications
- 21 September 2026
- Published by American Chemical Society (ACS) in The Journal of Physical Chemistry B
- Vol. 130 (39) , 10222-10232
Capturing Spatial Covariance for Efficient Multipolar Electrostatic Parameterization in DNA via Multioutput Learning
In DNA force field development, introducing high-rank atomic multipole moments (AMMs) to describe electron cloud anisotropy improves the accuracy of electrostatic energy calculations. However, high-rank AMMs have many components. Existing machine learning methods usually model each component independently, which increases model size and computational cost. This approach ignores the correlations among components and cannot balance accuracy and efficiency. In this study, a DNA molecular electrostatic energy data set was constructed for six base pairing types to obtain AMMs labels up to the hexadecapole moment. We evaluated two task grouping strategies alongside three multioutput Gaussian process regression (MOGPR) architectures (the Linear Model of Coregionalization (LMC), Intrinsic Coregionalization Model (ICM), and Semiparametric Latent Factor Model (SLFM)). Results showed that the Component strategy, which groups components by rank, reduced the total number of models by 80% for each data set. This grouping lowered the computational and maintenance costs of force field construction while reducing interference between components of different ranks. Among the MOGPR architectures, SLFM balanced accuracy and efficiency. It described the spatial covariance relationships among tasks through a low-rank structure, overcoming the limitations of independent prediction. In molecular electrostatic energy reconstruction, SLFM showed an even greater advantage. On the CG data set, which has high conformational dispersion, the mean absolute error of SLFM decreased by 46.5% compared with the single-task method. This study combined quantum chemical topology and multitask learning, providing a reference for predicting high-rank AMMs and developing DNA force fields.
- 21 September 2026
- Published by American Chemical Society (ACS) in The Journal of Physical Chemistry B
- Vol. 130 (39) , 10222-10232
Capturing Spatial Covariance for Efficient Multipolar Electrostatic Parameterization in DNA via Multioutput Learning
In DNA force field development, introducing high-rank atomic multipole moments (AMMs) to describe electron cloud anisotropy improves the accuracy of electrostatic energy calculations. However, high-rank AMMs have many components. Existing machine learning methods usually model each component independently, which increases model size and computational cost. This approach ignores the correlations among components and cannot balance accuracy and efficiency. In this study, a DNA molecular electrostatic energy data set was constructed for six base pairing types to obtain AMMs labels up to the hexadecapole moment. We evaluated two task grouping strategies alongside three multioutput Gaussian process regression (MOGPR) architectures (the Linear Model of Coregionalization (LMC), Intrinsic Coregionalization Model (ICM), and Semiparametric Latent Factor Model (SLFM)). Results showed that the Component strategy, which groups components by rank, reduced the total number of models by 80% for each data set. This grouping lowered the computational and maintenance costs of force field construction while reducing interference between components of different ranks. Among the MOGPR architectures, SLFM balanced accuracy and efficiency. It described the spatial covariance relationships among tasks through a low-rank structure, overcoming the limitations of independent prediction. In molecular electrostatic energy reconstruction, SLFM showed an even greater advantage. On the CG data set, which has high conformational dispersion, the mean absolute error of SLFM decreased by 46.5% compared with the single-task method. This study combined quantum chemical topology and multitask learning, providing a reference for predicting high-rank AMMs and developing DNA force fields.
- 12 August 2026
- Published by Springer Nature
- p. 68-76
Q-Matrix Learning Based on Graph Attention Neural Network for Knowledge Tracing
- 12 August 2026
- Published by Springer Nature
- p. 68-76
Q-Matrix Learning Based on Graph Attention Neural Network for Knowledge Tracing
- 25 March 2026
- Published by Royal Society of Chemistry (RSC) in Physical Chemistry Chemical Physics
- Vol. 28 (14) , 8889-8901
TransG4: an interpretable deep-learning approach for sequence-based G-quadruplex prediction
G4 formation propensity is predicted using CNN, transformer and BiGRU models on G4-seq and rG4-seq data. DNA mismatch and RNA RSR scores are estimated, and attention-derived motifs reveal sequence features linked to G4 structures.
- 25 March 2026
- Published by Royal Society of Chemistry (RSC) in Physical Chemistry Chemical Physics
- Vol. 28 (14) , 8889-8901
TransG4: an interpretable deep-learning approach for sequence-based G-quadruplex prediction
G4 formation propensity is predicted using CNN, transformer and BiGRU models on G4-seq and rG4-seq data. DNA mismatch and RNA RSR scores are estimated, and attention-derived motifs reveal sequence features linked to G4 structures.
- 13 March 2026
- Published by Royal Society of Chemistry (RSC) in Physical Chemistry Chemical Physics
- Vol. 28 (13) , 8115-8127
InChINet: a self-supervised molecular representation learning framework leveraging SMILES and InChI
Molecular representation, as one of the fundamental challenges in artificial intelligence-driven drug discovery, has attracted increasing attention due to its low cost and impressive speed while it is applied in molecular property prediction, drug molecule generation, drug–drug interactions, etc. Numerous models that integrate multi-modal representations have been proposed for molecular representation learning. However, existing methods have not yet considered the IUPAC International Chemical Identifier (InChI) as one of the multi-modal inputs. To address this issue, we propose InChINet, a self-supervised molecular representation learning framework that is pre-trained on 10 million unlabeled molecules. It leverages mutual information across the simplified molecular line input system (SMILES) and InChI. In addition, we present token reordering and token masking for SMILES. Combined with SMILES enumeration, these three strategies introduce domain knowledge and improve the model's stability against syntactic variations in SMILES representations. Benefiting from the introduction of InChI and augmentation strategies, InChINet achieves impressive performance on a wide range of downstream tasks, including molecular property prediction, drug–drug interaction (DDI) prediction, clustering analysis, zero-shot cross-lingual retrieval, and ablation study.
- 13 March 2026
- Published by Royal Society of Chemistry (RSC) in Physical Chemistry Chemical Physics
- Vol. 28 (13) , 8115-8127
InChINet: a self-supervised molecular representation learning framework leveraging SMILES and InChI
Molecular representation, as one of the fundamental challenges in artificial intelligence-driven drug discovery, has attracted increasing attention due to its low cost and impressive speed while it is applied in molecular property prediction, drug molecule generation, drug–drug interactions, etc. Numerous models that integrate multi-modal representations have been proposed for molecular representation learning. However, existing methods have not yet considered the IUPAC International Chemical Identifier (InChI) as one of the multi-modal inputs. To address this issue, we propose InChINet, a self-supervised molecular representation learning framework that is pre-trained on 10 million unlabeled molecules. It leverages mutual information across the simplified molecular line input system (SMILES) and InChI. In addition, we present token reordering and token masking for SMILES. Combined with SMILES enumeration, these three strategies introduce domain knowledge and improve the model's stability against syntactic variations in SMILES representations. Benefiting from the introduction of InChI and augmentation strategies, InChINet achieves impressive performance on a wide range of downstream tasks, including molecular property prediction, drug–drug interaction (DDI) prediction, clustering analysis, zero-shot cross-lingual retrieval, and ablation study.
- 23 December 2025
- Published by Royal Society of Chemistry (RSC) in Physical Chemistry Chemical Physics
- Vol. 28 (2) , 1747-1758
LMProtein: a protein language model based framework for protein structural property prediction
Recent advances in machine learning and self-supervised deep language modeling have made it possible to accurately predict protein structural properties. Most existing models and pretraining methods leverage evolutionary information in multiple sequence alignments (MSAs) to obtain very promising results in protein structural property prediction. However, methods which make use of MSAs are computationally intensive and time consuming, and these methods cannot be applied to proteins which lack sequence homologs. Here, we present LMProtein, a fast and accurate framework for predicting protein structural properties such as protein secondary structure, backbone dihedral angles, fluorescence landscape and stability landscape only using protein primary sequence. By combining the unsupervised pretrained language model ESM-2 with a convolutional neural network, long short-term memory neural networks and multilayer perceptron, LMProtein achieves better performances than recent MSA-based models and single-sequence-based models. The accuracy of the eight-state secondary structures (SS8) prediction is approximately 74%, the mean absolute error of dihedral angle prediction is 19° and 29° for Phi and Psi, respectively, and Spearman's ρ between the experimental and predicted values of fluorescence and stability is 0.69 and 0.79, respectively. We believe that our framework has broad potential for predicting protein structural characteristics, providing important opportunities to accelerate the progress of protein engineering and drug target identification.
- 23 December 2025
- Published by Royal Society of Chemistry (RSC) in Physical Chemistry Chemical Physics
- Vol. 28 (2) , 1747-1758
LMProtein: a protein language model based framework for protein structural property prediction
Recent advances in machine learning and self-supervised deep language modeling have made it possible to accurately predict protein structural properties. Most existing models and pretraining methods leverage evolutionary information in multiple sequence alignments (MSAs) to obtain very promising results in protein structural property prediction. However, methods which make use of MSAs are computationally intensive and time consuming, and these methods cannot be applied to proteins which lack sequence homologs. Here, we present LMProtein, a fast and accurate framework for predicting protein structural properties such as protein secondary structure, backbone dihedral angles, fluorescence landscape and stability landscape only using protein primary sequence. By combining the unsupervised pretrained language model ESM-2 with a convolutional neural network, long short-term memory neural networks and multilayer perceptron, LMProtein achieves better performances than recent MSA-based models and single-sequence-based models. The accuracy of the eight-state secondary structures (SS8) prediction is approximately 74%, the mean absolute error of dihedral angle prediction is 19° and 29° for Phi and Psi, respectively, and Spearman's ρ between the experimental and predicted values of fluorescence and stability is 0.69 and 0.79, respectively. We believe that our framework has broad potential for predicting protein structural characteristics, providing important opportunities to accelerate the progress of protein engineering and drug target identification.
- 14 August 2025
- Published by Springer Nature in Interdisciplinary Sciences: Computational Life Sciences
- Vol. 18 (2) , 392-418
Relevance of 3D Rotationally Equivariant Neural Networks for Predicting Protein-Ligand Binding Affinities
Proteins are fundamental to biological processes, mediating critical functions through precise molecular interactions. The rotational dynamics between ligand atoms and protein binding sites can significantly influence interaction efficacy by modifying spatial relationships. In our research, we present the PLAe (three-dimensional (3D) rotationally equivariant neural networks for predicting protein-ligand binding affinities) methodology. This novel model synergizes radial basis functions with e3nn networks to encapsulate the radial and angular dimensions of molecular features. Radial basis functions effectively measure interatomic distances, while e3nn-an advanced neural network utilizing spherical harmonics-maintains invariance under rotational and translational transformations. The Clebsch-Gordan coefficients are employed to integrate angular and atomic properties seamlessly. By merging radial basis and spherical harmonic elements with Clebsch-Gordan representations, our approach adeptly captures molecular rotational symmetries and interatomic interactions. The inclusion of an attention mechanism further refines the affinity predictions, ensuring a high level of precision. This integrative and sophisticated model sets a new benchmark to accurately predict protein-ligand binding affinities, leveraging intricate molecular details to enhance predictive performance.
- 14 August 2025
- Published by Springer Nature in Interdisciplinary Sciences: Computational Life Sciences
- Vol. 18 (2) , 392-418
Relevance of 3D Rotationally Equivariant Neural Networks for Predicting Protein-Ligand Binding Affinities
Proteins are fundamental to biological processes, mediating critical functions through precise molecular interactions. The rotational dynamics between ligand atoms and protein binding sites can significantly influence interaction efficacy by modifying spatial relationships. In our research, we present the PLAe (three-dimensional (3D) rotationally equivariant neural networks for predicting protein-ligand binding affinities) methodology. This novel model synergizes radial basis functions with e3nn networks to encapsulate the radial and angular dimensions of molecular features. Radial basis functions effectively measure interatomic distances, while e3nn-an advanced neural network utilizing spherical harmonics-maintains invariance under rotational and translational transformations. The Clebsch-Gordan coefficients are employed to integrate angular and atomic properties seamlessly. By merging radial basis and spherical harmonic elements with Clebsch-Gordan representations, our approach adeptly captures molecular rotational symmetries and interatomic interactions. The inclusion of an attention mechanism further refines the affinity predictions, ensuring a high level of precision. This integrative and sophisticated model sets a new benchmark to accurately predict protein-ligand binding affinities, leveraging intricate molecular details to enhance predictive performance.
- 14 June 2025
- Published by Wiley in Journal of Computational Chemistry
- Vol. 46 (16) , e70147
Virtual Bonding Enhanced Graph Self‐Supervised Learning for Molecular Property Prediction
Accurate prediction of molecular properties is essential for modern drug design and discovery. Self‐supervised learning (SSL) and Graph Neural Networks (GNNs) have been widely used in this field to learn molecular representations and predict molecular properties. However, previous graph‐based deep learning methods have overlooked the important weak interaction, that is, long‐range interatomic interaction, which is crucial in determining the molecular properties. This study presents a novel self‐supervised learning framework, Virtual Bonding Enhanced Molecular Property Prediction (VIBE‐MPP), to address the limitations of existing methods by incorporating weak interactions and 3D spatial information into the molecular representations. VIBE‐MPP utilizes a Virtual Bonding Graph Neural Network (VBGNN) to construct a virtual bonding enhanced graph that encodes molecules, and a Dual‐level Self‐supervised Boosted Pretraining (DSBP) approach to enhance representation learning through four designed pretext tasks. The framework introduces virtual bonds to represent atom interactions within a radius of 10 Å, enabling an atom to engage in message passing with multiple other neighboring atoms simultaneously. The model is evaluated on 10 benchmark datasets, demonstrating superior performance over state‐of‐the‐art methods in both classification and regression tasks. On average, it improves upon the best baseline models by 3.20% and achieves optimal performance on four regression datasets. Additionally, visualizations of the learned molecular representations in downstream datasets show that VIBE‐MPP effectively captures molecular properties and semantic information.
- 14 June 2025
- Published by Wiley in Journal of Computational Chemistry
- Vol. 46 (16) , e70147
Virtual Bonding Enhanced Graph Self‐Supervised Learning for Molecular Property Prediction
Accurate prediction of molecular properties is essential for modern drug design and discovery. Self‐supervised learning (SSL) and Graph Neural Networks (GNNs) have been widely used in this field to learn molecular representations and predict molecular properties. However, previous graph‐based deep learning methods have overlooked the important weak interaction, that is, long‐range interatomic interaction, which is crucial in determining the molecular properties. This study presents a novel self‐supervised learning framework, Virtual Bonding Enhanced Molecular Property Prediction (VIBE‐MPP), to address the limitations of existing methods by incorporating weak interactions and 3D spatial information into the molecular representations. VIBE‐MPP utilizes a Virtual Bonding Graph Neural Network (VBGNN) to construct a virtual bonding enhanced graph that encodes molecules, and a Dual‐level Self‐supervised Boosted Pretraining (DSBP) approach to enhance representation learning through four designed pretext tasks. The framework introduces virtual bonds to represent atom interactions within a radius of 10 Å, enabling an atom to engage in message passing with multiple other neighboring atoms simultaneously. The model is evaluated on 10 benchmark datasets, demonstrating superior performance over state‐of‐the‐art methods in both classification and regression tasks. On average, it improves upon the best baseline models by 3.20% and achieves optimal performance on four regression datasets. Additionally, visualizations of the learned molecular representations in downstream datasets show that VIBE‐MPP effectively captures molecular properties and semantic information.
- Open Access
- 4 June 2025
- Published by Springer Nature in Journal of Cheminformatics
- Vol. 17 (1) , 1-14
A 3D generation framework using diffusion model and reinforcement learning to generate multi-target compounds with desired properties
Deep generative models provide a powerful solution for the de novo design of molecules. However, the majority of existing methods only generate molecules for a single target. Generating molecules with biological activities against multiple specific targets and desired properties remains an extremely difficult challenge. In this study, we propose a novel 3D molecule generation framework based on reinforcement learning and diffusion model to generate molecules with predefined properties for given multiple targets. The proposed framework, MDRL, uses a diffusion model to understand the 3D chemical structure of molecules and employs Kolmogorov-Arnold Networks instead of Multilayer Perceptron to enhance model performance. Through reinforcement learning, the framework is able to generate molecules that simultaneously target two targets and further optimizes multiple molecular properties. Experimental results show that our model exhibits comparable performance to various state-of-the-art molecular generation models, and MDRL can effectively navigate chemical space to design polypharmacological compounds and control multiple molecular properties. In multiple case studies, we verify that the generated molecules can simultaneously target two targets through molecular docking and assess the model’s ability to control multiple molecular properties. The results in this study highlight the advantages and practicalities of our model in generating polypharmacological compounds with desired properties. This study introduces MDRL, a 3D molecular generation framework integrating diffusion models and reinforcement learning for joint optimization of multi-target binding and molecular properties. MDRL shows improvements over existing methods in controlling drug-relevant properties and enhancing multi-target affinity. Experimental results demonstrate that MDRL efficiently generates drug-like compounds with robust polypharmacological profiles, offering a novel strategy for multi-target drug design.
- 4 June 2025
- Published by Springer Nature in Journal of Cheminformatics
- Vol. 17 (1) , 1-14
Open AccessA 3D generation framework using diffusion model and reinforcement learning to generate multi-target compounds with desired properties
Deep generative models provide a powerful solution for the de novo design of molecules. However, the majority of existing methods only generate molecules for a single target. Generating molecules with biological activities against multiple specific targets and desired properties remains an extremely difficult challenge. In this study, we propose a novel 3D molecule generation framework based on reinforcement learning and diffusion model to generate molecules with predefined properties for given multiple targets. The proposed framework, MDRL, uses a diffusion model to understand the 3D chemical structure of molecules and employs Kolmogorov-Arnold Networks instead of Multilayer Perceptron to enhance model performance. Through reinforcement learning, the framework is able to generate molecules that simultaneously target two targets and further optimizes multiple molecular properties. Experimental results show that our model exhibits comparable performance to various state-of-the-art molecular generation models, and MDRL can effectively navigate chemical space to design polypharmacological compounds and control multiple molecular properties. In multiple case studies, we verify that the generated molecules can simultaneously target two targets through molecular docking and assess the model’s ability to control multiple molecular properties. The results in this study highlight the advantages and practicalities of our model in generating polypharmacological compounds with desired properties. This study introduces MDRL, a 3D molecular generation framework integrating diffusion models and reinforcement learning for joint optimization of multi-target binding and molecular properties. MDRL shows improvements over existing methods in controlling drug-relevant properties and enhancing multi-target affinity. Experimental results demonstrate that MDRL efficiently generates drug-like compounds with robust polypharmacological profiles, offering a novel strategy for multi-target drug design.
- 4 February 2025
- Published by Springer Nature in Multimedia Systems
- Vol. 31 (2) , 1-22
A spatial-temporal graph attention network for protein–ligand binding affinity prediction based on molecular geometry
- 4 February 2025
- Published by Springer Nature in Multimedia Systems
- Vol. 31 (2) , 1-22
A spatial-temporal graph attention network for protein–ligand binding affinity prediction based on molecular geometry