Zefeng Zhu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Zefeng Zhu has a total of 43 co-authors. Here are the top co-authors by number of shared publications:
- Chen Song
Center for Quantitative Biology Academy for Advanced Interdisciplinary Studies Peking University Beijing China
9 shared publications - Minghui Li
School of Basic Medical Science, Suzhou Medical College of Soochow University, Suzhou 215123, China
6 shared publications - Yuting Chen
Soochow University
4 shared publications - Haoyu Lu
Soochow University
3 shared publications - Ning Zhang
Soochow University
3 shared publications - Shuqin Wang
Soochow University
3 shared publications
Latest Publications
- 16 November 2026
- Published by Springer Nature in Nature
Architecture of a portal complex embedded in the poxvirus core
- 16 November 2026
- Published by Springer Nature in Nature
Architecture of a portal complex embedded in the poxvirus core
- Open Access
- 6 November 2026
- Published by Wiley in Quantitative Biology
- Vol. 14 (4)
Predicting the alternative conformation of a known protein structure based on the distogram of AlphaFold2
AlphaFold2 (AF2) has become one of the most established and leading tools for predicting protein structures, although it often favors a single conformation. While it has been shown that sampling different conformations is possible by manipulating the multiple sequence alignment (MSA), the potential for multi‐state information within a standard AF2 prediction remains underexplored. In this work, we demonstrate that even with a deep MSA, AF2 predictions encode structural heterogeneity. Specifically, we observe that the predicted residue–residue distance distributions often exhibit multiple probability peaks before recycling, suggesting the presence of alternative states. We derive state‐specific distance restraints by fitting and separating these multi‐peak distributions. These restraints are then incorporated into Rosetta as energy terms to model large‐scale conformational transitions. Across multiple protein systems undergoing diverse types of conformational changes, our method effectively samples alternative conformations starting from an input conformation, producing models comparable to those from existing approaches. Our findings highlight a previously underutilized feature of AF2 predictions and offer a novel strategy to take advantage of these insights to study conformational flexibility and dynamics in proteins. This work broadens the applicability of AF2 in structural biology and the modeling of multi‐state protein systems.
- 6 November 2026
- Published by Wiley in Quantitative Biology
- Vol. 14 (4)
Open AccessPredicting the alternative conformation of a known protein structure based on the distogram of AlphaFold2
AlphaFold2 (AF2) has become one of the most established and leading tools for predicting protein structures, although it often favors a single conformation. While it has been shown that sampling different conformations is possible by manipulating the multiple sequence alignment (MSA), the potential for multi‐state information within a standard AF2 prediction remains underexplored. In this work, we demonstrate that even with a deep MSA, AF2 predictions encode structural heterogeneity. Specifically, we observe that the predicted residue–residue distance distributions often exhibit multiple probability peaks before recycling, suggesting the presence of alternative states. We derive state‐specific distance restraints by fitting and separating these multi‐peak distributions. These restraints are then incorporated into Rosetta as energy terms to model large‐scale conformational transitions. Across multiple protein systems undergoing diverse types of conformational changes, our method effectively samples alternative conformations starting from an input conformation, producing models comparable to those from existing approaches. Our findings highlight a previously underutilized feature of AF2 predictions and offer a novel strategy to take advantage of these insights to study conformational flexibility and dynamics in proteins. This work broadens the applicability of AF2 in structural biology and the modeling of multi‐state protein systems.
- 3 July 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 66 (14) , 8594-8606
FoldDoF: Utilizing the Primary Degrees of Freedom of Protein Backbone for Geometric Modeling and Generation
Robust modeling of protein structures through internal coordinates requires modeling a mixture of bond angles and torsion angles. Smoothing out bond angles can lead to significant error accumulation during structure reconstruction. We propose using a concise yet precise representation for modeling and generating protein structures, which views the protein backbone structure as a sequence of 3D rotations of peptide units, unifying bond angles and torsion angles on a single rotation manifold. We therefore present differentiable algorithms for efficient conversion between such internal coordinates and Cartesian coordinates. We demonstrate that the relative 3D rotations between consecutive peptide units describe the primary degrees of freedom of protein backbone conformation and thus guarantee the fidelity of the reconstructed full-atomic backbone structures in an optimization-free manner. Given the advantage of this representation that allows for straightforward geometric reasoning and optimization in 3D space, we incorporated this peptide-unit-centric formulation into protein backbone generative models, resulting in FrameFlow variants with improved in silico design performance in terms of diversity, novelty, and length generalizability. The enhanced efficacy of probabilistic modeling of backbone conformations demonstrates the superiority of this backbone representation and lays the foundation for developing more effective generative models for both protein design and structure prediction.
- 3 July 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 66 (14) , 8594-8606
FoldDoF: Utilizing the Primary Degrees of Freedom of Protein Backbone for Geometric Modeling and Generation
Robust modeling of protein structures through internal coordinates requires modeling a mixture of bond angles and torsion angles. Smoothing out bond angles can lead to significant error accumulation during structure reconstruction. We propose using a concise yet precise representation for modeling and generating protein structures, which views the protein backbone structure as a sequence of 3D rotations of peptide units, unifying bond angles and torsion angles on a single rotation manifold. We therefore present differentiable algorithms for efficient conversion between such internal coordinates and Cartesian coordinates. We demonstrate that the relative 3D rotations between consecutive peptide units describe the primary degrees of freedom of protein backbone conformation and thus guarantee the fidelity of the reconstructed full-atomic backbone structures in an optimization-free manner. Given the advantage of this representation that allows for straightforward geometric reasoning and optimization in 3D space, we incorporated this peptide-unit-centric formulation into protein backbone generative models, resulting in FrameFlow variants with improved in silico design performance in terms of diversity, novelty, and length generalizability. The enhanced efficacy of probabilistic modeling of backbone conformations demonstrates the superiority of this backbone representation and lays the foundation for developing more effective generative models for both protein design and structure prediction.
- Open Access
- 26 June 2026
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 42 (7)
PepMCP: A Graph-Based Membrane Contact Probability Predictor for Membrane-Lytic Antimicrobial Peptides
Motivation: The membrane-lytic mechanism of antimicrobial peptides (AMPs) is often overlooked during their in silico discovery process, largely due to the lack of a suitable metric for the membrane-binding propensity of peptides. Previously, we proposed a characteristic called membrane contact probability (MCP) and applied it to the identification of membrane proteins and membrane-lytic AMPs. However, previous MCP predictors were not trained on short peptides targeting bacterial membranes, which may result in unsatisfactory performance for peptide studies. Results: In this study, we present PepMCP, a peptide-tailored model for predicting MCP values of short peptides. We collected more than 500 membrane-lytic AMPs from the literature, conducted coarse-grained molecular dynamics (MD) simulations for these AMPs, and extracted their residue MCP labels from MD trajectories to train PepMCP. PepMCP employs the GraphSAGE framework to address this node regression task, encoding each peptide sequence as a graph with 4-hop edges. PepMCP achieved a Pearson correlation coefficient of 0.883 and an RMSE of 0.123 on the node-level test set. It can recognize membrane-lytic AMPs with the predicted MCP values for each sequence, thereby facilitating mechanism-driven AMP discovery. Additionally, we provide a database, MemAMPdb, which includes the membrane-lytic AMPs, as well as the PepMCP web server for easy access. Availability and Implementation: The code and data are available at https://github.com/ComputBiophys/PepMCP. Supplementary Information: Supplementary data are available online.
- 26 June 2026
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 42 (7)
Open AccessPepMCP: A Graph-Based Membrane Contact Probability Predictor for Membrane-Lytic Antimicrobial Peptides
Motivation: The membrane-lytic mechanism of antimicrobial peptides (AMPs) is often overlooked during their in silico discovery process, largely due to the lack of a suitable metric for the membrane-binding propensity of peptides. Previously, we proposed a characteristic called membrane contact probability (MCP) and applied it to the identification of membrane proteins and membrane-lytic AMPs. However, previous MCP predictors were not trained on short peptides targeting bacterial membranes, which may result in unsatisfactory performance for peptide studies. Results: In this study, we present PepMCP, a peptide-tailored model for predicting MCP values of short peptides. We collected more than 500 membrane-lytic AMPs from the literature, conducted coarse-grained molecular dynamics (MD) simulations for these AMPs, and extracted their residue MCP labels from MD trajectories to train PepMCP. PepMCP employs the GraphSAGE framework to address this node regression task, encoding each peptide sequence as a graph with 4-hop edges. PepMCP achieved a Pearson correlation coefficient of 0.883 and an RMSE of 0.123 on the node-level test set. It can recognize membrane-lytic AMPs with the predicted MCP values for each sequence, thereby facilitating mechanism-driven AMP discovery. Additionally, we provide a database, MemAMPdb, which includes the membrane-lytic AMPs, as well as the PepMCP web server for easy access. Availability and Implementation: The code and data are available at https://github.com/ComputBiophys/PepMCP. Supplementary Information: Supplementary data are available online.
- Open Access
SteerAF: Distogram-based Steering of AlphaFold2 toward Alternative Conformations
End-to-end structure predictors, such as AlphaFold2, typically output only the dominant conformational state of a given protein, which is biased by the training dataset. Existing strategies for recovering alternative conformations are often computationally expensive and offer limited biological interpretability. Here, we present SteerAF, an inference-time optimization framework based on AlphaFold2 that leverages information encoded in the distogram derived from deep multiple sequence alignments (MSAs) to predict alternative protein conformations. Across four benchmark datasets, SteerAF matches or surpasses existing methods in predicting alternative conformations for the majority of systems. Sparse MSA-feature modifications generated via block gradient ascent exhibit a strong correlation with experimentally characterized functional residues, recovering them with approximately 50% precision in the tested proteins. Furthermore, SteerAF enables effective decoy selection in the absence of experimental structures, and its predictions can serve as seed structures for molecular dynamics simulations to map conformational landscapes. Thus, SteerAF provides an efficient and interpretable approach for predicting alternative conformations, offering a framework that can be extended to other similar predictors and problems.
Open AccessSteerAF: Distogram-based Steering of AlphaFold2 toward Alternative Conformations
End-to-end structure predictors, such as AlphaFold2, typically output only the dominant conformational state of a given protein, which is biased by the training dataset. Existing strategies for recovering alternative conformations are often computationally expensive and offer limited biological interpretability. Here, we present SteerAF, an inference-time optimization framework based on AlphaFold2 that leverages information encoded in the distogram derived from deep multiple sequence alignments (MSAs) to predict alternative protein conformations. Across four benchmark datasets, SteerAF matches or surpasses existing methods in predicting alternative conformations for the majority of systems. Sparse MSA-feature modifications generated via block gradient ascent exhibit a strong correlation with experimentally characterized functional residues, recovering them with approximately 50% precision in the tested proteins. Furthermore, SteerAF enables effective decoy selection in the absence of experimental structures, and its predictions can serve as seed structures for molecular dynamics simulations to map conformational landscapes. Thus, SteerAF provides an efficient and interpretable approach for predicting alternative conformations, offering a framework that can be extended to other similar predictors and problems.
- Open Access
Architecture of a portal complex embedded in the poxvirus core
Summary: Gene transcription within the viral core is a unique feature of poxvirus replication. Following entry into the cytoplasm, the poxvirus core dissociates from the lateral bodies and undergoes expansion, functioning as a compartment for early transcription. However, the mechanisms governing molecular exchange between the viral core and the host cytoplasm remain poorly understood. Here, we determine the structures of the portal complex and its pore on the poxvirus core at 7.1 Å and 4.9 Å resolution, respectively, using cryo-electron tomography and sub-tomogram averaging. The pore is assembled from three viral proteins, E8, E6, and L3, for which we constructed an atomic model. Structural and channel analyses reveal that the pore satisfies the geometric and electrostatic requirements for the transport of RNA and smaller negatively charged molecules, while excluding double-stranded DNA and cytosolic DNA sensors. Together, our findings establish a structural framework for understanding the assembly and function of the poxvirus portal complex and identify potential targets for antiviral intervention.
Open AccessArchitecture of a portal complex embedded in the poxvirus core
Summary: Gene transcription within the viral core is a unique feature of poxvirus replication. Following entry into the cytoplasm, the poxvirus core dissociates from the lateral bodies and undergoes expansion, functioning as a compartment for early transcription. However, the mechanisms governing molecular exchange between the viral core and the host cytoplasm remain poorly understood. Here, we determine the structures of the portal complex and its pore on the poxvirus core at 7.1 Å and 4.9 Å resolution, respectively, using cryo-electron tomography and sub-tomogram averaging. The pore is assembled from three viral proteins, E8, E6, and L3, for which we constructed an atomic model. Structural and channel analyses reveal that the pore satisfies the geometric and electrostatic requirements for the transport of RNA and smaller negatively charged molecules, while excluding double-stranded DNA and cytosolic DNA sensors. Together, our findings establish a structural framework for understanding the assembly and function of the poxvirus portal complex and identify potential targets for antiviral intervention.
- Open Access
PepMCP: A Graph-Based Membrane Contact Probability Predictor for Membrane-Lytic Antimicrobial Peptides
Motivation: The membrane-lytic mechanism of antimicrobial peptides (AMPs) is often overlooked during their in silico discovery process, largely due to the lack of a suitable metric for the membrane-binding propensity of peptides. Previously, we proposed a characteristic called membrane contact probability (MCP) and applied it to the identification of membrane proteins and membrane-lytic AMPs. However, previous MCP predictors were not trained on short peptides targeting bacterial membranes, which may result in unsatisfactory performance for peptide studies. Results: In this study, we present PepMCP, a peptide-tailored model for predicting MCP values of short peptides. We collected more than 500 membrane-lytic AMPs from the literature, conducted coarse-grained molecular dynamics (MD) simulations for these AMPs, and extracted their residue MCP labels from MD trajectories to train PepMCP. PepMCP employs the GraphSAGE framework to address this node regression task, encoding each peptide sequence as a graph with 4-hop edges. PepMCP achieved a Pearson correlation coefficient of 0. 883 and an RMSE of 0. 123 on the node-level test set. It can recognize membrane-lytic AMPs with the predicted MCP values for each sequence, thereby facilitating mechanism-driven AMP discovery. Additionally, we provide a database, MemAMPdb, which includes the membrane-lytic AMPs, as well as the PepMCP web server for easy access. Availability and Implementation: The code and data are available at https://github.com/ComputBiophys/PepMCP.
Open AccessPepMCP: A Graph-Based Membrane Contact Probability Predictor for Membrane-Lytic Antimicrobial Peptides
Motivation: The membrane-lytic mechanism of antimicrobial peptides (AMPs) is often overlooked during their in silico discovery process, largely due to the lack of a suitable metric for the membrane-binding propensity of peptides. Previously, we proposed a characteristic called membrane contact probability (MCP) and applied it to the identification of membrane proteins and membrane-lytic AMPs. However, previous MCP predictors were not trained on short peptides targeting bacterial membranes, which may result in unsatisfactory performance for peptide studies. Results: In this study, we present PepMCP, a peptide-tailored model for predicting MCP values of short peptides. We collected more than 500 membrane-lytic AMPs from the literature, conducted coarse-grained molecular dynamics (MD) simulations for these AMPs, and extracted their residue MCP labels from MD trajectories to train PepMCP. PepMCP employs the GraphSAGE framework to address this node regression task, encoding each peptide sequence as a graph with 4-hop edges. PepMCP achieved a Pearson correlation coefficient of 0. 883 and an RMSE of 0. 123 on the node-level test set. It can recognize membrane-lytic AMPs with the predicted MCP values for each sequence, thereby facilitating mechanism-driven AMP discovery. Additionally, we provide a database, MemAMPdb, which includes the membrane-lytic AMPs, as well as the PepMCP web server for easy access. Availability and Implementation: The code and data are available at https://github.com/ComputBiophys/PepMCP.
- Open Access
Predicting the alternative conformation of a known protein structure based on the distance map of AlphaFold2
With AlphaFold2 (AF2) becoming the top structural prediction tool, multiple studies have found that AF2 often favors one conformation state over others in high-precision structure predictions. Meanwhile, it has also been demonstrated that the prediction of multi-state structures from a given protein sequence is possible by subsampling multiple sequence alignment (MSA). In this work, we reveal that AF2 predictions contain information on multi-state structures even with the deepest MSA: protein distance maps extracted from AF2 often exhibit multi-peak signals in the distance probability distributions for residue pairs. By fitting and separating these multi-peak distributions of residue pairs, one can extract distinct distance information of two states, which can be incorporated into Rosetta as restraint energy functions to model large and complex conformational changes. Twenty protein systems with different types of conformational changes were selected for validation in modeling their alternative conformations. With our protocol, we successfully predicted the alternative conformations of 19 systems and achieved a template-based modeling score (TM-score) above 0.90 for the best-sampled models in nine cases. This work further expands the usage of AlphaFold2 in studying multi-state proteins.
Open AccessPredicting the alternative conformation of a known protein structure based on the distance map of AlphaFold2
With AlphaFold2 (AF2) becoming the top structural prediction tool, multiple studies have found that AF2 often favors one conformation state over others in high-precision structure predictions. Meanwhile, it has also been demonstrated that the prediction of multi-state structures from a given protein sequence is possible by subsampling multiple sequence alignment (MSA). In this work, we reveal that AF2 predictions contain information on multi-state structures even with the deepest MSA: protein distance maps extracted from AF2 often exhibit multi-peak signals in the distance probability distributions for residue pairs. By fitting and separating these multi-peak distributions of residue pairs, one can extract distinct distance information of two states, which can be incorporated into Rosetta as restraint energy functions to model large and complex conformational changes. Twenty protein systems with different types of conformational changes were selected for validation in modeling their alternative conformations. With our protocol, we successfully predicted the alternative conformations of 19 systems and achieved a template-based modeling score (TM-score) above 0.90 for the best-sampled models in nine cases. This work further expands the usage of AlphaFold2 in studying multi-state proteins.
- Open Access
Exploring the alternative conformation of a known protein structure based on contact map prediction
The rapid development of deep learning-based methods has considerably advanced the field of protein structure prediction. The accuracy of predicting the 3D structures of simple proteins is comparable to that of experimentally determined structures, providing broad possibilities for structure-based biological studies. Another critical question is whether and how multistate structures can be predicted from a given protein sequence. In this study, analysis of multiple two-state proteins demonstrated that deep learning-based contact map predictions contain structural information on both states, which suggests that it is probably appropriate to change the target of deep learningbased protein structure prediction from one specific structure to multiple likely structures. Furthermore, by combining deep learning- and physics-based computational methods, we developed a protocol for exploring alternative conformations from a known structure of a given protein, by which we successfully approached the holo-state conformation of a leucine-binding protein from its apo-state structure.
Open AccessExploring the alternative conformation of a known protein structure based on contact map prediction
The rapid development of deep learning-based methods has considerably advanced the field of protein structure prediction. The accuracy of predicting the 3D structures of simple proteins is comparable to that of experimentally determined structures, providing broad possibilities for structure-based biological studies. Another critical question is whether and how multistate structures can be predicted from a given protein sequence. In this study, analysis of multiple two-state proteins demonstrated that deep learning-based contact map predictions contain structural information on both states, which suggests that it is probably appropriate to change the target of deep learningbased protein structure prediction from one specific structure to multiple likely structures. Furthermore, by combining deep learning- and physics-based computational methods, we developed a protocol for exploring alternative conformations from a known structure of a given protein, by which we successfully approached the holo-state conformation of a leucine-binding protein from its apo-state structure.
- Open Access
- 19 November 2021
- Published by Springer Nature in Communications Biology
- Vol. 4 (1) , 1-11
PremPLI: a machine learning model for predicting the effects of missense mutations on protein-ligand interactions
Resistance to small-molecule drugs is the main cause of the failure of therapeutic drugs in clinical practice. Missense mutations altering the binding of ligands to proteins are one of the critical mechanisms that result in genetic disease and drug resistance. Computational methods have made a lot of progress for predicting binding affinity changes and identifying resistance mutations, but their prediction accuracy and speed are still not satisfied and need to be further improved. To address these issues, we introduce a structure-based machine learning method for quantitatively estimating the effects of single mutations on ligand binding affinity changes (named as PremPLI). A comprehensive comparison of the predictive performance of PremPLI with other available methods on two benchmark datasets confirms that our approach performs robustly and presents similar or even higher predictive accuracy than the approaches relying on first-principle statistical mechanics and mixed physics- and knowledge-based potentials while requires much less computational resources. PremPLI can be used for guiding the design of ligand-binding proteins, identifying and understanding disease driver mutations, and finding potential resistance mutations for different drugs. PremPLI is freely available athttps://lilab.jysw.suda.edu.cn/research/PremPLI/and allows to do large-scale mutational scanning.
- 19 November 2021
- Published by Springer Nature in Communications Biology
- Vol. 4 (1) , 1-11
Open AccessPremPLI: a machine learning model for predicting the effects of missense mutations on protein-ligand interactions
Resistance to small-molecule drugs is the main cause of the failure of therapeutic drugs in clinical practice. Missense mutations altering the binding of ligands to proteins are one of the critical mechanisms that result in genetic disease and drug resistance. Computational methods have made a lot of progress for predicting binding affinity changes and identifying resistance mutations, but their prediction accuracy and speed are still not satisfied and need to be further improved. To address these issues, we introduce a structure-based machine learning method for quantitatively estimating the effects of single mutations on ligand binding affinity changes (named as PremPLI). A comprehensive comparison of the predictive performance of PremPLI with other available methods on two benchmark datasets confirms that our approach performs robustly and presents similar or even higher predictive accuracy than the approaches relying on first-principle statistical mechanics and mixed physics- and knowledge-based potentials while requires much less computational resources. PremPLI can be used for guiding the design of ligand-binding proteins, identifying and understanding disease driver mutations, and finding potential resistance mutations for different drugs. PremPLI is freely available athttps://lilab.jysw.suda.edu.cn/research/PremPLI/and allows to do large-scale mutational scanning.