Dongbo Bu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Dongbo Bu has a total of 268 co-authors. Here are the top co-authors by number of shared publications:
Haicang Zhang
Chinese Academy of Sciences
18 shared publications- Shiwei Sun
Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences, 6 Kexueyuan South Road, Beijing 100080, China; University of Chinese Academy of Sciences, 19 Yuquan Road, Beijing 100049, China
15 shared publications Jianwei Zhu
Microsoft Research Asia, Beijing, China
14 shared publicationsShiwei Sun
Key Lab of Intelligent Information Processing, SKLP, Institute of Computing Technology, Chinese Academy of Sciences
14 shared publicationsChungong Yu
Institute of Computing Technology, Chinese Academy of Sciences
13 shared publications- Wei-Mou Zheng
Institute of Theoretical Physics, Chinese Academy of Sciences, Beijing, China
12 shared publications
Latest Publications
- Open Access
- 12 June 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
- Vol. 31 (6) , 1-28
AiDRC: Accelerating Detailed Routing by AI-Driven Design Rule Violation Prediction and Checking
Design rule violation (DRV) evaluation and optimization constitute a critical challenge in modern VLSI physical design. Fast and accurate assessments of routability and DRV have gained significant research attention due to their pivotal role in improving design closure efficiency. Traditional detailed routing and design rule checking (DRC) are computationally expensive. To address the above challenges, this work leverages AI models for the specific location prediction of DRVs in the pre-detailed routing stage and the checking of DRVs during detailed routing, which achieves fast and precise DRV evaluation. By integrating the crisscross attention mechanism and channel transformer within a ResNet backbone, our proposed DRV prediction and checking model gains the capability to learn non-local and cross-layer feature relationships and mitigates the negative impact of data imbalance. Experiments demonstrate the effectiveness of our prediction model and checking model, with area under curve (AUC) of 0.987 and F1-score of 0.934, respectively. Remarkably, when integrated into an open-source detailed routing tool, our DRV prediction model achieves 16× acceleration, while our DRV checking model achieves a 293× acceleration over the conventional DRC engine. By applying the predicted DRVs to the detailed routing tool, our framework eliminates an average of 44% DRVs of the initial detailed routing results, effectively bridging the gap between data-driven predictions and rule-based detailed routing workflows.
- 12 June 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
- Vol. 31 (6) , 1-28
Open AccessAiDRC: Accelerating Detailed Routing by AI-Driven Design Rule Violation Prediction and Checking
Design rule violation (DRV) evaluation and optimization constitute a critical challenge in modern VLSI physical design. Fast and accurate assessments of routability and DRV have gained significant research attention due to their pivotal role in improving design closure efficiency. Traditional detailed routing and design rule checking (DRC) are computationally expensive. To address the above challenges, this work leverages AI models for the specific location prediction of DRVs in the pre-detailed routing stage and the checking of DRVs during detailed routing, which achieves fast and precise DRV evaluation. By integrating the crisscross attention mechanism and channel transformer within a ResNet backbone, our proposed DRV prediction and checking model gains the capability to learn non-local and cross-layer feature relationships and mitigates the negative impact of data imbalance. Experiments demonstrate the effectiveness of our prediction model and checking model, with area under curve (AUC) of 0.987 and F1-score of 0.934, respectively. Remarkably, when integrated into an open-source detailed routing tool, our DRV prediction model achieves 16× acceleration, while our DRV checking model achieves a 293× acceleration over the conventional DRC engine. By applying the predicted DRVs to the detailed routing tool, our framework eliminates an average of 44% DRVs of the initial detailed routing results, effectively bridging the gap between data-driven predictions and rule-based detailed routing workflows.
- Open Access
Objective curriculum-guided design of multi-property proteins
Designing functional proteins that simultaneously possess multiple biochemical properties remains a significant challenge, as key protein properties, such as solubility, stability, binding affinity, and chemical resistance, are often interdependent or even conflicting. Current approaches typically attempt to jointly optimize multiple functional objectives in one shot, followed by extensive screening to identify rare feasible designs. Here, we introduce OCDesign, an objective curriculum–guided framework for multi–property protein design. OCDesign is based on the objective curriculum principle, i.e., the order in which objectives are introduced can shape the accessibility of functional solutions. In each design round of OCDesign, candidate sequences are generated in silico , assessed across multiple properties, selected based on Pareto-optimal trade–offs, and experimentally validated, with each experimental stage testing the role of the newly introduced objective within the curriculum. Using antibody–binding protein A as a model system, we show that one-shot optimization fails to yield functional designs, whereas a staged curriculum—progressing from solubility and structural consistency to binding affinity, and then to alkaline resistance—enables the design of proteins possessing multiple desired properties through substantially fewer wet-lab experiments. These results establish OCDesign as a practical computational–experimental strategy for organizing and integrating multiple objectives in protein design, and suggest that objective ordering is a key determinant of accessibility in high–dimensional design spaces.
Open AccessObjective curriculum-guided design of multi-property proteins
Designing functional proteins that simultaneously possess multiple biochemical properties remains a significant challenge, as key protein properties, such as solubility, stability, binding affinity, and chemical resistance, are often interdependent or even conflicting. Current approaches typically attempt to jointly optimize multiple functional objectives in one shot, followed by extensive screening to identify rare feasible designs. Here, we introduce OCDesign, an objective curriculum–guided framework for multi–property protein design. OCDesign is based on the objective curriculum principle, i.e., the order in which objectives are introduced can shape the accessibility of functional solutions. In each design round of OCDesign, candidate sequences are generated in silico , assessed across multiple properties, selected based on Pareto-optimal trade–offs, and experimentally validated, with each experimental stage testing the role of the newly introduced objective within the curriculum. Using antibody–binding protein A as a model system, we show that one-shot optimization fails to yield functional designs, whereas a staged curriculum—progressing from solubility and structural consistency to binding affinity, and then to alkaline resistance—enables the design of proteins possessing multiple desired properties through substantially fewer wet-lab experiments. These results establish OCDesign as a practical computational–experimental strategy for organizing and integrating multiple objectives in protein design, and suggest that objective ordering is a key determinant of accessibility in high–dimensional design spaces.
- Open Access
Sampling protein structural token space enables accurate prediction of multiple conformations
Protein function is fundamentally mediated by ensembles of distinct metastable states. However, existing methods typically exhibit a bias toward predicting a single dominant state, failing to capture alternative conformations or provide robust metrics for identifying high-quality multi-state conformations. Here, we present MultiStateFold (MSFold), a framework that integrates Parallel Tempering into the discrete structure token space of the ESM3 protein language model. By conceptualizing the model’s latent space as an implicit energy landscape, MSFold enables global exploration and barrier crossing, thereby overcoming the local sampling limitations inherent in base generative models. Across a benchmark of 313 multi-conformation pairs, MSFold sets a new performance standard: it achieves the highest success rate in modeling native states and substantially outperforms leading methods, including AlphaFold 3 and MSA Cluster, on challenging alternative conformations, while maintaining competitive accuracy for primary structures. Furthermore, we propose Sequence Log-Likelihood (SLL), a novel confidence metric derived from sequence-structure consistency. Our results demonstrate that SLL offers a modest improvement over standard metrics such as pTM and pLDDT. This work establishes a new paradigm for conformational sampling, bridging classical statistical physics with protein language models.
Open AccessSampling protein structural token space enables accurate prediction of multiple conformations
Protein function is fundamentally mediated by ensembles of distinct metastable states. However, existing methods typically exhibit a bias toward predicting a single dominant state, failing to capture alternative conformations or provide robust metrics for identifying high-quality multi-state conformations. Here, we present MultiStateFold (MSFold), a framework that integrates Parallel Tempering into the discrete structure token space of the ESM3 protein language model. By conceptualizing the model’s latent space as an implicit energy landscape, MSFold enables global exploration and barrier crossing, thereby overcoming the local sampling limitations inherent in base generative models. Across a benchmark of 313 multi-conformation pairs, MSFold sets a new performance standard: it achieves the highest success rate in modeling native states and substantially outperforms leading methods, including AlphaFold 3 and MSA Cluster, on challenging alternative conformations, while maintaining competitive accuracy for primary structures. Furthermore, we propose Sequence Log-Likelihood (SLL), a novel confidence metric derived from sequence-structure consistency. Our results demonstrate that SLL offers a modest improvement over standard metrics such as pTM and pLDDT. This work establishes a new paradigm for conformational sampling, bridging classical statistical physics with protein language models.
- Open Access
Designing faster mixed integer linear programming algorithm via learning the optimal path
Designing faster algorithms for solving Mixed-Integer Linear Programming (MILP) problems is highly desired across numerous practical domains, as a vast array of complex real-world challenges can be effectively modeled as MILP formulations. Solving these problems typically employs the branch-and-bound algorithm, the core of which can be conceived as searching for a path of nodes (or sub-problems) that contains the optimal solution to the original MILP problem. Traditional approaches to finding this path rely heavily on hand-crafted, intuition-based heuristic strategies, which often suffer from unstable and unpredictable performance across different MILP problem instances. To address this limitation, we introduce DeepBound, a deep learning-based node selection algorithm that automates the learning of such human intuition from data. The core of DeepBound lies in learning to prioritize nodes containing the optimal solution, thereby improving solving efficiency. DeepBound introduces a multi-level feature fusion network to capture the node representations. To tackle the inherent node imbalance in branch-and-bound trees, DeepBound employs a pairwise training paradigm that enhances the model's ability to discriminate between nodes. Extensive experiments on three NP-hard MILP benchmarks demonstrate that DeepBound achieves superior solving efficiency over conventional heuristic rules and existing learning-based approaches, obtaining optimal feasible solutions with significantly reduced computation time. Moreover, DeepBound demonstrates strong generalization capability on large and complex instances. The analysis of its learned features reveals that the method can automatically discover more flexible and robust feature selection, which may effectively improve and potentially replace human-designed heuristic rules.
Open AccessDesigning faster mixed integer linear programming algorithm via learning the optimal path
Designing faster algorithms for solving Mixed-Integer Linear Programming (MILP) problems is highly desired across numerous practical domains, as a vast array of complex real-world challenges can be effectively modeled as MILP formulations. Solving these problems typically employs the branch-and-bound algorithm, the core of which can be conceived as searching for a path of nodes (or sub-problems) that contains the optimal solution to the original MILP problem. Traditional approaches to finding this path rely heavily on hand-crafted, intuition-based heuristic strategies, which often suffer from unstable and unpredictable performance across different MILP problem instances. To address this limitation, we introduce DeepBound, a deep learning-based node selection algorithm that automates the learning of such human intuition from data. The core of DeepBound lies in learning to prioritize nodes containing the optimal solution, thereby improving solving efficiency. DeepBound introduces a multi-level feature fusion network to capture the node representations. To tackle the inherent node imbalance in branch-and-bound trees, DeepBound employs a pairwise training paradigm that enhances the model's ability to discriminate between nodes. Extensive experiments on three NP-hard MILP benchmarks demonstrate that DeepBound achieves superior solving efficiency over conventional heuristic rules and existing learning-based approaches, obtaining optimal feasible solutions with significantly reduced computation time. Moreover, DeepBound demonstrates strong generalization capability on large and complex instances. The analysis of its learned features reveals that the method can automatically discover more flexible and robust feature selection, which may effectively improve and potentially replace human-designed heuristic rules.
- Open Access
Predicting TCR-pMHC Binding by Reinforcement Learning
The binding between T cell receptors (TCRs) and peptide-major histocompatibility complexes (pMHCs) is fundamental to the immune system ability to recognize and eliminate pathogens. Accurate prediction of TCR-pMHC interactions holds significant promise for advancing cancer immunotherapy, vaccine design, and autoimmune disease research. However, existing approaches often treat the sequences, structures, and functions of TCRs, peptides, and MHC molecules in isolation, neglecting their interdependencies and hence limiting the prediction accuracy. In this study, we present ProTCR, a novel approach that integrates sequence, structural, and functional information within a reinforcement learning framework, offering a new paradigm for predicting TCR-pMHC binding. The reinforcement learning optimization enables ProTCR to generate TCR-pMHC sequences with enhanced binding propensity, thereby improving prediction accuracy. On benchmark datasets such as IEDB and VDJdb, ProTCR achieves an AUROC of 0.75, outperforming state-of-the-art methods by 32.7%, while offering interpretable insights into the structural and sequence determinants of binding. We further validate ProTCR using TCRs and neoantigens derived from a cervical cancer patient via proteogenomic profiling. Our analysis reveals a strong correlation between T cell clonal expansion and ProTCR-predicted TCR-peptide binding scores, supporting the biological relevance of the model. Additionally, ProTCR demonstrates robust performance in predicting SARS-CoV-2 TCR-pMHC complexes and generating MHC-specific peptides with potential applications in peptide-based immunotherapies. Collectively, these findings establish ProTCR as a powerful and interpretable tool for TCR-pMHC binding prediction, with broad utility across immunology research and translational applications.
Open AccessPredicting TCR-pMHC Binding by Reinforcement Learning
The binding between T cell receptors (TCRs) and peptide-major histocompatibility complexes (pMHCs) is fundamental to the immune system ability to recognize and eliminate pathogens. Accurate prediction of TCR-pMHC interactions holds significant promise for advancing cancer immunotherapy, vaccine design, and autoimmune disease research. However, existing approaches often treat the sequences, structures, and functions of TCRs, peptides, and MHC molecules in isolation, neglecting their interdependencies and hence limiting the prediction accuracy. In this study, we present ProTCR, a novel approach that integrates sequence, structural, and functional information within a reinforcement learning framework, offering a new paradigm for predicting TCR-pMHC binding. The reinforcement learning optimization enables ProTCR to generate TCR-pMHC sequences with enhanced binding propensity, thereby improving prediction accuracy. On benchmark datasets such as IEDB and VDJdb, ProTCR achieves an AUROC of 0.75, outperforming state-of-the-art methods by 32.7%, while offering interpretable insights into the structural and sequence determinants of binding. We further validate ProTCR using TCRs and neoantigens derived from a cervical cancer patient via proteogenomic profiling. Our analysis reveals a strong correlation between T cell clonal expansion and ProTCR-predicted TCR-peptide binding scores, supporting the biological relevance of the model. Additionally, ProTCR demonstrates robust performance in predicting SARS-CoV-2 TCR-pMHC complexes and generating MHC-specific peptides with potential applications in peptide-based immunotherapies. Collectively, these findings establish ProTCR as a powerful and interpretable tool for TCR-pMHC binding prediction, with broad utility across immunology research and translational applications.
- Open Access
ProAffinity++
Proteins are essential biological macromolecules that play a crucial role in living organisms. Protein-protein interactions, which govern various biological processes such as signal transduction, cell metabolism, and cell growth, are key aspect of protein function. The strength of these interactions, characterized by protein-protein binding affinity( Δ G), is a critical thermodynamic property of protein complexes. Accurate prediction of protein-protein binding affinity is meaningful for understanding the mechanisms of biological systems and improving the speed of binding affinity determination. However, current methods are limited in their accuracy and are primarily designed for binary complexes. To address these limitations, we propose ProAffinity++, an end-to-end algorithm that predicts binding affinities using a combined representation of sequence and structure. The core idea of ProAffinity++ is to model the binding regions of protein complexes and utilize graph neural networks to capture the local microenvironment of residues. Our experimental evaluations demonstrate that ProAffinity++ outperforms state-of-the-art methods in predicting protein-protein binding affinities. Moreover, it exhibits remarkable performance in challenging scenarios, such as antigen-antibody interactions and missense mutation problems, where existing methods have limitations.
Open AccessProAffinity++
Proteins are essential biological macromolecules that play a crucial role in living organisms. Protein-protein interactions, which govern various biological processes such as signal transduction, cell metabolism, and cell growth, are key aspect of protein function. The strength of these interactions, characterized by protein-protein binding affinity( Δ G), is a critical thermodynamic property of protein complexes. Accurate prediction of protein-protein binding affinity is meaningful for understanding the mechanisms of biological systems and improving the speed of binding affinity determination. However, current methods are limited in their accuracy and are primarily designed for binary complexes. To address these limitations, we propose ProAffinity++, an end-to-end algorithm that predicts binding affinities using a combined representation of sequence and structure. The core idea of ProAffinity++ is to model the binding regions of protein complexes and utilize graph neural networks to capture the local microenvironment of residues. Our experimental evaluations demonstrate that ProAffinity++ outperforms state-of-the-art methods in predicting protein-protein binding affinities. Moreover, it exhibits remarkable performance in challenging scenarios, such as antigen-antibody interactions and missense mutation problems, where existing methods have limitations.
- Open Access
- 23 October 2025
- Published by Oxford University Press (OUP) in Molecular Biology and Evolution
- Vol. 42 (11)
Accurate and Efficient Phylogenetic Inference through End-To-End Deep Learning
Accurate phylogenetic inference is crucial for understanding evolutionary relationships among species. Deep learning technique has been introduced for phylogenetic inference; however, the existing deep learning-based approaches either suffer from limited accuracy as they split inference into several disjoint stages, or exhibit low efficiency and hardly apply to the cases with over 20 species. Here, we present an accurate and efficient approach to phylogenetic inference. Our approach, called NeuralNJ, employs an end-to-end framework that directly constructs phylogenetic trees from the input taxa, thus effectively avoiding the inaccuracy incurred by the split inference stages. The key innovation of NeuralNJ lies in its learnable neighbor joining mechanism, which iteratively joins neighbors guided by learned priority scores and thereby achieves accurate tree reconstruction. The inference accuracy is further enhanced through incorporating reinforcement learning-based tree search. Using both simulated and empirical data as representatives, we demonstrate that NeuralNJ can effectively infer phylogenetic tree with improved computational efficiency and reconstruction accuracy. The study paves the way to accurate and efficient phylogenetic inference for hundreds of taxa in complex evolutionary scenarios.
- 23 October 2025
- Published by Oxford University Press (OUP) in Molecular Biology and Evolution
- Vol. 42 (11)
Open AccessAccurate and Efficient Phylogenetic Inference through End-To-End Deep Learning
Accurate phylogenetic inference is crucial for understanding evolutionary relationships among species. Deep learning technique has been introduced for phylogenetic inference; however, the existing deep learning-based approaches either suffer from limited accuracy as they split inference into several disjoint stages, or exhibit low efficiency and hardly apply to the cases with over 20 species. Here, we present an accurate and efficient approach to phylogenetic inference. Our approach, called NeuralNJ, employs an end-to-end framework that directly constructs phylogenetic trees from the input taxa, thus effectively avoiding the inaccuracy incurred by the split inference stages. The key innovation of NeuralNJ lies in its learnable neighbor joining mechanism, which iteratively joins neighbors guided by learned priority scores and thereby achieves accurate tree reconstruction. The inference accuracy is further enhanced through incorporating reinforcement learning-based tree search. Using both simulated and empirical data as representatives, we demonstrate that NeuralNJ can effectively infer phylogenetic tree with improved computational efficiency and reconstruction accuracy. The study paves the way to accurate and efficient phylogenetic inference for hundreds of taxa in complex evolutionary scenarios.
- Open Access
- 23 October 2025
- Published by Springer Nature in Scientific Reports
- Vol. 15 (1) , 1-13
Structural basis for antibiotic resistance by chloramphenicol acetyltransferase type A in Staphylococcus aureus
Methicillin-resistant Staphylococcus aureus (MRSA) exemplifies high-level antibiotic resistance in this major human pathogen. Its resistance to chloramphenicol is majorly conferred by enzymatic inactivation via chloramphenicol acetyltransferases (CATs). This modification sterically blocks the antibiotic’s ribosomal binding and thus neutralizes its inhibitory potency. Although CATs have been structurally studied across diverse bacteria species, the structures of S. aureus CATs (saCATs) have remained uncharacterized. To address this gap and elucidate species-specific resistance mechanisms, we determined the first high-resolution crystal structure of saCAT1, the prototypical saCAT enzyme. Structural analysis delineates the active site architecture and reveals the molecular basis for substrate recognition of both chloramphenicol and fusidic acid (FA). Further enzymatic assays demonstrated that the K m value against chloramphenicol is 16.9 µM, and the K i value of the inhibitor FA is 83.7 µM, indicating that the inhibitory capacity of FA is relatively limited. These findings provide an essential structural framework for understanding chloramphenicol resistance in S. aureus and facilitate the rational design of novel antimicrobial strategies to combat multidrug-resistant pathogens.
- 23 October 2025
- Published by Springer Nature in Scientific Reports
- Vol. 15 (1) , 1-13
Open AccessStructural basis for antibiotic resistance by chloramphenicol acetyltransferase type A in Staphylococcus aureus
Methicillin-resistant Staphylococcus aureus (MRSA) exemplifies high-level antibiotic resistance in this major human pathogen. Its resistance to chloramphenicol is majorly conferred by enzymatic inactivation via chloramphenicol acetyltransferases (CATs). This modification sterically blocks the antibiotic’s ribosomal binding and thus neutralizes its inhibitory potency. Although CATs have been structurally studied across diverse bacteria species, the structures of S. aureus CATs (saCATs) have remained uncharacterized. To address this gap and elucidate species-specific resistance mechanisms, we determined the first high-resolution crystal structure of saCAT1, the prototypical saCAT enzyme. Structural analysis delineates the active site architecture and reveals the molecular basis for substrate recognition of both chloramphenicol and fusidic acid (FA). Further enzymatic assays demonstrated that the K m value against chloramphenicol is 16.9 µM, and the K i value of the inhibitor FA is 83.7 µM, indicating that the inhibitory capacity of FA is relatively limited. These findings provide an essential structural framework for understanding chloramphenicol resistance in S. aureus and facilitate the rational design of novel antimicrobial strategies to combat multidrug-resistant pathogens.
- Open Access
Accurate and efficient phylogenetic inference through end-to-end deep learning
Accurate phylogenetic inference is crucial for understanding evolutionary relationships among species. Deep learning technique has been introduced for phylogenetic inference; however, the existing deep learning-based approaches either suffer from limited accuracy as they split inference into several disjoint stages, or exhibit low efficiency and hardly apply to the cases with over 20 species. Here we present an accurate and efficient approach to phylogenetic inference. Our approach, called NeuralNJ, employs an end-to-end framework that directly constructs phylogenetic trees from the input taxa, thus effectively avoiding the inaccuracy incurred by the split inference stages. The key innovation of NeuralNJ lies in its learnable neighbor joining mechanism, which iteratively joins neighbors guided by learned priority scores and thereby achieves accurate tree reconstruction. The inference accuracy is further enhanced through incorporating reinforcement learning-based tree search. Using both simulated and empirical data as representatives, we demonstrate that NeuralNJ can effectively infer phylogenetic tree with improved computational efficiency and reconstruction accuracy. The study paves the way to accurate and efficient phylogenetic inference for hundreds of taxa in complex evolutionary scenarios.
Open AccessAccurate and efficient phylogenetic inference through end-to-end deep learning
Accurate phylogenetic inference is crucial for understanding evolutionary relationships among species. Deep learning technique has been introduced for phylogenetic inference; however, the existing deep learning-based approaches either suffer from limited accuracy as they split inference into several disjoint stages, or exhibit low efficiency and hardly apply to the cases with over 20 species. Here we present an accurate and efficient approach to phylogenetic inference. Our approach, called NeuralNJ, employs an end-to-end framework that directly constructs phylogenetic trees from the input taxa, thus effectively avoiding the inaccuracy incurred by the split inference stages. The key innovation of NeuralNJ lies in its learnable neighbor joining mechanism, which iteratively joins neighbors guided by learned priority scores and thereby achieves accurate tree reconstruction. The inference accuracy is further enhanced through incorporating reinforcement learning-based tree search. Using both simulated and empirical data as representatives, we demonstrate that NeuralNJ can effectively infer phylogenetic tree with improved computational efficiency and reconstruction accuracy. The study paves the way to accurate and efficient phylogenetic inference for hundreds of taxa in complex evolutionary scenarios.
- Open Access
- 31 July 2025
- Published by Springer Nature in Research Square
Accurate Immune Protein Structure Prediction by Large Language Model and Transfer Learning
Accurate prediction of immune protein structures is critical for advancing immunotherapy. However, deep learning-based methods like AlphaFold and RosettaFold struggle with immune proteins due to the limited number of solved immune protein structures and the absence of homologous sequences in hypervariable regions. To address these challenges, we introduce ImmuneFold, a transfer-learning framework that leverages a large language model and low-rank adaptation (LoRA) for memory-efficient fine-tuning. ImmuneFold outperforms existing methods, including MSA-based AlphaFold3, in predicting the structures of T-cell receptors, antibodies, and nanobodies. Additionally, we pair ImmuneFold’s predictions with Rosetta energy scoring to develop a zero-shot protocol for TCR–epitope binding prediction, effectively mitigating overfitting issues common in supervised approaches. Experimental evaluations also confirm ImmuneFold’s robustness and accuracy in binding prediction. Beyond immune proteins, ImmuneFold provides a scalable framework for adapting advanced models, such as ESMFold and AlphaFold, to other protein families, thereby democratizing access to cutting-edge structural tools for researchers, even those with limited computational resources.
- 31 July 2025
- Published by Springer Nature in Research Square
Open AccessAccurate Immune Protein Structure Prediction by Large Language Model and Transfer Learning
Accurate prediction of immune protein structures is critical for advancing immunotherapy. However, deep learning-based methods like AlphaFold and RosettaFold struggle with immune proteins due to the limited number of solved immune protein structures and the absence of homologous sequences in hypervariable regions. To address these challenges, we introduce ImmuneFold, a transfer-learning framework that leverages a large language model and low-rank adaptation (LoRA) for memory-efficient fine-tuning. ImmuneFold outperforms existing methods, including MSA-based AlphaFold3, in predicting the structures of T-cell receptors, antibodies, and nanobodies. Additionally, we pair ImmuneFold’s predictions with Rosetta energy scoring to develop a zero-shot protocol for TCR–epitope binding prediction, effectively mitigating overfitting issues common in supervised approaches. Experimental evaluations also confirm ImmuneFold’s robustness and accuracy in binding prediction. Beyond immune proteins, ImmuneFold provides a scalable framework for adapting advanced models, such as ESMFold and AlphaFold, to other protein families, thereby democratizing access to cutting-edge structural tools for researchers, even those with limited computational resources.