Bozitao Zhong
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Bozitao Zhong has a total of 62 co-authors. Here are the top co-authors by number of shared publications:
- Liang Hong
Shanghai Jiao Tong University
6 shared publications - Lirong Zheng
Shanghai Jiao Tong University
5 shared publications - Mingchen Li
Shanghai Jiao Tong University
5 shared publications - Hai-Feng Chen
State Key Laboratory of Microbial Metabolism, Joint International Research Laboratory of Metabolic & Developmental Sciences, Department of Bioinformatics and Biostatistics, National Experimental Teaching Center for Life Sciences and Biotechnology, School of Life Sciences and Biotechnology, Shanghai Jiao Tong University, Shanghai, 200240, China
4 shared publications - Pan Tan
Shanghai Artificial Intelligence Laboratory, Shanghai 200232, China
4 shared publications Banghao Wu
Shanghai Jiao Tong University
3 shared publications
Latest Publications
- 21 October 2026
- Published byĀ Royal Society of Chemistry (RSC)Ā inĀ Physical Chemistry Chemical Physics
- Vol.Ā 28 Ā (40) , 25190-25199
Accurate m 6 A modification forcefield for conformation sampling of RNA
MODM6A force field could stabilize the local pairing of RNA.
- 21 October 2026
- Published byĀ Royal Society of Chemistry (RSC)Ā inĀ Physical Chemistry Chemical Physics
- Vol.Ā 28 Ā (40) , 25190-25199
Accurate m 6 A modification forcefield for conformation sampling of RNA
MODM6A force field could stabilize the local pairing of RNA.
- Open Access
- 6 September 2025
- Published byĀ Oxford University Press (OUP)Ā inĀ Briefings in Bioinformatics
- Vol.Ā 26 Ā (5)
Autoregressive enzyme function prediction with multi-scale multi-modality fusion
Accurate prediction of enzyme function is crucial for elucidating biological mechanisms and driving innovation across various sectors. Existing deep learning methods tend to rely solely on either sequence data or structural data and predict the Enzyme Commission (EC) number as a whole, neglecting the intrinsic hierarchical structure of EC numbers. To address these limitations, we introduce Multi-scale multi-modality Autoregressive Predictor (MAPred), a novel multi-modality and multi-scale model designed to autoregressively predict the EC number of proteins. MAPred integrates both the primary amino acid sequence and the 3D tokens of proteins, employing a dual-pathway approach to capture comprehensive protein characteristics and essential local functional sites. Additionally, MAPred utilizes an autoregressive prediction network to sequentially predict the digits of the EC number, leveraging the hierarchical organization of EC classifications. Evaluations on benchmark datasets, including New-392, Price, and New-815, demonstrate that our method outperforms existing models, marking a significant advance in the reliability and granularity of protein function prediction within bioinformatics.
- 6 September 2025
- Published byĀ Oxford University Press (OUP)Ā inĀ Briefings in Bioinformatics
- Vol.Ā 26 Ā (5)
Open AccessAutoregressive enzyme function prediction with multi-scale multi-modality fusion
Accurate prediction of enzyme function is crucial for elucidating biological mechanisms and driving innovation across various sectors. Existing deep learning methods tend to rely solely on either sequence data or structural data and predict the Enzyme Commission (EC) number as a whole, neglecting the intrinsic hierarchical structure of EC numbers. To address these limitations, we introduce Multi-scale multi-modality Autoregressive Predictor (MAPred), a novel multi-modality and multi-scale model designed to autoregressively predict the EC number of proteins. MAPred integrates both the primary amino acid sequence and the 3D tokens of proteins, employing a dual-pathway approach to capture comprehensive protein characteristics and essential local functional sites. Additionally, MAPred utilizes an autoregressive prediction network to sequentially predict the digits of the EC number, leveraging the hierarchical organization of EC classifications. Evaluations on benchmark datasets, including New-392, Price, and New-815, demonstrate that our method outperforms existing models, marking a significant advance in the reliability and granularity of protein function prediction within bioinformatics.
- Top Cited PapersOpen Access
- 5 July 2025
- Published byĀ Springer NatureĀ inĀ Nature Communications
- Vol.Ā 16 Ā (1) , 1-12
Harnessing protein language model for structure-based discovery of highly efficient and robust PET hydrolases
Plastic waste, particularly polyethylene terephthalate (PET), presents significant environmental challenges, driving extensive research into enzymatic biodegradation. However, existing PET hydrolases (PETases) are limited by narrow sequence diversity and suboptimal performance. This study introduces VenusMine, a protein discovery pipeline that integrates protein language models (PLMs) with a representation tree to identify PETases based on structural similarity using sequence information. Using the crystal structure of IsPETase as a template, VenusMine identifies and clusters target proteins. Candidates are further screened using PLM-based assessments of solubility and thermostability, leading to the selection of 34 proteins for biochemical validation. Results reveal that 14 candidates exhibit PET degradation activity across 30ā60 °C. Notably, a PET hydrolase from Kibdelosporangium banguiense (KbPETase) demonstrates a melting temperature (Tm) 32 °C higher than IsPETase and exhibits the highest PET degradation activity within 30 ā 65 °C among wild-type PETases. KbPETase also surpasses FastPETase and LCC in catalytic efficiency. X-ray crystallography and molecular dynamics simulations show that KbPETase possesses a conserved catalytic domain and enhanced intramolecular interactions, underpinning its improved functionality and thermostability. This work demonstrates a novel deep learning approach for discovering natural PETases with enhanced properties.
- 5 July 2025
- Published byĀ Springer NatureĀ inĀ Nature Communications
- Vol.Ā 16 Ā (1) , 1-12
Top Cited PapersOpen AccessHarnessing protein language model for structure-based discovery of highly efficient and robust PET hydrolases
Plastic waste, particularly polyethylene terephthalate (PET), presents significant environmental challenges, driving extensive research into enzymatic biodegradation. However, existing PET hydrolases (PETases) are limited by narrow sequence diversity and suboptimal performance. This study introduces VenusMine, a protein discovery pipeline that integrates protein language models (PLMs) with a representation tree to identify PETases based on structural similarity using sequence information. Using the crystal structure of IsPETase as a template, VenusMine identifies and clusters target proteins. Candidates are further screened using PLM-based assessments of solubility and thermostability, leading to the selection of 34 proteins for biochemical validation. Results reveal that 14 candidates exhibit PET degradation activity across 30ā60 °C. Notably, a PET hydrolase from Kibdelosporangium banguiense (KbPETase) demonstrates a melting temperature (Tm) 32 °C higher than IsPETase and exhibits the highest PET degradation activity within 30 ā 65 °C among wild-type PETases. KbPETase also surpasses FastPETase and LCC in catalytic efficiency. X-ray crystallography and molecular dynamics simulations show that KbPETase possesses a conserved catalytic domain and enhanced intramolecular interactions, underpinning its improved functionality and thermostability. This work demonstrates a novel deep learning approach for discovering natural PETases with enhanced properties.
- Open Access
Venus-MAXWELL: Efficient Learning of Protein-Mutation Stability Landscapes using Protein Language Models
In-silico prediction of protein mutant stability, measured by the difference in Gibbs free energy change (ĪĪ G ), is fundamental for protein engineering. Current sequence-to-label methods typically employ the two-stage pipeline: ( i ) encoding mutant sequences using neural networks ( e . g ., transformers), followed by ( ii ) the ĪĪ G regression from the latent representations. Although these methods have demonstrated promising performance, their dependence on specialized neural network encoders significantly increases the complexity. Additionally, the requirement to individually compute latent representations for each mutant site negatively impacts computational efficiency and poses the risk of overfitting . This work proposes the Venus-M axwell framework, which reformulates mutation ĪĪ G prediction as a sequence-to-landscape task. In Venus-M axwell , mutations of a protein and their corresponding ĪĪ G values are organized into a landscape matrix, allowing our framework to learn the ĪĪ G landscape of a protein with a single forward and backward pass during training. Besides, to facilitate future works, we also curated a large-scale ĪĪ G dataset with strict controls on data leakage and redundancy to ensure robust evaluation. Venus-M axwell is compatible with multiple protein language models and enables these models for accurate and efficient ĪĪ G prediction. For example, when integrated with the ESM-IF, Venus-M axwell achieves higher accuracy than ThermoMPNN with 10Ć faster in inference speed (despite having 50Ć more parameters than ThermoMPNN). The training codes, model weights, and datasets are publicly available at https://github.com/ai4protein/Venus-MAXWELL .
Open AccessVenus-MAXWELL: Efficient Learning of Protein-Mutation Stability Landscapes using Protein Language Models
In-silico prediction of protein mutant stability, measured by the difference in Gibbs free energy change (ĪĪ G ), is fundamental for protein engineering. Current sequence-to-label methods typically employ the two-stage pipeline: ( i ) encoding mutant sequences using neural networks ( e . g ., transformers), followed by ( ii ) the ĪĪ G regression from the latent representations. Although these methods have demonstrated promising performance, their dependence on specialized neural network encoders significantly increases the complexity. Additionally, the requirement to individually compute latent representations for each mutant site negatively impacts computational efficiency and poses the risk of overfitting . This work proposes the Venus-M axwell framework, which reformulates mutation ĪĪ G prediction as a sequence-to-landscape task. In Venus-M axwell , mutations of a protein and their corresponding ĪĪ G values are organized into a landscape matrix, allowing our framework to learn the ĪĪ G landscape of a protein with a single forward and backward pass during training. Besides, to facilitate future works, we also curated a large-scale ĪĪ G dataset with strict controls on data leakage and redundancy to ensure robust evaluation. Venus-M axwell is compatible with multiple protein language models and enables these models for accurate and efficient ĪĪ G prediction. For example, when integrated with the ESM-IF, Venus-M axwell achieves higher accuracy than ThermoMPNN with 10Ć faster in inference speed (despite having 50Ć more parameters than ThermoMPNN). The training codes, model weights, and datasets are publicly available at https://github.com/ai4protein/Venus-MAXWELL .
- Open Access
- 28 February 2025
- Published byĀ American Physical Society (APS)Ā inĀ Physical Review Research
- Vol.Ā 7 Ā (1)
Entropy-driven zero-shot deep learning model selection for viral proteins
Predicting the fitness of viral proteins is fundamental to understanding viral evolution and developing antiviral strategies. This study introduces the Venus-EEM, an entropy-driven ensemble model, aimed at improving the performance of zero-shot predictions for protein fitness across diverse viral datasets. We demonstrate that entropy serves as an effective criterion for selecting optimal zero-shot models, enabling adaptive model selection for different prediction tasks. By incorporating entropy-weighted ensemble learning from multiple protein language models, Venus-EEM achieves superior performance compared to existing methods. We validate the model's effectiveness through comprehensive evaluation on multiple viral datasets and a detailed case study of T7 RNA polymerase (T7 RNAP) activity. Our findings provide an effective approach for predicting viral protein mutations based on entropy, bridging fundamental physics principles with practical biological challenges.
- 28 February 2025
- Published byĀ American Physical Society (APS)Ā inĀ Physical Review Research
- Vol.Ā 7 Ā (1)
Open AccessEntropy-driven zero-shot deep learning model selection for viral proteins
Predicting the fitness of viral proteins is fundamental to understanding viral evolution and developing antiviral strategies. This study introduces the Venus-EEM, an entropy-driven ensemble model, aimed at improving the performance of zero-shot predictions for protein fitness across diverse viral datasets. We demonstrate that entropy serves as an effective criterion for selecting optimal zero-shot models, enabling adaptive model selection for different prediction tasks. By incorporating entropy-weighted ensemble learning from multiple protein language models, Venus-EEM achieves superior performance compared to existing methods. We validate the model's effectiveness through comprehensive evaluation on multiple viral datasets and a detailed case study of T7 RNA polymerase (T7 RNAP) activity. Our findings provide an effective approach for predicting viral protein mutations based on entropy, bridging fundamental physics principles with practical biological challenges.
- Open Access
- 16 December 2024
- Published byĀ Springer NatureĀ inĀ Research Square
Harnessing Protein Language Model for Structure-Based Discovery of Highly Efficient and Robust PET Hydrolases
Plastic waste, particularly polyethylene terephthalate (PET), poses significant environmental challenges, prompting extensive research into enzymatic biodegradation. However, existing PET hydrolases (PETases) are constrained to a narrow sequence space and exhibited limited performance for effective biodegradation. This study introduces a protein discovery pipeline, ProMine, which integrates protein language models (PLMs) with a representation tree to identify PETase based on structural similarity using sequence information. Using the crystal structure of IsPETase as a template, we employed ProMine to search for and cluster target proteins. PETase candidates were further screened using PLM-based assessments of solubility and thermostability, leading to the selection of 34 proteins for biochemical experiments. The results showed that 14 candidates exhibited PET degradation activity across a temperature range of 30-60 ā. Notably, we identified a PET hydrolase from Kibdelosporangium banguiense (KbPETase), which has a melting temperature 32 ā higher than that of IsPETase and exhibits the highest PET degradation activity within 30-60 ā compared to other wild-type PETases. KbPETase also shows higher catalytic efficiency than FastPETase. X-ray crystallography and molecular dynamics simulations revealed that KbPETase has a conserved catalytic domain and enhanced intramolecular interactions, contributing to its improved functionality and thermostability. This work demonstrates a novel deep learning approach for discovering natural PETases with enhanced properties.
- 16 December 2024
- Published byĀ Springer NatureĀ inĀ Research Square
Open AccessHarnessing Protein Language Model for Structure-Based Discovery of Highly Efficient and Robust PET Hydrolases
Plastic waste, particularly polyethylene terephthalate (PET), poses significant environmental challenges, prompting extensive research into enzymatic biodegradation. However, existing PET hydrolases (PETases) are constrained to a narrow sequence space and exhibited limited performance for effective biodegradation. This study introduces a protein discovery pipeline, ProMine, which integrates protein language models (PLMs) with a representation tree to identify PETase based on structural similarity using sequence information. Using the crystal structure of IsPETase as a template, we employed ProMine to search for and cluster target proteins. PETase candidates were further screened using PLM-based assessments of solubility and thermostability, leading to the selection of 34 proteins for biochemical experiments. The results showed that 14 candidates exhibited PET degradation activity across a temperature range of 30-60 ā. Notably, we identified a PET hydrolase from Kibdelosporangium banguiense (KbPETase), which has a melting temperature 32 ā higher than that of IsPETase and exhibits the highest PET degradation activity within 30-60 ā compared to other wild-type PETases. KbPETase also shows higher catalytic efficiency than FastPETase. X-ray crystallography and molecular dynamics simulations revealed that KbPETase has a conserved catalytic domain and enhanced intramolecular interactions, contributing to its improved functionality and thermostability. This work demonstrates a novel deep learning approach for discovering natural PETases with enhanced properties.
- Open Access
- 7 October 2024
- Published byĀ Springer NatureĀ inĀ Nature Communications
- Vol.Ā 15 Ā (1) , 1
Author Correction: Pretrainable geometric graph neural network for antibody affinity maturation
- 7 October 2024
- Published byĀ Springer NatureĀ inĀ Nature Communications
- Vol.Ā 15 Ā (1) , 1
Open AccessAuthor Correction: Pretrainable geometric graph neural network for antibody affinity maturation
- Open Access
- 10 September 2024
- Published byĀ Springer NatureĀ inĀ Cell Discovery
- Vol.Ā 10 Ā (1) , 1-12
A conditional protein diffusion model generates artificial programmable endonuclease sequences with enhanced activity
Deep learning-based methods for generating functional proteins address the growing need for novel biocatalysts, allowing for precise tailoring of functionalities to meet specific requirements. This advancement leads to the development of highly efficient and specialized proteins with diverse applications across scientific, technological, and biomedical fields. This study establishes a pipeline for protein sequence generation with a conditional protein diffusion model, namely CPDiffusion, to create diverse sequences of proteins with enhanced functions. CPDiffusion accommodates protein-specific conditions, such as secondary structures and highly conserved amino acids. Without relying on extensive training data, CPDiffusion effectively captures highly conserved residues and sequence features for specific protein families. We applied CPDiffusion to generate artificial sequences of Argonaute (Ago) proteins based on the backbone structures of wild-type (WT) Kurthia massiliensis Ago (KmAgo) and Pyrococcus furiosus Ago (PfAgo), which are complex multi-domain programmable endonucleases. The generated sequences deviate by up to nearly 400 amino acids from their WT templates. Experimental tests demonstrated that the majority of the generated proteins for both KmAgo and PfAgo show unambiguous activity in DNA cleavage, with many of them exhibiting superior activity as compared to the WT. These findings underscore CPDiffusionās remarkable success rate in generating novel sequences for proteins with complex structures and functions in a single step, leading to enhanced activity. This approach facilitates the design of enzymes with multi-domain molecular structures and intricate functions through in silico generation and screening, all accomplished without the need for supervision from labeled data.
- 10 September 2024
- Published byĀ Springer NatureĀ inĀ Cell Discovery
- Vol.Ā 10 Ā (1) , 1-12
Open AccessA conditional protein diffusion model generates artificial programmable endonuclease sequences with enhanced activity
Deep learning-based methods for generating functional proteins address the growing need for novel biocatalysts, allowing for precise tailoring of functionalities to meet specific requirements. This advancement leads to the development of highly efficient and specialized proteins with diverse applications across scientific, technological, and biomedical fields. This study establishes a pipeline for protein sequence generation with a conditional protein diffusion model, namely CPDiffusion, to create diverse sequences of proteins with enhanced functions. CPDiffusion accommodates protein-specific conditions, such as secondary structures and highly conserved amino acids. Without relying on extensive training data, CPDiffusion effectively captures highly conserved residues and sequence features for specific protein families. We applied CPDiffusion to generate artificial sequences of Argonaute (Ago) proteins based on the backbone structures of wild-type (WT) Kurthia massiliensis Ago (KmAgo) and Pyrococcus furiosus Ago (PfAgo), which are complex multi-domain programmable endonucleases. The generated sequences deviate by up to nearly 400 amino acids from their WT templates. Experimental tests demonstrated that the majority of the generated proteins for both KmAgo and PfAgo show unambiguous activity in DNA cleavage, with many of them exhibiting superior activity as compared to the WT. These findings underscore CPDiffusionās remarkable success rate in generating novel sequences for proteins with complex structures and functions in a single step, leading to enhanced activity. This approach facilitates the design of enzymes with multi-domain molecular structures and intricate functions through in silico generation and screening, all accomplished without the need for supervision from labeled data.
- Open Access
- 6 September 2024
- Published byĀ Springer NatureĀ inĀ Nature Communications
- Vol.Ā 15 Ā (1) , 1-14
Pretrainable geometric graph neural network for antibody affinity maturation
Increasing the binding affinity of an antibody to its target antigen is a crucial task in antibody therapeutics development. This paper presents a pretrainable geometric graph neural network, GearBind, and explores its potential in in silico affinity maturation. Leveraging multi-relational graph construction, multi-level geometric message passing and contrastive pretraining on mass-scale, unlabeled protein structural data, GearBind outperforms previous state-of-the-art approaches on SKEMPI and an independent test set. A powerful ensemble model based on GearBind is then derived and used to successfully enhance the binding of two antibodies with distinct formats and target antigens. ELISA EC50 values of the designed antibody mutants are decreased by up to 17 fold, and KD values by up to 6.1 fold. These promising results underscore the utility of geometric deep learning and effective pretraining in macromolecule interaction modeling tasks.
- 6 September 2024
- Published byĀ Springer NatureĀ inĀ Nature Communications
- Vol.Ā 15 Ā (1) , 1-14
Open AccessPretrainable geometric graph neural network for antibody affinity maturation
Increasing the binding affinity of an antibody to its target antigen is a crucial task in antibody therapeutics development. This paper presents a pretrainable geometric graph neural network, GearBind, and explores its potential in in silico affinity maturation. Leveraging multi-relational graph construction, multi-level geometric message passing and contrastive pretraining on mass-scale, unlabeled protein structural data, GearBind outperforms previous state-of-the-art approaches on SKEMPI and an independent test set. A powerful ensemble model based on GearBind is then derived and used to successfully enhance the binding of two antibodies with distinct formats and target antigens. ELISA EC50 values of the designed antibody mutants are decreased by up to 17 fold, and KD values by up to 6.1 fold. These promising results underscore the utility of geometric deep learning and effective pretraining in macromolecule interaction modeling tasks.
- Open Access
- 15 July 2024
- Published byĀ Springer NatureĀ inĀ Research Square
Precise Generation of Conformational Ensembles for Intrinsically Disordered Proteins via Fine-tuned Diffusion Models
Intrinsically disordered proteins (IDPs) play pivotal roles in various biological functions and are closely linked to many human diseases including cancer, diabetes and Alzheimer disease. Structural investigations of IDPs typically involve a combination of molecular dynamics (MD) simulations and experimental data to correct for intrinsic biases in simulation methods. However, these simulations are hindered by their high computational cost and a scarcity of experimental data, severely limiting their applicability. Despite the recent advancements in structure prediction for structured proteins, understanding the conformational properties of IDPs remains challenging partly due to the poor conservation of disordered protein sequences and limited experimental characterization. Here, we introduce IDPFold, a method capable of generating conformational ensembles for IDPs directly from their sequences using fine-tuned diffusion models. IDPFold bypasses the need for Multiple Sequence Alignments (MSA) or experimental data, achieving accurate predictions of ensemble properties across numerous IDPs. By sampling conformations at the backbone level, IDPFold provides more detailed structural features and more precise property estimation compared to other state-of-the-art methods. IDPFold is ready to be used in the elucidate the sequence-disorder-function paradigm of IDPs.
- 15 July 2024
- Published byĀ Springer NatureĀ inĀ Research Square
Open AccessPrecise Generation of Conformational Ensembles for Intrinsically Disordered Proteins via Fine-tuned Diffusion Models
Intrinsically disordered proteins (IDPs) play pivotal roles in various biological functions and are closely linked to many human diseases including cancer, diabetes and Alzheimer disease. Structural investigations of IDPs typically involve a combination of molecular dynamics (MD) simulations and experimental data to correct for intrinsic biases in simulation methods. However, these simulations are hindered by their high computational cost and a scarcity of experimental data, severely limiting their applicability. Despite the recent advancements in structure prediction for structured proteins, understanding the conformational properties of IDPs remains challenging partly due to the poor conservation of disordered protein sequences and limited experimental characterization. Here, we introduce IDPFold, a method capable of generating conformational ensembles for IDPs directly from their sequences using fine-tuned diffusion models. IDPFold bypasses the need for Multiple Sequence Alignments (MSA) or experimental data, achieving accurate predictions of ensemble properties across numerous IDPs. By sampling conformations at the backbone level, IDPFold provides more detailed structural features and more precise property estimation compared to other state-of-the-art methods. IDPFold is ready to be used in the elucidate the sequence-disorder-function paradigm of IDPs.