Chun Fang
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
Chun Fang has a total of 21 co-authors. Here are the top co-authors by number of shared publications:
- Fan Zhang
Beijing Institute of Petrochemical Technology, Beijing, 102617, China
4 shared publications - Jiasheng He
Department of Information Engineering, Beijing Institute of Petrochemical Technology, 19 Qingyuan North Road, Daxing District, Beijing 102617, P. R. China
4 shared publications - Shun Zhang
Beijing Institute of Petrochemical Technology, China
4 shared publications Zhenguo Wen
School of Information Engineering, Beijing Institute of Petrochemical Technology, Beijing, China
3 shared publicationsFan Mo
Faculty of Science and Engineering, Waseda University, Tokyo, Japan
2 shared publications- Hayato Yamana
Waseda University
2 shared publications
Latest Publications
- 8 July 2026
- Published by Springer Nature in Interdisciplinary Sciences: Computational Life Sciences
- p. 1-13
Identifying RNA ac4C Modification Sites via Pseudo-Nucleotide Fingerprint Encoding and Multi-Scale Feature Integration
RNA modifications, particularly N4-acetylcytidine (ac4C), play essential roles in gene regulation by influencing mRNA stability, translational efficiency, and cellular responses to stress. Aberrant ac4C modifications have been implicated in tumorigenesis and cancer progression, highlighting their potential as diagnostic biomarkers and therapeutic targets. However, existing computational approaches for predicting RNA modification sites often rely on fixed-scale features or shallow learning architectures, limiting their ability to model multiscale nucleotide interactions and long-range dependencies. To address these limitations, we introduce DFM-ac4C, a novel computational framework designed for the accurate prediction of ac4C modification sites. DFM-ac4C integrates pseudo-nucleotide fingerprint features derived from Kidera factors with global contextual embeddings generated by the Evo2 and DNAshape pretrained models. Leveraging an attention-based fusion architecture, the framework dynamically combines local sequence information, fingerprint descriptors, and global embeddings, enabling it to effectively capture the complex, multiscale characteristics of nucleotide interaction sites. Extensive performance evaluations demonstrate that DFM-ac4C significantly outperforms existing models, achieving outstanding predictive metrics, including an AUC of 96.76%, accuracy (ACC) of 90.49%, Matthew’s correlation coefficient (MCC) of 0.8098, sensitivity (SEN) of 90.22%, and specificity (SPE) of 90.76%. These results underscore DFM-ac4C’s effectiveness as a robust and efficient tool for RNA ac4C site identification. The complete dataset and source code supporting this study can be accessed freely from: https://github.com/BioAI-wym/DFM-ac4C.
- 8 July 2026
- Published by Springer Nature in Interdisciplinary Sciences: Computational Life Sciences
- p. 1-13
Identifying RNA ac4C Modification Sites via Pseudo-Nucleotide Fingerprint Encoding and Multi-Scale Feature Integration
RNA modifications, particularly N4-acetylcytidine (ac4C), play essential roles in gene regulation by influencing mRNA stability, translational efficiency, and cellular responses to stress. Aberrant ac4C modifications have been implicated in tumorigenesis and cancer progression, highlighting their potential as diagnostic biomarkers and therapeutic targets. However, existing computational approaches for predicting RNA modification sites often rely on fixed-scale features or shallow learning architectures, limiting their ability to model multiscale nucleotide interactions and long-range dependencies. To address these limitations, we introduce DFM-ac4C, a novel computational framework designed for the accurate prediction of ac4C modification sites. DFM-ac4C integrates pseudo-nucleotide fingerprint features derived from Kidera factors with global contextual embeddings generated by the Evo2 and DNAshape pretrained models. Leveraging an attention-based fusion architecture, the framework dynamically combines local sequence information, fingerprint descriptors, and global embeddings, enabling it to effectively capture the complex, multiscale characteristics of nucleotide interaction sites. Extensive performance evaluations demonstrate that DFM-ac4C significantly outperforms existing models, achieving outstanding predictive metrics, including an AUC of 96.76%, accuracy (ACC) of 90.49%, Matthew’s correlation coefficient (MCC) of 0.8098, sensitivity (SEN) of 90.22%, and specificity (SPE) of 90.76%. These results underscore DFM-ac4C’s effectiveness as a robust and efficient tool for RNA ac4C site identification. The complete dataset and source code supporting this study can be accessed freely from: https://github.com/BioAI-wym/DFM-ac4C.
- 24 October 2025
- Published by Springer Nature in Journal of Computer-Aided Molecular Design
- Vol. 39 (2) , 1-15
Kideraspa: designing variants of staphylococcal protein a based on a diffusion model with kidera factors
The interaction between staphylococcal protein A (SpA) and human immunoglobulin G (IgG) is pivotal in treating diseases such as cancer, inflammation, infections, and autoimmune disorders. However, acquiring natural SpA variants is labor-intensive, traditional protein design methods often depend on extensive datasets and detailed structural information, limiting their efficiency and applicability. To overcome these limitations, we propose a deep learning-based approach that directly targets desired binding functions by introducing mutations at selected SpA sites to optimize its properties. Specifically, we present a de novo protein design strategy that integrates a diffusion-based generative model with Kidera factor representations to create SpA variants. The framework comprises three modules: sequence generation, where protein sequences are encoded via Kidera factors and novel variants are generated using a diffusion model; computational screening, employing tools like AlphaFold3 to assess structural properties, solubility, and physicochemical characteristics, thereby selecting high-potential candidates; and experimental validation, involving wet-lab experiments to evaluate the biological activities and binding affinities of the designed proteins. The generated SpA variants demonstrated high success rates and strong binding affinities toward IgG. These findings confirm the effectiveness of our method in producing functional proteins comparable to natural counterparts, offering a scalable and data-efficient alternative to protein engineering.
- 24 October 2025
- Published by Springer Nature in Journal of Computer-Aided Molecular Design
- Vol. 39 (2) , 1-15
Kideraspa: designing variants of staphylococcal protein a based on a diffusion model with kidera factors
The interaction between staphylococcal protein A (SpA) and human immunoglobulin G (IgG) is pivotal in treating diseases such as cancer, inflammation, infections, and autoimmune disorders. However, acquiring natural SpA variants is labor-intensive, traditional protein design methods often depend on extensive datasets and detailed structural information, limiting their efficiency and applicability. To overcome these limitations, we propose a deep learning-based approach that directly targets desired binding functions by introducing mutations at selected SpA sites to optimize its properties. Specifically, we present a de novo protein design strategy that integrates a diffusion-based generative model with Kidera factor representations to create SpA variants. The framework comprises three modules: sequence generation, where protein sequences are encoded via Kidera factors and novel variants are generated using a diffusion model; computational screening, employing tools like AlphaFold3 to assess structural properties, solubility, and physicochemical characteristics, thereby selecting high-potential candidates; and experimental validation, involving wet-lab experiments to evaluate the biological activities and binding affinities of the designed proteins. The generated SpA variants demonstrated high success rates and strong binding affinities toward IgG. These findings confirm the effectiveness of our method in producing functional proteins comparable to natural counterparts, offering a scalable and data-efficient alternative to protein engineering.
- 4 September 2025
- Published by Springer Nature in Journal of Computer-Aided Molecular Design
- Vol. 39 (1) , 1-10
Protein A-like peptide generation based on generalized diffusion model
With the rapid advancement of biotechnology, protein generation and design based on generative models have demonstrated extensive applications in drug development, vaccine research, and biocatalysis. This research proposes a protein generation method based on the generalized diffusion model, which breaks through the traditional diffusion model’s reliance on Gaussian noise, enables more flexible protein sequence generation, and preliminarily verifies its advantages. Specifically, protein sequences were first encoded using one-hot encoding and input into the diffusion model to generate novel sequences. Subsequently, the tertiary structures of the generated proteins were predicted using AlphaFold, followed by structural alignment and backbone distance calculation via PyMOL to select the optimal sequences. The predicted derivative protein sequence A_005 was screened from the generated sequences and subjected to an affinity assay with Protein A parental. Experimental results revealed that A_005 exhibited remarkably high affinity, as well as a satisfactory dissociation rate and association rate. The findings demonstrate that the protein generation method based on the generalized diffusion model can effectively design protein sequences with high structural and functional similarity to target sequences. While prior studies have shown that both DDPM and generalized diffusion models achieve high generation quality, the generalized diffusion model outperforms in terms of task adaptability. Our research not only opens new technological pathways for protein design but also lays a solid foundation for future applications in biomedicine, providing significant theoretical and experimental evidence for subsequent drug development.
- 4 September 2025
- Published by Springer Nature in Journal of Computer-Aided Molecular Design
- Vol. 39 (1) , 1-10
Protein A-like peptide generation based on generalized diffusion model
With the rapid advancement of biotechnology, protein generation and design based on generative models have demonstrated extensive applications in drug development, vaccine research, and biocatalysis. This research proposes a protein generation method based on the generalized diffusion model, which breaks through the traditional diffusion model’s reliance on Gaussian noise, enables more flexible protein sequence generation, and preliminarily verifies its advantages. Specifically, protein sequences were first encoded using one-hot encoding and input into the diffusion model to generate novel sequences. Subsequently, the tertiary structures of the generated proteins were predicted using AlphaFold, followed by structural alignment and backbone distance calculation via PyMOL to select the optimal sequences. The predicted derivative protein sequence A_005 was screened from the generated sequences and subjected to an affinity assay with Protein A parental. Experimental results revealed that A_005 exhibited remarkably high affinity, as well as a satisfactory dissociation rate and association rate. The findings demonstrate that the protein generation method based on the generalized diffusion model can effectively design protein sequences with high structural and functional similarity to target sequences. While prior studies have shown that both DDPM and generalized diffusion models achieve high generation quality, the generalized diffusion model outperforms in terms of task adaptability. Our research not only opens new technological pathways for protein design but also lays a solid foundation for future applications in biomedicine, providing significant theoretical and experimental evidence for subsequent drug development.
- 8 May 2025
- Published by Elsevier in Analytical Biochemistry
- Vol. 704, 115881
AFP-MCDF: Multi and cross-dimensional feature fusion methods for antifreeze protein prediction
- 8 May 2025
- Published by Elsevier in Analytical Biochemistry
- Vol. 704, 115881
AFP-MCDF: Multi and cross-dimensional feature fusion methods for antifreeze protein prediction
- 24 February 2025
- Published by Elsevier in Journal of Molecular Graphics and Modelling
- Vol. 137, 108982
DeepTree-AAPred: Binary tree-based deep learning model for anti-angiogenic peptides prediction
- 24 February 2025
- Published by Elsevier in Journal of Molecular Graphics and Modelling
- Vol. 137, 108982
DeepTree-AAPred: Binary tree-based deep learning model for anti-angiogenic peptides prediction
- 16 October 2024
- Published by Elsevier in Computational Biology and Chemistry
- Vol. 113, 108240
Cycle-ESM: Generation-assisted classification of antifungal peptides using ESM protein language model
- 16 October 2024
- Published by Elsevier in Computational Biology and Chemistry
- Vol. 113, 108240
Cycle-ESM: Generation-assisted classification of antifungal peptides using ESM protein language model
- 17 May 2024
- Published by Elsevier in Computational Biology and Chemistry
- Vol. 111, 108098
FusPB-ESM2: Fusion model of ProtBERT and ESM-2 for cell-penetrating peptide prediction
- 17 May 2024
- Published by Elsevier in Computational Biology and Chemistry
- Vol. 111, 108098
FusPB-ESM2: Fusion model of ProtBERT and ESM-2 for cell-penetrating peptide prediction
- 1 April 2024
- Published by World Scientific Pub Co Pte Ltd in Journal of Bioinformatics and Computational Biology
- Vol. 22 (02) , 2450006
MoRF_ESM: Prediction of MoRFs in disordered proteins based on a deep transformer protein language model
Molecular recognition features (MoRFs) are particular functional segments of disordered proteins, which play crucial roles in regulating the phase transition of membrane-less organelles and frequently serve as central sites in cellular interaction networks. As the association between disordered proteins and severe diseases continues to be discovered, identifying MoRFs has gained growing significance. Due to the limited number of experimentally validated MoRFs, the performance of existing MoRF’s prediction algorithms is not good enough and still needs to be improved. In this research, we present a model named MoRF_ESM, which utilizes deep-learning protein representations to predict MoRFs in disordered proteins. This approach employs a pretrained ESM-2 protein language model to generate embedding representations of residues in the form of attention map matrices. These representations are combined with a self-learned TextCNN model for feature extraction and prediction. In addition, an averaging step was incorporated at the end of the MoRF_ESM model to refine the output and generate final prediction results. In comparison to other impressive methods on benchmark datasets, the MoRF_ESM approach demonstrates state-of-the-art performance, achieving [ ] higher AUC than other methods when tested on TEST1 and achieving [ ] higher AUC than other methods when tested on TEST2. These results imply that the combination of ESM-2 and TextCNN can effectively extract deep evolutionary features related to protein structure and function, along with capturing shallow pattern features located in protein sequences, and is well qualified for the prediction task of MoRFs. Given that ESM-2 is a highly versatile protein language model, the methodology proposed in this study can be readily applied to other tasks involving the classification of protein sequences.
- 1 April 2024
- Published by World Scientific Pub Co Pte Ltd in Journal of Bioinformatics and Computational Biology
- Vol. 22 (02) , 2450006
MoRF_ESM: Prediction of MoRFs in disordered proteins based on a deep transformer protein language model
Molecular recognition features (MoRFs) are particular functional segments of disordered proteins, which play crucial roles in regulating the phase transition of membrane-less organelles and frequently serve as central sites in cellular interaction networks. As the association between disordered proteins and severe diseases continues to be discovered, identifying MoRFs has gained growing significance. Due to the limited number of experimentally validated MoRFs, the performance of existing MoRF’s prediction algorithms is not good enough and still needs to be improved. In this research, we present a model named MoRF_ESM, which utilizes deep-learning protein representations to predict MoRFs in disordered proteins. This approach employs a pretrained ESM-2 protein language model to generate embedding representations of residues in the form of attention map matrices. These representations are combined with a self-learned TextCNN model for feature extraction and prediction. In addition, an averaging step was incorporated at the end of the MoRF_ESM model to refine the output and generate final prediction results. In comparison to other impressive methods on benchmark datasets, the MoRF_ESM approach demonstrates state-of-the-art performance, achieving [ ] higher AUC than other methods when tested on TEST1 and achieving [ ] higher AUC than other methods when tested on TEST2. These results imply that the combination of ESM-2 and TextCNN can effectively extract deep evolutionary features related to protein structure and function, along with capturing shallow pattern features located in protein sequences, and is well qualified for the prediction task of MoRFs. Given that ESM-2 is a highly versatile protein language model, the methodology proposed in this study can be readily applied to other tasks involving the classification of protein sequences.
- Open Access
- 26 January 2024
- Published by Association for Computing Machinery (ACM)
- p. 375-379
Prediction of DNA enhancers based on multi-species genomic base model DNABERT-2 and BiGRU network
DNA enhancers are important regulatory elements in the genome, and accurate prediction is crucial for understanding gene regulation. In this study, a DNA enhancer prediction algorithm, DNABERT-Enhancer, is proposed by combining the DNABERT-2 model and the BiGRU network. The algorithm analogizes DNA sequences to sentences, and strengthens the capture of contextual information of DNA sequences by integrating the structural strengths of the DNABERT-2 pre-training model and the BiGRU network, thus realizing the DNA enhancers prediction. The experimental results show an AUC value of 0.943 on the test set. This study provides a new method for DNA enhancer prediction in the field of bioinformatics, which is expected to enhance our understanding of gene regulation in practical applications, for example, it may be used to more accurately identify disease markers, customize personalized medical treatment plans or finding new targets in drug discovery that advances the field of genomics.
- 26 January 2024
- Published by Association for Computing Machinery (ACM)
- p. 375-379
Open AccessPrediction of DNA enhancers based on multi-species genomic base model DNABERT-2 and BiGRU network
DNA enhancers are important regulatory elements in the genome, and accurate prediction is crucial for understanding gene regulation. In this study, a DNA enhancer prediction algorithm, DNABERT-Enhancer, is proposed by combining the DNABERT-2 model and the BiGRU network. The algorithm analogizes DNA sequences to sentences, and strengthens the capture of contextual information of DNA sequences by integrating the structural strengths of the DNABERT-2 pre-training model and the BiGRU network, thus realizing the DNA enhancers prediction. The experimental results show an AUC value of 0.943 on the test set. This study provides a new method for DNA enhancer prediction in the field of bioinformatics, which is expected to enhance our understanding of gene regulation in practical applications, for example, it may be used to more accurately identify disease markers, customize personalized medical treatment plans or finding new targets in drug discovery that advances the field of genomics.
- Open Access
- 26 January 2024
- Published by Association for Computing Machinery (ACM)
- p. 7-11
Deep Learning Neural Network-Based Weibo Sentiment Analysis
Sentiment analysis plays a crucial role in current social sentiment research, as it effectively reflects the dynamics of public sentiment through the analysis of Weibo content. However, traditional sentiment analysis models are constrained by challenges in context integration and feature sparsity, which limit their accuracy. Deep learning models such as TextCNN and BiLSTM have made some progress in sentiment analysis, but they still face issues related to long-range dependencies and fine-grained correlations, which constrain their performance.To address these challenges, this study introduces self-attention mechanisms and incorporates them into TextCNN and BiLSTM models. Experimental results demonstrate that the combination of BiLSTM with self-attention mechanism excels in Weibo sentiment analysis, achieving an accuracy rate of 90.35% on the test set. This research provides an efficient deep learning approach for Weibo sentiment analysis, with the potential to play a significant role in sentiment research
- 26 January 2024
- Published by Association for Computing Machinery (ACM)
- p. 7-11
Open AccessDeep Learning Neural Network-Based Weibo Sentiment Analysis
Sentiment analysis plays a crucial role in current social sentiment research, as it effectively reflects the dynamics of public sentiment through the analysis of Weibo content. However, traditional sentiment analysis models are constrained by challenges in context integration and feature sparsity, which limit their accuracy. Deep learning models such as TextCNN and BiLSTM have made some progress in sentiment analysis, but they still face issues related to long-range dependencies and fine-grained correlations, which constrain their performance.To address these challenges, this study introduces self-attention mechanisms and incorporates them into TextCNN and BiLSTM models. Experimental results demonstrate that the combination of BiLSTM with self-attention mechanism excels in Weibo sentiment analysis, achieving an accuracy rate of 90.35% on the test set. This research provides an efficient deep learning approach for Weibo sentiment analysis, with the potential to play a significant role in sentiment research