Changhui Yan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Changhui Yan has a total of 30 co-authors. Here are the top co-authors by number of shared publications:
- Noreen Fayyaz Khan
Department of Computer Science, North Dakota State University, Fargo, ND, USA
2 shared publications - Caroline S. Moffat
Centre for Crop Disease and Management, School of Molecular and Life Sciences, Curtin University, Bentley, Western Australia, Australia
1 shared publication Deewan Bajracharya
1 shared publicationGayan K. Kariyawasam
Department of Plant Pathology, North Dakota State University, Fargo, ND 58108, USA
1 shared publicationGongjun Shi
Department of Plant Pathology, North Dakota State University, Fargo, ND 58108, USA
1 shared publication- Guoqing Gui
School of Architecture and Civil Engineering, Jinggangshan University, Ji’an, Jiangxi, China
1 shared publication
Latest Publications
- 29 May 2026
- Published by Springer Nature
- p. 28-40
A Structure-Aware Attention Mechanism for Protein Function Prediction
- 29 May 2026
- Published by Springer Nature
- p. 28-40
A Structure-Aware Attention Mechanism for Protein Function Prediction
- Open Access
- 23 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of Engineering in Medicine and Biology
- Vol. 7, 70-78
Leveraging Vision Transformers for High-Precision Classification of Cancer Cell Cultures: A Comparative Study on MDA-MB-231 and PC3 Datasets
Goal: Accurate classification of cancer cell cultures is critical for understanding tumor behavior, drug responses, and disease progression. Traditional manual evaluation methods are often subjective and prone to errors, necessitating automated approaches based on deep learning. Methods: We design a pipeline using Otsu thresholding, morphological filtering, and watershed segmentation, coupled with class-balanced augmentation. A baseline CNN, two attention-augmented variants: CNN-SE (squeeze-and-excitation) and CNN-CBAM (channel-spatial attention) and a ViT are benchmarked. All models are tuned through a 5-fold cross-validation grid search and evaluated on MDA-MB-231 (triple negative breast cancer) and PC3 (prostate cancer) cell images. The study employed both quantitative and qualitative approaches to comprehensively assess the effectiveness and reliability of the proposed model. Results: Attention modules substantially strengthen CNN performance but ViT achieves the best overall accuracy and generalization, particularly reflecting an advantage in modeling long-range dependencies. Conclusions: Channel and spatial attention narrows the gap between CNNs and transformers, while transformers provide the highest end-to-end performance. These results support attention-augmented CNNs and ViTs as robust complementary tools for automated analysis of cancer cell cultures.
- 23 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of Engineering in Medicine and Biology
- Vol. 7, 70-78
Open AccessLeveraging Vision Transformers for High-Precision Classification of Cancer Cell Cultures: A Comparative Study on MDA-MB-231 and PC3 Datasets
Goal: Accurate classification of cancer cell cultures is critical for understanding tumor behavior, drug responses, and disease progression. Traditional manual evaluation methods are often subjective and prone to errors, necessitating automated approaches based on deep learning. Methods: We design a pipeline using Otsu thresholding, morphological filtering, and watershed segmentation, coupled with class-balanced augmentation. A baseline CNN, two attention-augmented variants: CNN-SE (squeeze-and-excitation) and CNN-CBAM (channel-spatial attention) and a ViT are benchmarked. All models are tuned through a 5-fold cross-validation grid search and evaluated on MDA-MB-231 (triple negative breast cancer) and PC3 (prostate cancer) cell images. The study employed both quantitative and qualitative approaches to comprehensively assess the effectiveness and reliability of the proposed model. Results: Attention modules substantially strengthen CNN performance but ViT achieves the best overall accuracy and generalization, particularly reflecting an advantage in modeling long-range dependencies. Conclusions: Channel and spatial attention narrows the gap between CNNs and transformers, while transformers provide the highest end-to-end performance. These results support attention-augmented CNNs and ViTs as robust complementary tools for automated analysis of cancer cell cultures.
- Open Access
VirusT5: Harnessing Large Language Models to Predicting SARS-CoV-2 Evolution
During a virus's evolution,various regions of the genome are subjected to distinct levels of functional constraints.Combined with factors like codon bias and DNA repair efficiency,these constraints contribute to unique mutation patterns within the genome or a specific gene. In this project, we harnessed the power of Large Language Models(LLMs) to predict the evolution of SARS-CoV-2. By treating the mutation process from one generation to the next as a translation task, we trained a transformer model, called VirusT5, to capture the mutation patterns underlying SARS-CoV-2 evolution. We evaluated the VirusT5's ability to detect these mutation patterns including its ability to identify mutation hotspots and explored the potential of using VirusT5 to predict future virus variants. Our findings demonstrate the feasibility of using a large language model to model viral evolution as a translation process. This study establishes the groundbreaking concept of "mutation-as-translation," paving the way for new methodologies and tools for combating virus threats
Open AccessVirusT5: Harnessing Large Language Models to Predicting SARS-CoV-2 Evolution
During a virus's evolution,various regions of the genome are subjected to distinct levels of functional constraints.Combined with factors like codon bias and DNA repair efficiency,these constraints contribute to unique mutation patterns within the genome or a specific gene. In this project, we harnessed the power of Large Language Models(LLMs) to predict the evolution of SARS-CoV-2. By treating the mutation process from one generation to the next as a translation task, we trained a transformer model, called VirusT5, to capture the mutation patterns underlying SARS-CoV-2 evolution. We evaluated the VirusT5's ability to detect these mutation patterns including its ability to identify mutation hotspots and explored the potential of using VirusT5 to predict future virus variants. Our findings demonstrate the feasibility of using a large language model to model viral evolution as a translation process. This study establishes the groundbreaking concept of "mutation-as-translation," paving the way for new methodologies and tools for combating virus threats
- Open Access
- 14 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 154080-154091
Detection of Violent Scenes in Cartoon Movies Using a Deep Learning Approach
Cartoon movies are a primary source of entertainment for children. However, concerns arise when these movies inadvertently expose children to violent scenes. This paper addresses the challenge of detecting subtle instances of violence within cartoon movies. The main difficulty in this task is the sequential nature of movies, where a sequence of frames must be considered in their given order. Existing methods have not effectively addressed the issue. In this study, we tackled this challenge by employing a sequential model. The research comprises three key steps. Initially, a histogram technique was implemented to select pivotal frames from the video sequences. Subsequently, convolutional neural networks (CNNs) were utilized to extract prominent features from these selected key frames. In the third phase, the acquired features were utilized to train a sequential model using sequence-based learning. The model was then refined through transfer learning, using a dataset containing scenes devoid of violence, as well as scenes depicting varying forms of violence, including bloodshed, fights, gunshots, and explosions. A significant contribution of this study is the meticulous categorization of violent scenes into four distinct types, allowing for further investigation into the diverse effects of different violence categories. Furthermore, the study introduces an innovative approach by integrating a dense layer into the sequential model to enhance final classification. The trained model’s performance was comprehensively evaluated using metrics such as F1 score, precision, accuracy, and recall. To validate the effectiveness of the proposed model, it was benchmarked against state-of-the-art methods. This study presents an innovative deep-learning methodology for the identification of violent scenes in cartoon movies. Its potential applications encompass a wide range, including safeguarding children from inappropriate content.
- 14 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 154080-154091
Open AccessDetection of Violent Scenes in Cartoon Movies Using a Deep Learning Approach
Cartoon movies are a primary source of entertainment for children. However, concerns arise when these movies inadvertently expose children to violent scenes. This paper addresses the challenge of detecting subtle instances of violence within cartoon movies. The main difficulty in this task is the sequential nature of movies, where a sequence of frames must be considered in their given order. Existing methods have not effectively addressed the issue. In this study, we tackled this challenge by employing a sequential model. The research comprises three key steps. Initially, a histogram technique was implemented to select pivotal frames from the video sequences. Subsequently, convolutional neural networks (CNNs) were utilized to extract prominent features from these selected key frames. In the third phase, the acquired features were utilized to train a sequential model using sequence-based learning. The model was then refined through transfer learning, using a dataset containing scenes devoid of violence, as well as scenes depicting varying forms of violence, including bloodshed, fights, gunshots, and explosions. A significant contribution of this study is the meticulous categorization of violent scenes into four distinct types, allowing for further investigation into the diverse effects of different violence categories. Furthermore, the study introduces an innovative approach by integrating a dense layer into the sequential model to enhance final classification. The trained model’s performance was comprehensively evaluated using metrics such as F1 score, precision, accuracy, and recall. To validate the effectiveness of the proposed model, it was benchmarked against state-of-the-art methods. This study presents an innovative deep-learning methodology for the identification of violent scenes in cartoon movies. Its potential applications encompass a wide range, including safeguarding children from inappropriate content.
- 1 July 2021
- Published by Elsevier in Fungal Genetics and Biology
- Vol. 152, 103571
A genome-wide genetic linkage map and reference quality genome sequence for a new race in the wheat pathogen Pyrenophora tritici-repentis
- 1 July 2021
- Published by Elsevier in Fungal Genetics and Biology
- Vol. 152, 103571
A genome-wide genetic linkage map and reference quality genome sequence for a new race in the wheat pathogen Pyrenophora tritici-repentis
- Open Access
The selection of reference genome and the search for the origin of SARS-CoV-2
S: The pandemic caused by SARS-CoV-2 has a great impact on the whole world. In a theory of the origin of SARS-CoV-2, pangolins were considered a potential intermediate host. To assemble the coronavirus found in pangolins, SARS-CoV-2 were used a reference genome in most of studies, assuming that pangolins CoV and SARS-CoV-2 are the closest neighbors in the evolution. However, this assumption may not be true. We investigated how the selection of reference genome affect the resulting CoV genome assembly. We explored various representative CoV as reference genome, and found significant differences in the resulting assemblies. The assembly obtained using RaTG13 as reference showed better statistics in total length and N50 than the assembly guided by SARS-CoV-2, indicating that RaTG13 maybe a better reference for assembling CoV in pangolin or other potential intermediate hosts.
Open AccessThe selection of reference genome and the search for the origin of SARS-CoV-2
S: The pandemic caused by SARS-CoV-2 has a great impact on the whole world. In a theory of the origin of SARS-CoV-2, pangolins were considered a potential intermediate host. To assemble the coronavirus found in pangolins, SARS-CoV-2 were used a reference genome in most of studies, assuming that pangolins CoV and SARS-CoV-2 are the closest neighbors in the evolution. However, this assumption may not be true. We investigated how the selection of reference genome affect the resulting CoV genome assembly. We explored various representative CoV as reference genome, and found significant differences in the resulting assemblies. The assembly obtained using RaTG13 as reference showed better statistics in total length and N50 than the assembly guided by SARS-CoV-2, indicating that RaTG13 maybe a better reference for assembling CoV in pangolin or other potential intermediate hosts.
- Open Access
- 1 March 2018
- Published by Elsevier in KSCE Journal of Civil Engineering
- Vol. 22 (3) , 928-940
Deep BBN Learning for Health Assessment toward Decision-Making on Structures under Uncertainties
- 1 March 2018
- Published by Elsevier in KSCE Journal of Civil Engineering
- Vol. 22 (3) , 928-940
Open AccessDeep BBN Learning for Health Assessment toward Decision-Making on Structures under Uncertainties
- Open Access
- 1 January 2017
- Published by SAGE Publications in Journal of Computational Biology
- Vol. 24 (1) , 31-39
A Graph Approach to Mining Biological Patterns in the Binding Interfaces
Protein–RNA interactions play important roles in the biological systems. Searching for regular patterns in the Protein–RNA binding interfaces is important for understanding how protein and RNA recognize each other and bind to form a complex. Herein, we present a graph-mining method for discovering biological patterns in the protein–RNA interfaces. We represented known protein–RNA interfaces using graphs and then discovered graph patterns enriched in the interfaces. Comparison of the discovered graph patterns with UniProt annotations showed that the graph patterns had a significant overlap with residue sites that had been proven crucial for the RNA binding by experimental methods. Using 200 patterns as input features, a support vector machine method was able to classify protein surface patches into RNA-binding sites and non-RNA-binding sites with 84.0% accuracy and 88.9% precision. We built a simple scoring function that calculated the total number of the graph patterns that occurred in a protein–RNA interface. That scoring function was able to discriminate near-native protein–RNA complexes from docking decoys with a performance comparable with that of a state-of-the-art complex scoring function. Our work also revealed possible patterns that might be important for binding affinity.
- 1 January 2017
- Published by SAGE Publications in Journal of Computational Biology
- Vol. 24 (1) , 31-39
Open AccessA Graph Approach to Mining Biological Patterns in the Binding Interfaces
Protein–RNA interactions play important roles in the biological systems. Searching for regular patterns in the Protein–RNA binding interfaces is important for understanding how protein and RNA recognize each other and bind to form a complex. Herein, we present a graph-mining method for discovering biological patterns in the protein–RNA interfaces. We represented known protein–RNA interfaces using graphs and then discovered graph patterns enriched in the interfaces. Comparison of the discovered graph patterns with UniProt annotations showed that the graph patterns had a significant overlap with residue sites that had been proven crucial for the RNA binding by experimental methods. Using 200 patterns as input features, a support vector machine method was able to classify protein surface patches into RNA-binding sites and non-RNA-binding sites with 84.0% accuracy and 88.9% precision. We built a simple scoring function that calculated the total number of the graph patterns that occurred in a protein–RNA interface. That scoring function was able to discriminate near-native protein–RNA complexes from docking decoys with a performance comparable with that of a state-of-the-art complex scoring function. Our work also revealed possible patterns that might be important for binding affinity.
- Open Access
- 1 January 2014
- Published by Springer Nature in BMC Systems Biology
- Vol. 8 (Suppl 4) , S10
A graph kernel method for DNA-binding site prediction
Protein-DNA interactions play important roles in many biological processes. Computational methods that can accurately predict DNA-binding sites on proteins will greatly expedite research on problems involving protein-DNA interactions.
- 1 January 2014
- Published by Springer Nature in BMC Systems Biology
- Vol. 8 (Suppl 4) , S10
Open AccessA graph kernel method for DNA-binding site prediction
Protein-DNA interactions play important roles in many biological processes. Computational methods that can accurately predict DNA-binding sites on proteins will greatly expedite research on problems involving protein-DNA interactions.
- Open Access
- 29 July 2011
- Published by Springer Nature in Journal of Biomedical Semantics
- Vol. 2 (1) , 3
A shortest-path graph kernel for estimating gene product semantic similarity
Background: Existing methods for calculating semantic similarity between gene products using the Gene Ontology (GO) often rely on external resources, which are not part of the ontology. Consequently, changes in these external resources like biased term distribution caused by shifting of hot research topics, will affect the calculation of semantic similarity. One way to avoid this problem is to use semantic methods that are "intrinsic" to the ontology, i.e. independent of external knowledge. Results: We present a shortest-path graph kernel (spgk) method that relies exclusively on the GO and its structure. In spgk, a gene product is represented by an induced subgraph of the GO, which consists of all the GO terms annotating it. Then a shortest-path graph kernel is used to compute the similarity between two graphs. In a comprehensive evaluation using a benchmark dataset, spgk compares favorably with other methods that depend on external resources. Compared with simUI, a method that is also intrinsic to GO, spgk achieves slightly better results on the benchmark dataset. Statistical tests show that the improvement is significant when the resolution and EC similarity correlation coefficient are used to measure the performance, but is insignificant when the Pfam similarity correlation coefficient is used. Conclusions: Spgk uses a graph kernel method in polynomial time to exploit the structure of the GO to calculate semantic similarity between gene products. It provides an alternative to both methods that use external resources and "intrinsic" methods with comparable performance.
- 29 July 2011
- Published by Springer Nature in Journal of Biomedical Semantics
- Vol. 2 (1) , 3
Open AccessA shortest-path graph kernel for estimating gene product semantic similarity
Background: Existing methods for calculating semantic similarity between gene products using the Gene Ontology (GO) often rely on external resources, which are not part of the ontology. Consequently, changes in these external resources like biased term distribution caused by shifting of hot research topics, will affect the calculation of semantic similarity. One way to avoid this problem is to use semantic methods that are "intrinsic" to the ontology, i.e. independent of external knowledge. Results: We present a shortest-path graph kernel (spgk) method that relies exclusively on the GO and its structure. In spgk, a gene product is represented by an induced subgraph of the GO, which consists of all the GO terms annotating it. Then a shortest-path graph kernel is used to compute the similarity between two graphs. In a comprehensive evaluation using a benchmark dataset, spgk compares favorably with other methods that depend on external resources. Compared with simUI, a method that is also intrinsic to GO, spgk achieves slightly better results on the benchmark dataset. Statistical tests show that the improvement is significant when the resolution and EC similarity correlation coefficient are used to measure the performance, but is insignificant when the Pfam similarity correlation coefficient is used. Conclusions: Spgk uses a graph kernel method in polynomial time to exploit the structure of the GO to calculate semantic similarity between gene products. It provides an alternative to both methods that use external resources and "intrinsic" methods with comparable performance.