Dongsheng Zhao
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
Dongsheng Zhao has a total of 136 co-authors. Here are the top co-authors by number of shared publications:
Fan Tong
Academy of Military Medical Sciences, Beijing, China
6 shared publicationsXiangwen Zheng
Academy of Military Medical Sciences, Beijing, China
6 shared publications- Haifeng Xu
Information Center, Academy of Military Medical Sciences, Beijing, China
5 shared publications - Xiaochen Bo
Academy of Military Medical Sciences, Beijing 100850, China
5 shared publications Xiaowei Luo
Academy of Military Medical Sciences, Beijing, China
5 shared publications- Jianfei Pang
Information Center, Academy of Military Medical Sciences, Beijing, China
4 shared publications
Latest Publications
- Open Access
- 22 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 364-374
TCM-KSS: A System for Traditional Chinese Medicine Knowledge Services
- 22 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 364-374
Open AccessTCM-KSS: A System for Traditional Chinese Medicine Knowledge Services
- Top Cited PapersOpen Access
- 26 July 2025
- Published by Springer Nature in Nature Communications
- Vol. 16 (1) , 1-16
Evidential deep learning-based drug-target interaction prediction
Drug-target interaction (DTI) prediction is a crucial component of drug discovery. Recent deep learning methods show great potential in this field but also encounter substantial challenges. These include generating reliable confidence estimates for predictions, enhancing robustness when handling novel, unseen DTIs, and mitigating the tendency toward overconfident and incorrect predictions. To solve these problems, we propose EviDTI, a novel approach utilizing evidential deep learning (EDL) for uncertainty quantification in neural network-based DTI prediction. EviDTI integrates multiple data dimensions, including drug 2D topological graphs and 3D spatial structures, and target sequence features. Through EDL, EviDTI provides uncertainty estimates for its predictions. Experimental results on three benchmark datasets demonstrate the competitiveness of EviDTI against 11 baseline models. In addition, our study shows that EviDTI can calibrate prediction errors. More importantly, well-calibrated uncertainty information enhances the efficiency of drug discovery by prioritizing DTIs with higher confident predictions for experimental validation. In a case study focused on tyrosine kinase modulators, uncertainty-guided predictions identify novel potential modulators targeting tyrosine kinase FAK and FLT3. These results underscore the potential of evidential deep learning as a robust tool for uncertainty quantification in DTI prediction and its broader implications for accelerating drug discovery.
- 26 July 2025
- Published by Springer Nature in Nature Communications
- Vol. 16 (1) , 1-16
Top Cited PapersOpen AccessEvidential deep learning-based drug-target interaction prediction
Drug-target interaction (DTI) prediction is a crucial component of drug discovery. Recent deep learning methods show great potential in this field but also encounter substantial challenges. These include generating reliable confidence estimates for predictions, enhancing robustness when handling novel, unseen DTIs, and mitigating the tendency toward overconfident and incorrect predictions. To solve these problems, we propose EviDTI, a novel approach utilizing evidential deep learning (EDL) for uncertainty quantification in neural network-based DTI prediction. EviDTI integrates multiple data dimensions, including drug 2D topological graphs and 3D spatial structures, and target sequence features. Through EDL, EviDTI provides uncertainty estimates for its predictions. Experimental results on three benchmark datasets demonstrate the competitiveness of EviDTI against 11 baseline models. In addition, our study shows that EviDTI can calibrate prediction errors. More importantly, well-calibrated uncertainty information enhances the efficiency of drug discovery by prioritizing DTIs with higher confident predictions for experimental validation. In a case study focused on tyrosine kinase modulators, uncertainty-guided predictions identify novel potential modulators targeting tyrosine kinase FAK and FLT3. These results underscore the potential of evidential deep learning as a robust tool for uncertainty quantification in DTI prediction and its broader implications for accelerating drug discovery.
- Open Access
- 13 November 2024
- Published by Springer Nature in Research Square
Evidential Deep Learning for Guided Drug-Target Interaction Prediction
Drug-target interaction (DTI) prediction is a crucial component of drug discovery. Recent deep learning methods show great potential in this field but also encounter substantial challenges. These include generating reliable confidence estimates for predictions, enhancing robustness when handling novel, unseen DTIs, and mitigating the tendency toward overconfident and incorrect predictions. To solve these problems, we propose EviDTI, a novel approach utilizing evidential deep learning (EDL) for uncertainty quantification in neural network-based DTI prediction. EviDTI integrates multiple data dimensions, including drug 2D topological graphs and 3D spatial structures, and target sequence features. Through EDL, EviDTI provides uncertainty estimates for its predictions. Experimental results on three benchmark datasets demonstrated the competitiveness of EviDTI against five state-of-the-art baseline models. In addition, our study shows that EviDTI can calibrate prediction errors. More importantly, well-calibrated uncertainty information enhanced the efficiency of drug discovery by prioritizing DTIs with higher confident predictions for experimental validation. In a case study focused on tyrosine kinase inhibitors (TKIs), uncertainty-guided predictions identified novel potential inhibitors targeting FAK and FLT3. These results underscore the potential of evidential deep learning as a robust tool for uncertainty quantification in DTI prediction and its broader implications for accelerating drug discovery.
- 13 November 2024
- Published by Springer Nature in Research Square
Open AccessEvidential Deep Learning for Guided Drug-Target Interaction Prediction
Drug-target interaction (DTI) prediction is a crucial component of drug discovery. Recent deep learning methods show great potential in this field but also encounter substantial challenges. These include generating reliable confidence estimates for predictions, enhancing robustness when handling novel, unseen DTIs, and mitigating the tendency toward overconfident and incorrect predictions. To solve these problems, we propose EviDTI, a novel approach utilizing evidential deep learning (EDL) for uncertainty quantification in neural network-based DTI prediction. EviDTI integrates multiple data dimensions, including drug 2D topological graphs and 3D spatial structures, and target sequence features. Through EDL, EviDTI provides uncertainty estimates for its predictions. Experimental results on three benchmark datasets demonstrated the competitiveness of EviDTI against five state-of-the-art baseline models. In addition, our study shows that EviDTI can calibrate prediction errors. More importantly, well-calibrated uncertainty information enhanced the efficiency of drug discovery by prioritizing DTIs with higher confident predictions for experimental validation. In a case study focused on tyrosine kinase inhibitors (TKIs), uncertainty-guided predictions identified novel potential inhibitors targeting FAK and FLT3. These results underscore the potential of evidential deep learning as a robust tool for uncertainty quantification in DTI prediction and its broader implications for accelerating drug discovery.
- Open Access
- 27 August 2024
- Published by MDPI AG in International Journal of Molecular Sciences
- Vol. 25 (17) , 9280
A Point Cloud Graph Neural Network for Protein–Ligand Binding Site Prediction
Predicting protein–ligand binding sites is an integral part of structural biology and drug design. A comprehensive understanding of these binding sites is essential for advancing drug innovation, elucidating mechanisms of biological function, and exploring the nature of disease. However, accurately identifying protein–ligand binding sites remains a challenging task. To address this, we propose PGpocket, a geometric deep learning-based framework to improve protein–ligand binding site prediction. Initially, the protein surface is converted into a point cloud, and then the geometric and chemical properties of each point are calculated. Subsequently, the point cloud graph is constructed based on the inter-point distances, and the point cloud graph neural network (GNN) is applied to extract and analyze the protein surface information to predict potential binding sites. PGpocket is trained on the scPDB dataset, and its performance is verified on two independent test sets, Coach420 and HOLO4K. The results show that PGpocket achieves a 58% success rate on the Coach420 dataset and a 56% success rate on the HOLO4K dataset. These results surpass competing algorithms, demonstrating PGpocket’s advancement and practicality for protein–ligand binding site prediction.
- 27 August 2024
- Published by MDPI AG in International Journal of Molecular Sciences
- Vol. 25 (17) , 9280
Open AccessA Point Cloud Graph Neural Network for Protein–Ligand Binding Site Prediction
Predicting protein–ligand binding sites is an integral part of structural biology and drug design. A comprehensive understanding of these binding sites is essential for advancing drug innovation, elucidating mechanisms of biological function, and exploring the nature of disease. However, accurately identifying protein–ligand binding sites remains a challenging task. To address this, we propose PGpocket, a geometric deep learning-based framework to improve protein–ligand binding site prediction. Initially, the protein surface is converted into a point cloud, and then the geometric and chemical properties of each point are calculated. Subsequently, the point cloud graph is constructed based on the inter-point distances, and the point cloud graph neural network (GNN) is applied to extract and analyze the protein surface information to predict potential binding sites. PGpocket is trained on the scPDB dataset, and its performance is verified on two independent test sets, Coach420 and HOLO4K. The results show that PGpocket achieves a 58% success rate on the Coach420 dataset and a 56% success rate on the HOLO4K dataset. These results surpass competing algorithms, demonstrating PGpocket’s advancement and practicality for protein–ligand binding site prediction.
- Open Access
- 25 July 2024
- Published by Oxford University Press (OUP) in Briefings in Bioinformatics
- Vol. 25 (5)
ReadCurrent: a VDCNN-based tool for fast and accurate nanopore selective sequencing
Nanopore selective sequencing allows the targeted sequencing of DNA of interest using computational approaches rather than experimental methods such as targeted multiplex polymerase chain reaction or hybridization capture. Compared to sequence-alignment strategies, deep learning (DL) models for classifying target and nontarget DNA provide large speed advantages. However, the relatively low accuracy of these DL-based tools hinders their application in nanopore selective sequencing. Here, we present a DL-based tool named ReadCurrent for nanopore selective sequencing, which takes electric currents as inputs. ReadCurrent employs a modified very deep convolutional neural network (VDCNN) architecture, enabling significantly lower computational costs for training and quicker inference compared to conventional VDCNN. We evaluated the performance of ReadCurrent across 10 nanopore sequencing datasets spanning human, yeasts, bacteria, and viruses. We observed that ReadCurrent achieved a mean accuracy of 98.57% for classification, outperforming four other DL-based selective sequencing methods. In experimental validation that selectively sequenced microbial DNA from human DNA, ReadCurrent achieved an enrichment ratio of 2.85, which was higher than the 2.7 ratio achieved by MinKNOW using the sequence-alignment strategy. In summary, ReadCurrent can rapidly classify target and nontarget DNA with high accuracy, providing an alternative in the toolbox for nanopore selective sequencing. ReadCurrent is available at https://github.com/Ming-Ni-Group/ReadCurrent.
- 25 July 2024
- Published by Oxford University Press (OUP) in Briefings in Bioinformatics
- Vol. 25 (5)
Open AccessReadCurrent: a VDCNN-based tool for fast and accurate nanopore selective sequencing
Nanopore selective sequencing allows the targeted sequencing of DNA of interest using computational approaches rather than experimental methods such as targeted multiplex polymerase chain reaction or hybridization capture. Compared to sequence-alignment strategies, deep learning (DL) models for classifying target and nontarget DNA provide large speed advantages. However, the relatively low accuracy of these DL-based tools hinders their application in nanopore selective sequencing. Here, we present a DL-based tool named ReadCurrent for nanopore selective sequencing, which takes electric currents as inputs. ReadCurrent employs a modified very deep convolutional neural network (VDCNN) architecture, enabling significantly lower computational costs for training and quicker inference compared to conventional VDCNN. We evaluated the performance of ReadCurrent across 10 nanopore sequencing datasets spanning human, yeasts, bacteria, and viruses. We observed that ReadCurrent achieved a mean accuracy of 98.57% for classification, outperforming four other DL-based selective sequencing methods. In experimental validation that selectively sequenced microbial DNA from human DNA, ReadCurrent achieved an enrichment ratio of 2.85, which was higher than the 2.7 ratio achieved by MinKNOW using the sequence-alignment strategy. In summary, ReadCurrent can rapidly classify target and nontarget DNA with high accuracy, providing an alternative in the toolbox for nanopore selective sequencing. ReadCurrent is available at https://github.com/Ming-Ni-Group/ReadCurrent.
- Open Access
Novel Scaffold Agonists of the α2A Adrenergic Receptor Identified via Ensemble-Based Strategy
The α2A adrenergic receptor (α2A-AR) serves as a critical molecular target for sedatives and analgesics. However, α2A-AR ligands with an imidazole ring also interact with an imidazoline receptor as well as other proteins and lead to undesirable effects, motivating us to develop more novel scaffold α2A-AR ligands. For this purpose, we employed an ensemble-based ligand discovery strategy, integrating long-term molecular dynamics (MD) simulations and virtual screening, to identify new potential α2A-AR agonists with novel scaffold. Our results showed that compounds SY-15 and SY-17 exhibited significant biological effects in the preliminary evaluation of protein kinase A (PKA) redistribution assays. They also reduced levels of intracellular cyclic adenosine monophosphate (cAMP) in a dose-dependent manner. Upon treatment of the cells with 100 μM concentrations of SY-15 and SY-17, there was a respective decrease in the intracellular cAMP levels by 63.43% and 53.83%. Subsequent computational analysis was conducted to elucidate the binding interactions of SY-15 and SY-17 with the α2A-AR. The binding free energies of SY-15 and SY-17 calculated by MD simulations were −45.93 and −71.97 kcal/mol. MD simulations also revealed that both compounds act as bitopic agonists, occupying the orthosteric site and a novel exosite of the receptor simultaneously. Our findings of integrative computational and experimental approaches could offer the potential to enhance ligand affinity and selectivity through dual-site occupancy and provide a novel direction for the rational design of sedatives and analgesics.
Open AccessNovel Scaffold Agonists of the α2A Adrenergic Receptor Identified via Ensemble-Based Strategy
The α2A adrenergic receptor (α2A-AR) serves as a critical molecular target for sedatives and analgesics. However, α2A-AR ligands with an imidazole ring also interact with an imidazoline receptor as well as other proteins and lead to undesirable effects, motivating us to develop more novel scaffold α2A-AR ligands. For this purpose, we employed an ensemble-based ligand discovery strategy, integrating long-term molecular dynamics (MD) simulations and virtual screening, to identify new potential α2A-AR agonists with novel scaffold. Our results showed that compounds SY-15 and SY-17 exhibited significant biological effects in the preliminary evaluation of protein kinase A (PKA) redistribution assays. They also reduced levels of intracellular cyclic adenosine monophosphate (cAMP) in a dose-dependent manner. Upon treatment of the cells with 100 μM concentrations of SY-15 and SY-17, there was a respective decrease in the intracellular cAMP levels by 63.43% and 53.83%. Subsequent computational analysis was conducted to elucidate the binding interactions of SY-15 and SY-17 with the α2A-AR. The binding free energies of SY-15 and SY-17 calculated by MD simulations were −45.93 and −71.97 kcal/mol. MD simulations also revealed that both compounds act as bitopic agonists, occupying the orthosteric site and a novel exosite of the receptor simultaneously. Our findings of integrative computational and experimental approaches could offer the potential to enhance ligand affinity and selectivity through dual-site occupancy and provide a novel direction for the rational design of sedatives and analgesics.
- Open Access
- 19 December 2023
- Published by Springer Nature in BMC Bioinformatics
- Vol. 24 (1) , 1-22
BioEGRE: a linguistic topology enhanced method for biomedical relation extraction based on BioELECTRA and graph pointer neural network
Background: Automatic and accurate extraction of diverse biomedical relations from literature is a crucial component of bio-medical text mining. Currently, stacking various classification networks on pre-trained language models to perform fine-tuning is a common framework to end-to-end solve the biomedical relation extraction (BioRE) problem. However, the sequence-based pre-trained language models underutilize the graphical topology of language to some extent. In addition, sequence-oriented deep neural networks have limitations in processing graphical features. Results: In this paper, we propose a novel method for sentence-level BioRE task, BioEGRE (BioELECTRA and Graph pointer neural net-work for Relation Extraction), aimed at leveraging the linguistic topological features. First, the biomedical literature is preprocessed to retain sentences involving pre-defined entity pairs. Secondly, SciSpaCy is employed to conduct dependency parsing; sentences are modeled as graphs based on the parsing results; BioELECTRA is utilized to generate token-level representations, which are modeled as attributes of nodes in the sentence graphs; a graph pointer neural network layer is employed to select the most relevant multi-hop neighbors to optimize representations; a fully-connected neural network layer is employed to generate the sentence-level representation. Finally, the Softmax function is employed to calculate the probabilities. Our proposed method is evaluated on three BioRE tasks: a multi-class (CHEMPROT) and two binary tasks (GAD and EU-ADR). The results show that our method achieves F1-scores of 79.97% (CHEMPROT), 83.31% (GAD), and 83.51% (EU-ADR), surpassing the performance of existing state-of-the-art models. Conclusion: The experimental results on 3 biomedical benchmark datasets demonstrate the effectiveness and generalization of BioEGRE, which indicates that linguistic topology and a graph pointer neural network layer explicitly improve performance for BioRE tasks.
- 19 December 2023
- Published by Springer Nature in BMC Bioinformatics
- Vol. 24 (1) , 1-22
Open AccessBioEGRE: a linguistic topology enhanced method for biomedical relation extraction based on BioELECTRA and graph pointer neural network
Background: Automatic and accurate extraction of diverse biomedical relations from literature is a crucial component of bio-medical text mining. Currently, stacking various classification networks on pre-trained language models to perform fine-tuning is a common framework to end-to-end solve the biomedical relation extraction (BioRE) problem. However, the sequence-based pre-trained language models underutilize the graphical topology of language to some extent. In addition, sequence-oriented deep neural networks have limitations in processing graphical features. Results: In this paper, we propose a novel method for sentence-level BioRE task, BioEGRE (BioELECTRA and Graph pointer neural net-work for Relation Extraction), aimed at leveraging the linguistic topological features. First, the biomedical literature is preprocessed to retain sentences involving pre-defined entity pairs. Secondly, SciSpaCy is employed to conduct dependency parsing; sentences are modeled as graphs based on the parsing results; BioELECTRA is utilized to generate token-level representations, which are modeled as attributes of nodes in the sentence graphs; a graph pointer neural network layer is employed to select the most relevant multi-hop neighbors to optimize representations; a fully-connected neural network layer is employed to generate the sentence-level representation. Finally, the Softmax function is employed to calculate the probabilities. Our proposed method is evaluated on three BioRE tasks: a multi-class (CHEMPROT) and two binary tasks (GAD and EU-ADR). The results show that our method achieves F1-scores of 79.97% (CHEMPROT), 83.31% (GAD), and 83.51% (EU-ADR), surpassing the performance of existing state-of-the-art models. Conclusion: The experimental results on 3 biomedical benchmark datasets demonstrate the effectiveness and generalization of BioEGRE, which indicates that linguistic topology and a graph pointer neural network layer explicitly improve performance for BioRE tasks.
- Open Access
- 27 November 2023
- Published by Taylor & Francis in mAbs
- Vol. 15 (1) , 2285904
BERT2DAb: a pre-trained model for antibody representation based on amino acid sequences and 2D-structure
Prior research has generated a vast amount of antibody sequences, which has allowed the pre-training of language models on amino acid sequences to improve the efficiency of antibody screening and optimization. However, compared to those for proteins, there are fewer pre-trained language models available for antibody sequences. Additionally, existing pre-trained models solely rely on embedding representations using amino acids or k-mers, which do not explicitly take into account the role of secondary structure features. Here, we present a new pre-trained model called BERT2DAb. This model incorporates secondary structure information based on self-attention to learn representations of antibody sequences. Our model achieves state-of-the-art performance on three downstream tasks, including two antigen-antibody binding classification tasks (precision: 85.15%/94.86%; recall:87.41%/86.15%) and one antigen-antibody complex mutation binding free energy prediction task (Pearson correlation coefficient: 0.77). Moreover, we propose a novel method to analyze the relationship between attention weights and contact states of pairs of subsequences in tertiary structures. This enhances the interpretability of BERT2DAb. Overall, our model demonstrates strong potential for improving antibody screening and design through downstream applications.
- 27 November 2023
- Published by Taylor & Francis in mAbs
- Vol. 15 (1) , 2285904
Open AccessBERT2DAb: a pre-trained model for antibody representation based on amino acid sequences and 2D-structure
Prior research has generated a vast amount of antibody sequences, which has allowed the pre-training of language models on amino acid sequences to improve the efficiency of antibody screening and optimization. However, compared to those for proteins, there are fewer pre-trained language models available for antibody sequences. Additionally, existing pre-trained models solely rely on embedding representations using amino acids or k-mers, which do not explicitly take into account the role of secondary structure features. Here, we present a new pre-trained model called BERT2DAb. This model incorporates secondary structure information based on self-attention to learn representations of antibody sequences. Our model achieves state-of-the-art performance on three downstream tasks, including two antigen-antibody binding classification tasks (precision: 85.15%/94.86%; recall:87.41%/86.15%) and one antigen-antibody complex mutation binding free energy prediction task (Pearson correlation coefficient: 0.77). Moreover, we propose a novel method to analyze the relationship between attention weights and contact states of pairs of subsequences in tertiary structures. This enhances the interpretability of BERT2DAb. Overall, our model demonstrates strong potential for improving antibody screening and design through downstream applications.
- Open Access
- 22 August 2023
- Published by Springer Nature in Research Square
BioEGRE: A Linguistic Topology Enhanced Method for Biomedical Relation Extraction based on BioELECTRA and Graph Pointer Neural Network
Background: Automatic and accurate extraction of various biomedical relations from literature is a crucial subtask of bio-medical text mining. Currently, stacking various classification networks on pre-trained language models to perform fine-tuning is a common framework to end-to-end solve the biomedical relation extraction (BioRE) problem. However, the sequence-based pre-trained language models underutilize the graphical topology of language to some extent. In addition, sequence-oriented deep neural networks have limitations in processing graphical features. Results: In this paper, we propose a novel method for sentence-level BioRE task, BioEGRE (BioELECTRA & Graph pointer neural net-work for Relation Extraction), which can capitalize the topological features of language. First, biomedical literature is preprocessed, which preserves sentences containing pre-fetched entity pair. Second, SciSpaCy is used to perform dependency parsing; sentences are modeled as graphs based on the parsing results; BioELECTRA is used to generate token-level representation, which is modeled as the attribute of nodes in sentence graphs; a graph pointer neural network layer is utilized to select the most relevant multi-hop neighbors to optimize the representation; a full-connected neural network layer is used to generate the sentence-level representation. Finally, a Softmax function is utilized to calculate probabilities. Our method is evaluated on a multi-type (CHEMPROT) and 2 binary (GAD and EU-ADR) BioRE tasks respectively, and achieves 79.97% (CHEMPROT), 83.31% (GAD) and 83.51% (EU-ADR) of F1-score, which outperforms existing state-of-the-art models. Conclusion: The experimental results on 3 biomedical benchmark datasets demonstrate the effectiveness and generalization of BioEGRE, which indicates that linguistic topology and a graph pointer neural network layer explicitly improve performance for BioRE tasks.
- 22 August 2023
- Published by Springer Nature in Research Square
Open AccessBioEGRE: A Linguistic Topology Enhanced Method for Biomedical Relation Extraction based on BioELECTRA and Graph Pointer Neural Network
Background: Automatic and accurate extraction of various biomedical relations from literature is a crucial subtask of bio-medical text mining. Currently, stacking various classification networks on pre-trained language models to perform fine-tuning is a common framework to end-to-end solve the biomedical relation extraction (BioRE) problem. However, the sequence-based pre-trained language models underutilize the graphical topology of language to some extent. In addition, sequence-oriented deep neural networks have limitations in processing graphical features. Results: In this paper, we propose a novel method for sentence-level BioRE task, BioEGRE (BioELECTRA & Graph pointer neural net-work for Relation Extraction), which can capitalize the topological features of language. First, biomedical literature is preprocessed, which preserves sentences containing pre-fetched entity pair. Second, SciSpaCy is used to perform dependency parsing; sentences are modeled as graphs based on the parsing results; BioELECTRA is used to generate token-level representation, which is modeled as the attribute of nodes in sentence graphs; a graph pointer neural network layer is utilized to select the most relevant multi-hop neighbors to optimize the representation; a full-connected neural network layer is used to generate the sentence-level representation. Finally, a Softmax function is utilized to calculate probabilities. Our method is evaluated on a multi-type (CHEMPROT) and 2 binary (GAD and EU-ADR) BioRE tasks respectively, and achieves 79.97% (CHEMPROT), 83.31% (GAD) and 83.51% (EU-ADR) of F1-score, which outperforms existing state-of-the-art models. Conclusion: The experimental results on 3 biomedical benchmark datasets demonstrate the effectiveness and generalization of BioEGRE, which indicates that linguistic topology and a graph pointer neural network layer explicitly improve performance for BioRE tasks.
- Open Access
- 1 January 2023
- Published by American Association for the Advancement of Science (AAAS) in Computational and Structural Biotechnology Journal
- Vol. 21, 5839-5850
Improving antibody optimization ability of generative adversarial network through large language model
Generative adversarial networks (GANs) have successfully generated functional protein sequences. However, traditional GANs often suffer from inherent randomness, resulting in a lower probability of obtaining desirable sequences. Due to the high cost of wet-lab experiments, the main goal of computer-aided antibody optimization is to identify high-quality candidate antibodies from a large range of possibilities, yet improving the ability of GANs to generate these desired antibodies is a challenge. In this study, we propose and evaluate a new GAN called the Language Model Guided Antibody Generative Adversarial Network (AbGAN-LMG). This GAN uses a language model as an input, harnessing such models’ powerful representational capabilities to improve the GAN’s generation of high-quality antibodies. We conducted a comprehensive evaluation of the antibody libraries and sequences generated by AbGAN-LMG for COVID-19 (SARS-CoV-2) and Middle East Respiratory Syndrome (MERS-CoV). Results indicate that AbGAN-LMG has learned the fundamental characteristics of antibodies and that it improved the diversity of the generated libraries. Additionally, when generating sequences using AZD-8895 as the target antibody for optimization, over 50% of the generated sequences exhibited better developability than AZD-8895 itself. Through molecular docking, we identified 70 antibodies that demonstrated higher affinity for the wild-type receptor-binding domain (RBD) of SARS-CoV-2 compared to AZD-8895. In conclusion, AbGAN-LMG demonstrates that language models used in conjunction with GANs can enable the generation of higher-quality libraries and candidate sequences, thereby improving the efficiency of antibody optimization. AbGAN-LMG is available at http://39.102.71.224:88/.
- 1 January 2023
- Published by American Association for the Advancement of Science (AAAS) in Computational and Structural Biotechnology Journal
- Vol. 21, 5839-5850
Open AccessImproving antibody optimization ability of generative adversarial network through large language model
Generative adversarial networks (GANs) have successfully generated functional protein sequences. However, traditional GANs often suffer from inherent randomness, resulting in a lower probability of obtaining desirable sequences. Due to the high cost of wet-lab experiments, the main goal of computer-aided antibody optimization is to identify high-quality candidate antibodies from a large range of possibilities, yet improving the ability of GANs to generate these desired antibodies is a challenge. In this study, we propose and evaluate a new GAN called the Language Model Guided Antibody Generative Adversarial Network (AbGAN-LMG). This GAN uses a language model as an input, harnessing such models’ powerful representational capabilities to improve the GAN’s generation of high-quality antibodies. We conducted a comprehensive evaluation of the antibody libraries and sequences generated by AbGAN-LMG for COVID-19 (SARS-CoV-2) and Middle East Respiratory Syndrome (MERS-CoV). Results indicate that AbGAN-LMG has learned the fundamental characteristics of antibodies and that it improved the diversity of the generated libraries. Additionally, when generating sequences using AZD-8895 as the target antibody for optimization, over 50% of the generated sequences exhibited better developability than AZD-8895 itself. Through molecular docking, we identified 70 antibodies that demonstrated higher affinity for the wild-type receptor-binding domain (RBD) of SARS-CoV-2 compared to AZD-8895. In conclusion, AbGAN-LMG demonstrates that language models used in conjunction with GANs can enable the generation of higher-quality libraries and candidate sequences, thereby improving the efficiency of antibody optimization. AbGAN-LMG is available at http://39.102.71.224:88/.