摘要
目前抗原⁃抗体结合亲和力预测的人工智能方法都基于序列或结构的单一建模,难以捕获更全面的信息.因此,提出了一种基于结构的多模态特征融合方法.该方法主要包括三个模块:多模态抗体信息挖掘模块、多模态抗原信息挖掘模块和融合预测模块.在多模态抗体信息挖掘模块中,引入Roformer网络分别提取抗体的重链和轻链信息,同时采用GearNet挖掘抗体的结构信息.随后,借助交叉力注意机制自适应地融合序列和结构信息.在多模态抗原信息挖掘模块中,使用蛋白质大语言模型ESM2(Evolutionary Scale Modeling v2)来实现抗原序列信息的抽取.在融合预测模块中,为了实现全面且高效的亲和力预测协同,基于卷积神经网络(Convolutional Neural Networks,CNN)构建了一个多尺度特征协同提取网络,进一步提升抗体和抗原表示的可判别性.最后,将多尺度的抗体和抗原表示输入融合层,生成稳健的亲和力预测结果.实验结果表明,提出的模型在基准数据集上的表现超越了所有的基线模型,并在独立测试集上表现优异,证明该方法具有强大的预测与泛化能力.
Currently,artificial intelligence methods for predicting antigen⁃antibody binding affinity predominantly rely on either sequence⁃based or structure⁃based unimodal modeling approaches,which limit their ability to capture comprehensive interaction information.Therefore,we present a structure⁃based multi⁃modal feature fusion framework for antigen⁃antibody affinity prediction.The framework mainly consists of three components:a multi⁃modal antibody information mining module,a multi⁃modal antigen information mining module,and a fusion prediction module.In the antibody information mining module,we employ the Roformer network to separately extract heavy chain and light chain information while utilizing the GearNet to capture structural information of antibodies.The sequence and structural information is then adaptively fused through a cross⁃attention mechanism.For antigen information extraction,we implement the ESM2(Evolutionary Scale Modeling v2)protein language model to process antigen sequence data.The fusion prediction module incorporates a multi⁃scale feature extraction network based on CNN(Convolutional Neural Networks)to enhance the discriminative power of antibody and antigen representations,enabling comprehensive and efficient affinity prediction.Finally,the multi⁃scale representations of antibodies and antigens are fed into a fusion layer to generate robust affinity prediction results.Experimental results demonstrate that our proposed model surpasses all baseline methods on benchmark datasets and shows excellent performance on independent test sets,demonstrating the method's robust predictive power and generalization capability.
作者
施佳豪
姜舒
鞠恒荣
丁卫平
Shi Jiahao;Jiang Shu;Ju Hengrong;Ding Weiping(School of Artificial Intelligence and Computer Science,Nantong University,Nantong,226019,China)
出处
《南京大学学报(自然科学版)》
北大核心
2026年第1期36-47,共12页
Journal of Nanjing University(Natural Science)
基金
国家自然科学基金(62406153)
江苏省高等学校自然科学研究(23KJB520031)
江苏省高校自然科学研究项目(24KJB520032)。
关键词
抗原-抗体结合亲和力预测
多模态融合
结构信息
注意力机制
antigen-antibody binding affinity prediction
multi-modal fusion
structural information
attention mechanism
