基于多模态对比扩散的抗菌肽生成模型

An Antimicrobial Peptide Generation Model Based on Multimodal Contrastive Diffusion

摘要: 目的：构建一种融合序列、结构与表面信息的抗菌肽生成模型，以提高候选肽的新颖性、稳定性和功能相关性。方法：整合公开抗菌肽数据库资源，经序列标准化、去冗余、结构补全和质量控制后构建20104条抗菌肽样本，并配对构建等量非抗菌肽样本。提出基于多模态对比扩散的抗菌肽生成模型SSSCD (Sequence-Structure-Surface Contrastive Diffusion)，将20维氨基酸独热编码、替换矩阵与疏水性特征组成序列表征，将主链坐标、AAindex理化描述符和残基层面表面几何拓扑特征组成结构表征；序列分支采用Transformer进行离散扩散去噪，结构分支采用等变图神经网络进行连续坐标去噪，并引入跨模态对比学习和能量引导采样。结果：SSSCD在Similarity、Instability、Antimicrobial和Docking score上分别达到24.93、40.05、0.87和1740，整体优于LSTM-RNN、AMPGAN、HydrAMP和DiffAB等模型。结论：SSSCD能够在统一框架下联合建模抗菌肽序列语义、三维骨架和表面特征，提升结构合理且功能相关的抗菌肽候选物的生成，为抗菌肽的筛选与后续研究提供了一种可行的计算策略。

Abstract: Objective: To develop an antimicrobial peptide generation model based on multimodal contrastive diffusion by integrating peptide sequence, backbone structure and residue-level surface descriptors, thereby improving the novelty, stability and functional relevance of generated candidates. Methods: Public antimicrobial peptide resources were integrated and subjected to sequence normalization, redundancy removal, structural completion and quality control. A dataset containing 20,104 antimicrobial peptide samples was constructed, together with an equal number of non-antimicrobial peptide samples. A multimodal contrastive diffusion model termed SSSCD (Sequence-Structure-Surface Contrastive Diffusion) was proposed. Sequence representations were constructed using amino acid one-hot encodings, substitution matrices and hydrophobicity-related features, whereas structural representations incorporated backbone coordinates, AAindex physicochemical descriptors and residue-level surface geometric-topological features. A Transformer-based discrete diffusion branch was employed for sequence denoising, while an equivariant graph neural network was used for continuous coordinate denoising. Cross-modal contrastive learning and energy-guided sampling were further introduced to improve sequence-structure consistency and local geometric rationality. Results: SSSCD achieved Similarity, Instability, Antimicrobial and Docking scores of 24.93, 40.05, 0.87 and 1740, respectively, outperforming LSTM-RNN, AMPGAN, HydrAMP and DiffAB. Conclusion: SSSCD effectively integrates sequence semantics, backbone geometry and surface-level descriptors within a unified contrastive diffusion framework. The model improves the generation of structurally plausible and functionally relevant antimicrobial peptide candidates, providing a feasible computational strategy for antimicrobial peptide generation and downstream screening.