计算机科学
可解释性
人工智能
机器学习
图形
推论
化学空间
突出
理论计算机科学
限制
杠杆(统计)
深度学习
一般化
模式
虚拟筛选
优先次序
显著性(神经科学)
航程(航空)
启发式
图同构
词汇
编码器
特征学习
概率逻辑
作者
Jiahui Guan,Qianhui Jiang,Peilin Xie,Xuxin He,Dan Yu,Hexian Zhang,Lantian Yao,Junwen Wang,Jiahui Guan,Qianhui Jiang,Peilin Xie,Xuxin He,Dan Yu,Hexian Zhang,Lantian Yao,Junwen Wang
标识
DOI:10.1021/acs.jcim.5c02222
摘要
Chemical allergens are prevalent in both consumer and industrial products, often triggering hypersensitivity reactions with significant public health and regulatory implications. Traditional experimental screening is time-consuming and labor-intensive, hindering the pace of allergen discovery and risk assessment. Existing computational approaches often rely on handcrafted molecular fingerprints and shallow classifiers, which struggle to adequately capture molecular topology or cross-modal dependencies, limiting generalization and interpretability. To address these challenges, we propose MGCL-CAP, a deep learning framework for chemical allergenicity prediction that leverages masked graph contrastive learning and gated cross-attention fusion. MGCL-CAP performs random subgraph masking within a shared graph isomorphism network encoder to learn structure-invariant graph embeddings, enhancing resilience to missing or noisy substructures. These embeddings are then integrated with one-dimensional molecular fingerprints via multihead gated cross-attention to align modalities and emphasize salient chemical cues. Experimental results show that MGCL-CAP outperforms state-of-the-art allergenicity predictors and remains stable across a range of hyperparameters. Interpretability analyses highlight substructures suggestive of sensitization-related mechanisms, providing valuable mechanistic insights for future chemical safety assessments. Overall, MGCL-CAP offers a reliable tool for the computational assessment of chemical allergenicity, enabling efficient candidate prioritization and supporting safer formulation design while reducing experimental burdens.
科研通智能强力驱动
Strongly Powered by AbleSci AI