多对齐对比学习用于酶-反应检索
Multi-Alignment Contrastive Learning for Enzyme--Reaction Retrieval
AI总结：
提出多对齐对比学习框架，通过联合建模酶-反应跨域兼容性及功能注释驱动的域内关系，并引入Gromov-Wasserstein正则化项，提升酶虚拟筛选和双向检索性能。
摘要：
识别催化目标生化反应的酶是计算酶发现和生物催化剂设计的关键步骤。最近的表示学习方法将这一问题表述为酶-反应匹配，其中配对的酶和反应被嵌入到共享空间中。然而，大多数现有方法主要依赖于成对的酶-反应监督，并且对反应集或酶家族内部关系的利用有限。本文介绍了一种用于生化检索的多对齐对比学习框架。该框架联合建模酶与反应之间的跨域兼容性以及由功能注释诱导的域内关系。此外，受Gromov-Wasserstein启发的正则化目标鼓励学习的酶和反应表示空间之间的几何一致性。通过将成对的催化监督与高阶关系对齐相结合，该模型捕获了直接的酶-反应关联以及更广泛的功能组织。我们在酶虚拟筛选和双向酶-反应检索任务上评估了该方法。在EnzymeMap上的实验表明，与强对比基线相比，在BEDROC和富集因子指标下，早期识别性能有所提高。在ReactZyme上，该方法在基于时间、酶相似性和反应相似性的划分中均取得了一致的增益，展示了对未见酶和未见反应的鲁棒性。消融研究进一步表明，域内对齐、功能监督和几何正则化项各自对观察到的改进有所贡献。这些结果表明，建模多种形式的对齐可以改进用于酶发现、反应注释及相关计算生物学应用的对比检索模型。
英文摘要：
Identifying enzymes that catalyze target biochemical reactions is a key step in computational enzyme discovery and biocatalyst design. Recent representation-learning methods formulate this problem as enzyme--reaction matching, where paired enzymes and reactions are embedded into a shared space. However, most existing approaches primarily rely on pairwise enzyme--reaction supervision and make limited use of the relationships within reaction sets or enzyme families. This work introduces a multi-alignment contrastive learning framework for biochemical retrieval. The framework jointly models cross-domain compatibility between enzymes and reactions and within-domain relationships induced by functional annotations. In addition, a Gromov--Wasserstein-inspired regularization objective encourages geometric consistency between the learned enzyme and reaction representation spaces. By combining pairwise catalytic supervision with higher-order relational alignment, the model captures both direct enzyme--reaction associations and broader functional organization. We evaluate the approach on enzyme virtual screening and bidirectional enzyme--reaction retrieval tasks. Experiments on EnzymeMap show improved early-recognition performance under BEDROC and enrichment-factor metrics compared with strong contrastive baselines. On ReactZyme, the method achieves consistent gains across time-based, enzyme-similarity, and reaction-similarity splits, demonstrating robustness to unseen enzymes and unseen reactions. Ablation studies further indicate that within-domain alignment, functional supervision, and the geometric regularization term each contribute to the observed improvements. These results suggest that modeling multiple forms of alignment can improve contrastive retrieval models for enzyme discovery, reaction annotation, and related computational biology applications.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
