Abstract: To address the over-reliance on global features and insufficient task adaptability of molecular representations in cross-modal molecular retrieval, a reinforcement learning-based fine-grained cross-modal molecular retrieval method was proposed. For the text end, the pre-trained SciBERT model was used to extract token-level and sentence-level representations. For the molecular end, Graph Convolutional Network (GCN) was adopted to obtain group-level and molecule-level representations. The method introduced the Proximal Policy Optimization (PPO) algorithm in reinforcement learning to achieve accurate alignment between tokens and groups, and dynamically generates token representations matching the semantics of groups. Progressive strategy was applied in overall training, and retrieval learning was conducted by maximizing the similarity of matched text-molecule pairs through the contrastive loss function. Experimental results on the ChEBI-20 dataset show that this method outperforms current mainstream methods in metrics such as mean reciprocal rank and Hits@1, and providing a new solution for cross-modal molecular retrieval tasks.

Key words: Proximal Policy Optimization &, #40

摘要： 针对跨模态分子检索中过度依赖全局特征和分子表征缺乏任务自适应能力的问题，提出一种基于强化学习的细粒度跨模态分子检索方法。该方法文本端采用预训练SciBERT模型提取Token级和句子级表征，分子端采用图卷积网络(GCN)获取基团级和分子级表征。方法引入强化学习中的近端策略优化(PPO)算法实现Token与基团的精准对齐，动态生成与基团语义相匹配的Token表征。整体训练采用渐进式策略，通过对比损失函数最大化匹配文本-分子对的相似度进行检索学习。在数据集ChEBI-20上的实验结果表明，该方法在平均互逆排名和Hits@1等指标均优于目前的主流方法，为跨模态分子检索任务提供了新的解决方案。