MolExplain：用于可解释分子性质预测的交互式工具

中文摘要 AI 辅助

机器学习在分子性质预测中的应用在药物发现中日益普遍，然而大多数模型作为黑箱运行，仅返回预测结果而不揭示哪些结构特征驱动了该预测。MolExplain通过将性质预测与子结构级别的可视化可解释性结合在交互式网页界面中，弥补了这一差距。该系统将分子特征化为摩根指纹，使用训练好的XGBoost模型进行分类，并应用SHAP归因生成平滑的热力图叠加，指示分子中哪些区域对预测性质有贡献或反对。应用于环肽膜通透性时，该工具的归因独立地恢复了骨架N-甲基化在改善被动膜扩散中的已知作用，与既定化学一致。尽管在环肽上进行了演示，该框架旨在推广到其他分子性质，将MolExplain定位为交互式、可解释性驱动的分子设计平台。

英文摘要

The application of machine learning to molecular property prediction has become increasingly prevalent in drug discovery, yet most models operate as black boxes, returning a prediction without revealing which structural features drive it. MolExplain addresses this gap by combining property prediction with sub-structure level visual explainability in an interactive web interface. The system featurizes molecules as Morgan fingerprints, classifies them using a trained XGBoost model, and applies SHAP attribution to produce a smooth heatmap overlay indicating which regions of the molecule contribute for or against the predicted property. Applied to cyclic peptide membrane permeability, the tool's attribution independently recovers the known role of backbone N-methylation in improving passive membrane diffusion, consistent with established chemistry. While demonstrated on cyclic peptides, the framework is designed to generalize to other molecular properties, positioning MolExplain as a platform for interactive, explainability-driven molecular design.

发表机构

• Brandeis University（布兰迪斯大学）

机构由 AI 辅助整理，请以论文原文为准。