材料研究中的可解释机器学习

王冠杰, 刘盛咸, 周健, 孙志梅

Explainable Machine Learning in the Research of Materials Science

WANG Guanjie, LIU Shengxian, ZHOU Jian, SUN Zhimei

图7 蛋白质-配体复合物3F7H的三维结构，基于生物学知识的蛋白质-配体相互作用，及PotentialNet模型和InteractionNet模型通过LRP方法获得的原子对解离常数预测贡献的热图[96]

Fig.7 Three-dimensional structure of the protein-ligand complex, 3F7H (The protein is depicted in a cartoon (green), and the ligand is depicted in color-coded ball-and-stick. Atom colors: gray (C), red (O), and blue (N)) (a); knowledge-based estimation of protein-ligand interactions (Hydrogen bonds are depicted in red dashed lines, and hydrophobic contacts are depicted in gray dashed lines) (b); heat map for the atomic contributions on the prediction of interaction, obtained from the LRP on PotentialNet (c); and heat map for the atomic contributions on the prediction of interaction, obtained from the LRP on InteractionNet (d) (The contributions are illustrated with color intensity of red (positive influence), white (zero influence), and blue (negative influence) colors)[96]