Hierarchical Molecular Attention Network: Improving Molecular Property Prediction Through Substructure Identification.

Few-shot molecular property prediction is a persisting challenge in many biology-related tasks, because the same molecule may exhibit different properties (e.g., active or inactive) in different tasks. Existing methods view all atoms as equally important and attend to predict the properties by averaging the features of similar molecules, which ignores key substructures within molecules and leads to poor prediction performance. Since a molecule implicitly includes key substructures, which determines the properties of the molecular, and the atom combination forms key substructures, we focus on the atom combination in this paper. With this, we propose the Hierarchical Molecular Attention Network (HMAN) to predict the molecular properties through combining atoms. First, we utilize the average pooling to extract both the prototype and the molecular features, and use Graph Neural Networks (GNN) to extract the atomic features, then concatenates the prototype and molecular features with all atomic features as input to the self-attention mechanism to calculate the different weights for atoms. Here, we select the top $B$ atoms with the highest attention scores to form the key substructures. Second, we view the formed key substructures as the query vectors and regard the molecular features as the key-value pairs, and then feed them into another self-attention to obtain the scores of key substructures. We still select $k$ key substructures with the highest scores, and weight the sum of a molecular feature and $k$ key substructures to predict the molecular properties. To train the HMAN, we design a new loss function, which includes Binary Cross-Entropy and a weighted negative log-likelihood. The former is to predict the molecular properties, and the last one is to optimize the weight distribution, rendering that the weight distribution matches the predicted probability distribution, with back-propagation. Theoretical analysis proves the convergence of the new loss function, and extensive experimental results demonstrate that HMAN significantly outperforms SOTA baseline models in molecular property prediction tasks.