Key Substructure Learning with Chemical Intuition for Material Property Prediction

Substructures are crucial factors influencing material molecular properties, and the effective identification of molecular substructures can accelerate novel material discovery. The chemical intuition of irregularity and contribution variability existing in substructures present challenges in modeling molecular structural information, resulting in unsatisfactory performance in molecular property prediction. In this paper, we propose Key Substructure Learning with Chemical Intuition for Material Property Prediction (KSCI), including two modules. (1) Graph self-attention-based irregular substructure representation (GSR) adaptively extracts irregular molecular substructures using substructure feature entropy coefficient and multi-hop neighbor feature weighted aggregation. GSR utilizes a novel substructure graph self-attention mechanism, which captures the molecular graph’s topology to achieve explicit encoding of substructures. (2) Gain-driven key substructure learning (GKL) calculates the mutual information gain of all molecular substructures to material molecular representation. In this way, the substructures with higher gain to molecular properties can be identified and direct the generation of molecular graph-level embedded representations. Experimental results on four datasets demonstrate that the Mean Absolute Error of KSCI outperforms eleven state-of-the-art material property prediction models. Visualization results show that KSCI exhibits satisfactory performance in irregular key substructure identification.