Molecular Representation Learning by Leveraging Chemical Information
Molecular property prediction is of great importance in AI drug design due to its high experimental efficiency compared with biological experiments. As graph neural networks have achieved great success in many domains, some studies apply graph neural networks to molecular property prediction and regard each molecule as a graph. A molecule’s atom is regarded as a node of the graph, while its bond is regarded as an edge of the graph. However, most existing methods simply apply general graph neural networks without considering the domain knowledge. As chemical information is highly related to molecular functions, it is critical for accurate property prediction. Thus, we leverage chemical information to learn molecular representation by integrating molecular fingerprints, i.e., the presence or absence of particular chemical substructures. We compare our proposed method to several strong baselines, and our proposed method significantly surpasses other methods. Up to now, our method ranks first in the Open Graph Benchmark(OGB) leaderboard for ogbg-molhiv.
PDF
OGB
