Molformer: Motif-based Roto-Translation Invariant Transformer on 3D Heterogeneous Molecular Graphs

Procuring expressive molecular representations underpins AI-driven molecule design and scientific discovery. The research to date mainly focuses on atom-level homogeneous molecular graphs, ignoring the rich information in subgraphs or motifs. As for 3D structures, previous studies fail to efficiently capture long-range dependencies nor consider the non-uniformity of interatomic distances. To address such issues, we formulate heterogeneous molecular graphs, and introduce Molformer to exploit both molecular motifs and 3D geometry. Specifically, we extract motifs based on functional groups for small molecules and use reinforcement learning for proteins respectively, and construct heterogeneous molecular graphs composed of both atom-level and motif-level nodes. To utilize 3D spatial information, Molformer adopts a roto-translation invariant convolutional position encoding. It is coupled with a multi-scale self-attention mechanism to capture local fine-grained patterns with increasing contextual scales, and an attentive farthest point sampling algorithm to obtain the molecular representations. We validate Molformer across a few domains including quantum chemistry, physiology, and biophysics. Experiments show that Molformer outperforms state-of-the-art baselines. Our work provides a promising way to utilize informative motifs and amalgamate 3D geometric information.