Abstract
In graphical-based chemistry machine learning (ML) models, a precise atomic correlation depiction, known as molecular topology, can significantly increase the model's performance. However, owing to the high acquisition cost, existing chemical graph data often use chemical bonds or atomic distances as molecular topology proxies. Such approximation, though practical, inevitably introduces inaccurate information that amplifies the learning difficulty of models and ultimately detracts from their performance. Despite various available data preprocessing techniques that can mitigate potential damage from inaccurate topology, those complex and data-type-tailored methods limit their generalizability in dealing with different types of chemical graph data. This paper introduces the dynamic connection layer (DCL) to address this challenge in a more general framework. This innovative graph neural layer dynamically modifies the input "topological information" (whether represented by chemical bonds or distances) to obtain a more accurate molecular topology description while preserving the generality as a trainable neural layer. We assessed the efficacy of the DCL layer using a data set derived from QM9 and demonstrated its efficiency in processing chemical graph data.