Weighted Molecular Views with Global Attention Tokens
Year: 2026
Type: preprint
Abstract: Accurate molecular property prediction strongly depends on the quality of molecular representations used as input to machine learning models. [1] proposed a molecular featurization method based on atomic coordinates that generates weighted views of a molecule. [2] proposedmodifications in [1] to the featurizationmethod for dimensionality reduction, which led to reduced empirical performance. In this work, we propose an alternative approach for dimensionality reduction and computational efficiency. Our proposed modifications increase the representativeness of molecular representations by generating additional molecular views, while reducing the dimensionality of these representations through the introduction of a global node into the architecture. The proposed modifications were evaluated on the QM9 dataset, where they demonstrated improved predictive performance. This work also provides a theoretical perspective to support and explain the proposed modifications.
Source: ChemRxiv
Authors: Muhammad Shahzeb Ali, Shahtaj Bano
Institutions: Ohio University, RMIT University
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7162784543