AIMETRY: GEOMETRY-SENSITIVE 3D MOLECULAR REPRESENTATIONS FROM QUANTUM MECHANICAL LEARNED FEATURES
Quantifying molecular similarity depends on both chemical composition and 3D structure, which are difficult to capture in a single fixed-length representation. Contrastive learning is a common machine learning approach to generate such representations (molecular embeddings), pulling each molecule toward a positive anchor of itself. These anchors typically come from graph augmentations that alter chemical identity or from other conformers of the same molecule. However, using conformers as positive anchors in standard contrastive learning suppresses geometric distinctions and graded structural relationships. Here, we introduce AIMetry, a geometry-sensitive representation learning scheme in which a graph transformer aggregates spatially and chemically aware atomic features from AIMNet2, a machine-learned interatomic potential trained on quantum-chemical data, into a molecular embedding. We guide standard contrastive learning using fingerprint similarity, which modulates embedding similarity between different-molecule pairs, and root-mean-square deviation, which modulates embedding similarity between conformer pairs. Trained on approximately 44 million conformers of 2.8 million ChEMBL molecules, AIMetry demonstrates higher retention of fingerprint-defined structural relationships and greater geometric sensitivity than a reference model pretrained with standard contrastive learning. AIMetry also improves linear-probe performance relative to the reference model across the evaluated datasets, while benefits under nonlinear readout and fine-tuning are task-specific. Chemical and conformational analyses provide complementary evidence about the information retained in the representation.
Authors
- Olexandr Isayev (ORCID: https://orcid.org/0000-0001-7581-8497)
- Shuhao Zhang (ORCID: https://orcid.org/0000-0002-5174-1875)
- Sheng Shu (ORCID: https://orcid.org/0009-0007-6518-7485)
- Kaiya Taura (ORCID: https://orcid.org/0009-0003-8470-9155)
Institutions
- University of Pittsburgh (US)
- Management Science Associates (United States) (US)
- Carnegie Mellon University (US)
Publication Details
- Journal
- ChemRxiv
- Published
- 2026-11-18
- DOI
- https://doi.org/10.26434/chemrxiv.15012428/v1
- Primary Topic
- Computational Drug Discovery Methods
- Type
- preprint