Year: 2026
Type: preprint
Abstract: Most deep learning models of mixtures encode each molecule separately and then combine the resulting vectors. This combination step—per-molecule pooling—silently discards molecular topology. Treating the mixture as a unified atomic system, without molecular boundaries, avoids this loss. To illustrate this loss, we compare molecular-level and atomic-level modeling on two tasks with different information requirements: one dominated by functional groups, one by topology. Molecule-level baselines each excel on one but fail on the other; a single boundary-free architecture matches the strongest baselines on both, and where the two kinds of information must act together—olfactory mixtures—it reaches macro-AUROC 0.9658, far exceeding previous methods. Modeling a mixture as its atoms, and nothing more, captures the underlying physics without task-specific compromises.
Source: ChemRxiv
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7203743323