Abstract
<jats:p>The properties of a molecular mixture—the scent of a perfume, the thermodynamic behavior of a solvent blend—emerge from cross-molecule atomic interactions, not from whole-molecule properties alone. Yet dominant machine learning paradigms encode mixture components as independent molecules, discarding cross-molecule interactions before they can be modeled. We introduce a unified paradigm that treats all constituent molecules as one interacting system using graph convolutions as a molecular tokenizer and Transformer self-attention for cross-molecule interaction. The same backbone achieves macro-AUROC 0.930 on olfactory mixture prediction and Pearson R 0.937 on excess enthalpy prediction with direct end-to-end training. Systematic ablation confirms that removing molecular boundaries and enabling cross-molecule attention are both necessary, with bidirectional self-attention providing the decisive gain.</jats:p>