Pretrained foundation models offer new representations for molecular structures but have predominantly been applied to single-component property prediction. Here, we develop a concentration-weighted average of the latent representations of graph neural network and transformer-based molecular foundation models to predict the critical micelle concentrations of single surfactant, biphasic separation of binary mixtures, and liposome formation in 7-component amphiphile mixtures. Simple random forest and feed-forward neural network models trained with these representations are comparable to or exceed prior work using bespoke graph neural network methods, physicochemical features, and binary fingerprinting. High-throughput automated experiments validate the ability of our models to extrapolate to novel surfactant mixtures and to guide the iterative exploration of liposome design.
Liu et al. (Tue,) studied this question.