Liquid–liquid phase separation occurs in young mammalian eye lenses and in concentrated solutions of isolated eye lens proteins, and has been linked to some forms of cataract. Here we study theoretically the protein compositions and cloud temperatures of two separated equilibrium phases that form out of concentrated mixtures of model proteins, chosen to have properties similar to those which reproduce experimental data on mixtures of two of the prevalent mammalian eye lens proteins, γ- and α-crystallin. We use a thermodynamic perturbation theory that has previously been shown to provide a quantitative model for key features of the experimentally observed neutron scattering, phase boundary and tie line data, and that is also consistent with corresponding model, coarse-grained molecular dynamics simulations. In so doing we find an extremely sensitive dependence of protein partitioning on mutual attraction that is likely to have implications for many other protein, colloid, and other soft condensed matter systems. Previously, we found that a model square well attraction between the proteins of well depth uαγ ≈ 0.5 kBT protects concentrated γ-α mixtures against thermodynamic instability and is thus essential for their transparency. Furthermore, the dependence of the mixture phase separation on uαγ was found to be highly non-monotonic, in that either weakening or increasing uαγ by 0.5 kBT can lead to considerably enhanced phase separation that occurs at much higher temperatures. In the present work we show that the compositions of the separated protein phases are even more dramatically sensitive to the magnitude of uαγ. Specifically, increasing uαγ by just 0.2 kBT can change the phase separation of α–γ mixtures from one that is primarily compositional in nature to one of protein density separation, in which the two phases in equilibrium differ principally in overall protein concentration. Further, for the square-well widths investigated, we find that the phase separation properties change relatively rapidly in response to changes in square well depth, in comparison with their response to changes in the diameter ratio of the model proteins. We discuss potential ways in which sensitive connections between changes in molecular attraction and their macroscopic consequences, a hallmark of concentrated liquid mixtures, can lead to potential molecular mechanisms for hereditary and other forms of cataract, and can be applied to other colloidal and physiological systems.