Hemoglobin is a complex system that undergoes conformational changes in response to oxygen, allosteric effectors, mutations, and environmental changes. Here, we study allostery and polymerization of hemoglobin and its variants by application of two previously described methods: (i) AllosMod for simulating allostery dynamics given two allosterically related input structures and (ii) a machine-learning method for dynamics- and structure-based prediction of the mutation impact on allostery (Weinkam et al. J. Mol. Biol. 2013, 425, 647-661), now applicable to systems with multiple coupled binding sites, such as hemoglobin. First, we predict the relative stabilities of substates and microstates of hemoglobin, which are determined primarily by entropy within our model. Next, we predict the impact of 866 annotated mutations on hemoglobin's oxygen binding equilibrium. We then discuss a subset of 30 mutations that occur in the presence of the sickle cell mutation and whose effects on polymerization have been measured. Seven of these HbS mutations occur in three predicted druggable binding pockets that might be exploited to directly inhibit polymerization; one of these binding pockets is not apparent in the crystal structure, but only in structures generated by AllosMod. For the 30 mutations, we predict that mutation-induced conformational changes within a single tetramer tend not to significantly impact polymerization; instead, these mutations more likely impact polymerization by directly perturbing a polymerization interface. Finally, our analysis of allostery allows us to hypothesize why hemoglobin evolved to have multiple subunits and a persistent low frequency sickle cell mutation.