Proteins can sense signals and--in a process called allostery - transmit information to distant sites. Such information is often not encoded by a protein's average structure, but rather by its dynamics in a way that remains unclear. We show that maximum information tree networks learned from microseconds-long molecular dynamics simulations provide mechanistically-detailed maps of information transmission within proteins in a ligand- and mutation-sensitive manner. On a PDZ domain and the entire human steroid receptor family, these networks quantitatively predict functionally relevant experimental datasets spanning multiple scales, including allosteric sensitivity across a saturation mutagenesis library, calorimetric binding entropies, and phylogenetic trees. These results suggest that a sparse network of entropic couplings encodes the dynamics-to-function map; functional reprogramming and diversification by ligand binding and evolution can modify this network without changing protein structure.