Abstract A general model of protein folding is proposed that links together concepts from molecular physics, thermodynamics and chemical kinetics. The model is based on a description of the potential hypersurface through a three-level taxonomy: conformational substates, conformational states, and free-energy levels. The required parameters are either estimated or taken from experimental or computational literature data. The free-energy barriers between conformational states are shown to be essentially negentropic in the folding direction. A consistent picture emerges, which answers Levinthal's paradox. It consists in a biased random walk with a very high degree of percolation. Finally, the model leads to a sketch of a possible algorithm for an ab initio simulation of protein folding.