The aim of this work is to better understand the factors and mechanisms leading to the supramolecular organization of collagen layers adsorbed on polymers. Native collagen adsorption on polystyrene (PS) and plasma-oxidized polystyrene (PSox) substrates revealed that the adsorbed layer consists in two parts: a dense and thin sheet (~ 10 nm) in which fibrils are formed, as revealed by atomic force microscopy, and an overlying thick layer (~ 200 nm) which contains protruding molecules, as revealed by quartz crystal microbalance with dissipation monitoring. The protruding molecules are in low density but modify noticeably the local viscosity. Faster and enhanced fibril formation takes place on hydrophobic compared to hydrophilic substrate. As a result of drastic thermal denaturation, the ability of collagen to assemble into fibrils is lost and the number of protruding molecules responsible for higher viscosity is reduced. Radiochemical measurements showed that collagen molecules are more easily displaced when adsorbed on a hydrophobic substrate compared to a hydrophilic substrate. This may explain why fibril formation occurs more readily on the more hydrophobic substrate, but is in contrast with higher surface affinity. The possible explanation of this paradox by the quick formation of a dense layer of collagen molecules having a smaller number of contact points with a very hydrophobic surface could not be demonstrated by a comparison of adsorption procedures. Comparing different collagen sources revealed various modes of aggregation with different characteristics regarding size and order (large fibers in solution, smaller fibrils to featureless underneath layer in the adsorbed phase). Moreover, collagen aggregation in the solution is a process competing with adsorption: more aggregated solutions behave like less concentrated solutions regarding the adsorbed amount and fibril formation in the adsorbed phase. It must be emphasized that interpretation of the QCM-D data, which is based on fitting physical quantities according to a model, has to be performed very carefully, and requires the examination of the sensitivity of the fitted data to the fitting parameters.