Globular proteins, the work house molecules of life, are linear chains of amino-acids which fold rapidly and reproducibly into their native conformations. An important goal in the protein problem is the prediction of the native state structure given the amino acid sequence of a protein. A vast semplification arises because the number of distinct folds adopted by proteins is limited to just a few thousand. This has led to a powerful method, called threading, for predicting the native state of a sequence: one mounts the sequence on candidate structures obtained as pieces of all known folds and determines the best fit structure through a scoring function which provides a measure of the interactions among the aminoacids. Here we will present a maximum entropy approach for inferring amino-acid interactions in proteins subject to constraints pertaining to the mean numbers of various types of equilibrium contacts. We will illustrate the method on simply models where we obtained very promising results. We will also show the flexibility of the method that can work very well even when the mean numbers of contacts are not known exactly. These results are suggesting that our maxent approach has the possibilty of being successfull when applied to real proteins