The ability of proteins to fold to well defined compact structures is one of the most remarkable examples of the effect of natural selection on biological molecules. To understand their properties, including the stability, the mechanism of folding, and the possibilities of misfolding and association, it is necessary to know the protein free energy landscape. We use NMR data as restraints in a Monte Carlo sampling procedure to determine the ensemble of structures populated by human ␣-lactalbumin in the presence of increasing concentrations of urea. The ensembles of structures that represent the partially folded states of the protein show that two structural cores, corresponding to portions of the ␣ and  domains of the native protein, are preserved even when the native-like interactions that define their existence are substantially weakened. Analysis of the network of residual contacts reveals the presence of a complex interface region between the two structural cores and indicates that the development of specific interactions within this interface is the key step in achieving the native structure. The relative probabilities of the conformations determined from the NMR data are used to construct a coarse-grained free energy landscape for ␣-lactalbumin in the absence of urea. The form of the landscape, together with the existence of distinct cores, supports the concept that robustness and modularity are the properties that make possible the folding of complex proteins.T he folding of small single-domain proteins is now relatively well understood as a result of recent progress in describing the structural properties of the transition state ensembles of a range of representative proteins that fold with two-state kinetics (1-3). A particularly important conclusion of these studies is that the protein fold is encoded by the sequence through key nucleation sites (1,4,5). By contrast, the folding mechanism of larger proteins, in which the population of one or more partially folded states can be significant, still remains to be resolved. We show here that it is possible to use experimental data to determine the structures and relative free energies of partially folded states for such more complex systems, which include the vast majority of proteins. These results provide the information necessary for the construction of the accessible portions of the free energy surface (3, 6) or ''landscape'' (7) from which the thermodynamic description of the folding process can be determined. The present approach incorporates experimental restraints into the energy function used in Monte Carlo simulations to bias the conformations that are sampled to the portions of space compatible with the measurements (5). In this way the conformation space accessible to rather complex proteins can be fully characterized. The work described here thus extends the theoretical analyses of energy landscapes for smaller proteins (see, for example, ref. 8).To illustrate the approach we apply it to the extensively studied protein human ␣-lactalbumin (HLA)...