A novel approach to hierarchical peptide–protein and protein–protein docking is described and evaluated. Modeling procedure starts from a reduced space representation of proteins and peptides. Polypeptide chains are represented by strings of alpha-carbon beads restricted to a fine-mesh cubic lattice. Side chains are represented by up to two centers of interactions, corresponding to beta-carbons and the centers of mass of the remaining portions of the side groups, respectively. Additional pseudoatoms are located in the centers of the virtual bonds connecting consecutive alpha carbons. These pseudoatoms support a model of main-chain hydrogen bonds. Docking starts from a collection of random configurations of modeled molecules. Interacting molecules are flexible; however, higher accuracy models are obtained when the conformational freedom of one (the larger one) of the assembling molecules is limited by a set of weak distance restraints extracted from the experimental (or theoretically predicted) structures. Sampling is done by means of Replica Exchange Monte Carlo method. Afterwards, the set of obtained structures is subject to a hierarchical clustering. Then, the centroids of the resulting clusters are used as scaffolds for the reconstruction of the atomic details. Finally, the all-atom models are energy minimized and scored using classical tools of molecular mechanics. The method is tested on a set of macromolecular assemblies consisting of proteins and peptides. It is demonstrated that the proposed approach to the flexible docking could be successfully applied to prediction of protein–peptide and protein–protein interactions. The obtained models are almost always qualitatively correct, although usually of relatively low (or moderate) resolution. In spite of this limitation, the proposed method opens new possibilities of computational studies of macromolecular recognition and mechanisms of assembly of macromolecular complexes.