We propose in this paper a modular learning environment for protein modeling. In this system, the protein modeling problem is tackled in two successive phases. First, partial structural informations are determined via numerical learning techniques. Then, in the second phase, the multiple available informations are combined in paaem matching searches via dynamic programming. It is shown on real problems that various protein structure predictions can be improved in this way, such as secondary structure prediction, alignment of weakly homologous protein sequences or protein model evaluations.