In a recently proposed contextual alignment model, efficient algorithms exist for global and local pairwise alignment of protein sequences. Preliminary results obtained for biological data are very promising. Our main motivation was to adopt the idea of context dependency to the multiple‐alignment setting. To this aim the relaxation of the model was developed (we call this new model averaged contextual alignment) and a new family of amino acids substitution matrices are constructed. In this paper we present a contextual multiple‐alignment algorithm and report the outcomes of experiments performed for the BAliBASE test set. The contextual approach turned out to give much better results for the set of sequences containing orphan genes.