It is a common task in biology to predict the function of a protein from its proteic sequence. A classical approach is to use predictive models which are estimated from a set of aligned proteic sequences sharing the same function. The models assigns a score to amino acids for each column of the alignment which tells how conserved they are in this column. Such score can be interpreted as expressing how important for the function each amino acids are at a given position. In this report we propose an approach that takes distances between pairs of sequences into account for the computation of the conservation score of amino acids at each position.