Why do people suggest that translated sequences be used to search for relatives in databanks?

Why Protein Sequence

Protein sequences are composed of a 20 aa alphabet determined by 61 degenerate codons. When the DNA sequences are translated into 21 different types of codons (20 aa and a terminator), the information is sharpened up considerably. The 'wrong-frame' information is discarded, and third-base degeneracies are consolidated. All in all, the signal-to-noise ratio is greatly improved for the specific purpose of identifying protein relatives. It is accepted that convergence phenomena in aa sequences are very rare and thus aa similarity almost always means homology. Furthermore, aa sequences may still show a similarity derived from common folding patterns and function of the proteins, even while their coding DNA sequences might have strongly diverged due to other selective pressures existent at the genome level (e.g., G+C pressure, preferential usage of synonymous codons, etc.). Protein evolution is governed by the constraint of maintaining a characteristic fold which enables some function. Thus, it is possible to infer relationships between proteins that last shared a common ancestor 1-2.5 billion years ago by conducting protein searches, doubling the lookback time obtained performing DNA database searches.

link