ESTIMATING PROTEIN FUNCTION FROM COMBINATORIAL SEQUENCE DATA USING DECISION ALGORITHMS AND NEURAL NETWORKS

Correlations between protein sequences and phenotypes were explored us ing databases of combinatorial cassette mutants of pigment-protein com plexes. Heuristically formulated decision algorithms and computer impl emented neural networks were compared to determine their accuracy in c lassification of phenotypic categories. For the databases examined, de cision algorithms employing very simple rules were able to properly cl assify mutants 80-84% of the time, based only on the amino acid sequen ce of the mutagenized region. Such decision algorithms did not require the formulation of any rules that involved site-to-site interactions, but rather, performed well based on the stringency of specific critic al sites in the protein that accept only a restricted set of amino aci ds. In some cases, neural networks scored almost 10% higher than decis ion algorithms on the same databases (i.e., 94%). However, the success of the primitive decision algorithms and perceptrons at sorting seque nces into categories suggests that linear effects predominate in the c lassification of a mutant's phenotype. Such methods should be generall y applicable to the broad spectrum of databases that are currently bei ng generated in combinatorial chemistry and biology experiments. (C) 1 994 Wiley-Liss, Inc.