Averaging interaction energies over homologs improves protein fold recognition in gapless threading

Protein structure prediction is limited by the inaccuracy of the simplified energy functions necessary for efficient sorting over many conformations, It was recently suggested (Finkelstein, Phys Rev Lett 1998;80:4823-4825) th at these errors can be reduced by energy averaging over a set of homologous sequences. This conclusion is confirmed in this study by testing protein s tructure recognition in gapless threading. The accuracy of recognition was estimated by the Z-score values obtained in gapless threading tests. For th reading, we used 20 target proteins, each having from 20 to 70 homologs tak en from the HSSP sequence base. The energy of the native structures was com pared with the energy from 34 to 75 thousand of alternative structures gene rated by threading. The energy calculations were done with our recently dev eloped C-alpha atom-based phenomenological potentials. We show that averagi ng of protein energies over homologs reduces the Z-score from similar to -6 .1 (average Z-score for individual chains) to similar to -8.1. This means t hat a correct fold can be found among 3*10(9) random folds in the first cas e and among 3*10(15) in the second. Such increase in selectivity is importa nt for recognition of protein folds. (C) 1999 Wiley-Liss, Inc.