Prediction of β-turn in proteins using neural networks

Protein Engineering vol.3 no.5 pp.459-460, 1990 Corrigendum Prediction of /3-turn in proteins using neural networks by M.J.McGregor, T.P.Flores and MJ.E.Sternberg (d) 1 Protein Engineering, 2, 521-526 (1989) Reference WUmot.C.M. and ThomtonJ.M. (1990). Prot. Engng, in press. Table I . Results for the neural network with (a, c) eight hidden units and (b, d) no hidden units C No. of predictions Correct Incorrect (a) 1 2 3 4 5 6 Mean 37 38 43 45 44 44 41.8 105 110 105 88 114 103 104.2 26.1 25.1 29.1 33.8 27.8 29.9 28.7 0.173 0.173 0.210 0.247 0.204 0.219 0.204 (b) 1 2 3 4 5 6 Mean 44 38 43 46 36 38 40.8 116 115 120 107 114 120 115.3 27.5 24.8 26.4 30.1 24.0 24.1 26.1 0.201 0.167 0.190 0.225 0.156 0.161 0.183 67 76 66 52.8 48.6 55.4 0.393 0.366 0.426 (c) 1 2 3 75 72 82 © Oxford University Press 2 3 4 5 6 Mean 78 87 77 78.5 55 71 70 67.5 58.6 55.1 52.4 53.8 0.442 0.432 0.404 0.411 97 81 91 89 81 88 87.8 63 72 72 64 69 70 68.3 60.6 52.9 55.8 58.2 54.0 55.7 56.2 0.475 0.403 0.439 0.455 0.406 0.427 0.434 (c) and (d) count a prediction which is displaced by one or two residues as correct. C is Matthew's correlation coefficient. Table n . Weights and biases for the network with no hidden units trained on the whole data set Position 1 2 3 4 Type I turns (bias = -1.449189) ALA -0.243087 -0.142091 -0.406034 ARG 0.141068 0.930008 ASN 0.150789 0.683654 ASP 0.300384 0.640047 CYS -0.058365 -0.090925 GLN -0.270385 -0.232212 GLU 0.224860 -0.180140 GLY -0.824957 0.499984 HIS -0.573601 -0.578029 ILE -0.331232 -0.409766 LEU -0.340384 -0.136094 LYS 0.099007 -0.543506 MET 0.027302 -0.060225 PHE -0.647743 -0.329642 PRO -0.593015 0.534037 SER 0.624545 0.136937 THR 0.111164 -0.004853 TRP -0.209019 -0.381581 TYR -0.632678 -0.698976 VAL -0.857536 -0.109024 0.176669 0.165416 0.856229 0.145938 -0.046035 0.106304 -1.015797 -0.214866 -0.376639 -0.385285 0.112464 -0.081694 -0.217963 -0.417375 0.278746 0.226478 0.015674 -0.237570 -0.627830 -0.505421 -0.223996 0.062782 0.254206 0.034707 -0.078962 -0.089487 0.764431 -0.101168 -0.465837 -0.015671 -0.280728 0.225133 0.374096 -0.871687 -0.193818 0.028503 0.253646 -0.036036 -0.489945 Typen turns (bias = -2.785452) ALA 0.080881 0.200580 ARG -0.248750 -0.161799 ASN -0.477788 -0.091148 -0.298290 ASP -0.197392 -0.367305 CYS -0.197055 0.191025 GLN 0.143987 -0.269154 GLU -0.057999 -0.192140 GLY -0.291102 -0.690576 HIS 0.162537 -0.269901 ILE -0.461945 -0.035772 LEU -0.350905 -0.392525 LYS -0.250948 0.206530 MET 0.222486 0.163443 PHE 0.072575 0.403400 PRO 1.041997 0.067392 SER 0.009001 -0.019577 THR -0.518423 -0.166333 TRP -0.094064 -0.139075 TYR 0.008944 -0.243054 VAL -0.311138 -0.399618 -0.310081 0.183620 -0.253403 -0.265602 -0.024288 -0.273195 2.070249 -0.098164 -0.416577 -0.321496 -0.223949 -0.084831 -0.301763 -0.029515 -0.384652 -0.466651 -0.167270 -0.144744 -0.543210 -0.163486 0.239793 -0.196367 0.155113 -0.017332 0.097946 -0.174807 -0.201066 -0.223062 -0.413987 -0.451829 0.116696 -0.203805 -0.454057 -0.469359 -0.188710 -0.129632 -0.306867 -0.223803 -0.337433 459 An error was found in the original data set which resulted in some /S-tum sequences being classified as non-turns. The results using a corrected data set are presented here. The method is unchanged from before with the exception of the sampling of the data. Previously, the ratio of output types produced after the training stage was very sensitive to the ratio of sampling of the input-output pairs. In the present method the sampling ratio of the input—output pairs is adjusted during training to give the optimum ratio of outputs. For example, if the network predicts a proportion of type I turns which is too high, the number of type I turns sampled for each iteration during training is reduced, and vice versa (similarly for the other types). Two neural networks were developed, one with no hidden units and one with eight hidden units. Table I gives the revised accuracies of prediction and Table II gives the revised weights and biases for the network with no hidden units. Parts (a) and (b) of Table I give the accuracies for predictions with eight and with no hidden units respectively; parts (c) and (d) give the results for eight and no hidden units in which a prediction is considered correct if the turn location is displaced by one or two residues. These values can be compared with the accuracy of prediction by the revised method of Wilmot and Thornton (1990) that obtained 26% accuracy on a single test data set for a correct location of a turn and 43 % if the turn position could be displaced by one or two residues. Thus the neural network approach is generally more accurate than a sample statistical treatment. Set 4 5 6 Mean Corrigendum Table E[. Continued Position 1 4 Non-specific turns (bias = -1.664808) ALA -0.480207 -0.314464 ARG 0.174041 0.180753 ASN -0.012317 0.561298 ASP 0.060205 0.375278 CYS 0.135668 -0.182610 GLN -0.300550 0.109253 GLU 0.131456 -0.038217 GLY -0.223319 0.786988 HIS -0.067205 -0.325387 ILE 0.210573 0.044735 LEU -0.401093 -0.490781 LYS 0.120655 0.155493 MET 0.160298 -0.320116 PHE -0.598026 -0.661481 PRO -0.243341 -0.496534 SER 0.037685 0.375594 THR -0.250224 -0.154122 TRP -0.096621 -0.152782 TYR 0.193099 0.140795 VAL -0.310746 -0.248186 -0.518980 0.230967 0.836332 0.270065 -0.373159 0.056229 -0.172797 0.850369 0.224775 -0.424143 -0.544969 -0.378395 -0.153883 -0.367062 -0.448775 0.372670 -0.469267 -0.324497 0.142521 -0.754627 -0.288990 -0.134497 0.087782 -0.357116 -0.369412 -0.194394 -0.301124 -0.243850 -0.284595 -0.145925 -0.096433 -0.172368 -0.254940 -0.229200 0.803885 0.343048 -0.047064 -0.320580 0.068894 -0.197625 Non turns ALA ARG ASN ASP CYS GLN GLU GLY HIS ILE LEU LYS MET PHE PRO SER THR TRP TYR VAL 0.324966 -0.394566 -1.265182 -0.863000 0.126676 -0.173760 -0.182687 -1.497146 -0.365602 0.721478 0.581402 0.078282 0.259375 0.359250 0.670346 -0.581634 0.143260 0.230658 -0.204291 1.077674 0.398388 0.039500 -0.138101 -0.187986 0.065920 -0.223480 0.162262 -0.761628 0.332103 0.484412 0.137195 0.021264 -0.080825 0.000183 0.020226 -0.265058 -0.030253 0.147034 -0.167003 0.473166 460 (bias = 0.836405) 0.461901 0.085202 0.223159 -0.198649 -0.755655 -0.567240 -0.545307 -0.564479 -0.513999 0.322215 0.124781 0.062026 0.019090 -0.204367 -0.269952 -0.002373 -0.118517 0.421921 0.337538 0.461370 0.442933 0.649051 0.229464 -0.332395 0.248576 0.513678 0.376568 1.008598 0.203038 -0.656973 -0.540746 -0.670116 -0.051439 0.069756 0.246985 0.606565 0.021291 0.409969 0.698816 0.808408 3 2  (...truncated)