Dunker Ak

Identifying disordered regions in proteins from amino acid sequence

A rule-based and several neural network predictors are developed for identifying disordered regions in proteins. The rule-based predictor, which relied on the observation that disordered regions contain few aromatic amino acids, was suitable only for very long disordered regions, whereas the neural network predictors were developed separately for short-, medium-, and long-disordered regions (LDRs), The out-of-sample prediction accuracies on a residue-by-residue basis ranged from 69 to 74% for the neural network predictors when applied to the same length class, but fell to 59 to 67% when applied to different length classes. Testing the rule-based predictor on a residue-by-residue basis using out-of-sample LDRs gave a success rate of 70%. Application of both the rule-based and LDR neural network predictors to large databases of protein sequences provide strong evidence that disordered regions are very common in nature. These results are consistent with our recent proposal that disordered regions are crucial for the evolution of molecular recognition.

THE PROTEIN NON-FOLDING PROBLEM: AMINO ACID DETERMINANTS OF INTRINSIC ORDER AND DISORDER

To investigate the determinants of protein order and disorder, three primary and one derivative database of intrinsically disordered proteins were compiled. The segments in each primary database were characterized by one of the following: X-ray crystallography, nuclear magnetic resonance (NMR), or circular dichroism (CD). The derivative database was based on homology. The three primary disordered databases have a combined total of 157 proteins or segments of length.30 with 18,010 residues, while the derivative database contains 572 proteins from 32 families with 52,688 putatively disordered residues. For the four disordered databases, the amino acid compositions were compared with those from a database of ordered structure. Relative to the ordered protein, the intrinsically disordered segments in all four databases were significantly depleted in W, C, F, I, Y, V, L and N, significantly enriched in A, R, G, Q, S, P, E and K, and inconsistently different in H, M, T, and D, suggesting that the first set be called order-promoting and the second set disorder-promoting. Also, 265 amino acid properties were ranked by their ability to discriminate order and disorder and then pruned to remove the most highly correlated pairs. The 10 highest-ranking properties after pruning consisted of 2 residue contact scales, 4 hydrophobicity scales, 3 scales associated with.-sheets and one polarity scale. Using these 10 properties for comparisons of the 3 primary databases suggests that disorder in all 3 databases is very similar, but with those characterized by NMR and CD being the most similar, those by CD and X-ray being next, and those by NMR and X-ray being the least similar.

Folding minimal sequences: the lower bound for sequence complexity of globular proteins.

Alphabet size and informational entropy, two formal measures of sequence complexity, are herein applied to two prior studies on the folding of minimal proteins. These measures show a designed four‐helix bundle to be unlike its natural counterparts but rather more like a coiled‐coil dimer. Segments from a simplified sarc homology 3 domain and more than 2u2008000u2008000 segments from globular proteins both have lower bounds for alphabet size of 10 and for entropy near 2.9. These values are therefore suggested to be necessary and sufficient for folding into globular proteins having both rigid side chain packing and biological function.

Methods for improving protein disorder prediction

In this paper we propose several methods for improving prediction of protein disorder. These include attribute construction from protein sequence, choice of classifier and postprocessing. While ensembles of neural networks achieved the higher accuracy, the difference as compared to logistic regression classifiers was smaller than 1%. Bagging of neural networks, where moving averages over windows of length 61 were used for attribute construction, combined with postprocessing by averaging predictions over windows of length 81 resulted in 82.6% accuracy for a larger set of ordered and disordered proteins than used previously. This result was a significant improvement over previous methodology, which gave an accuracy of 70.2%. Moreover, unlike the previous methodology, the modified attribute construction allowed prediction at protein ends.

The benefit of intrinsic disorder information in neural network prediction of calmodulin binding targets

Calmodulin is an important calcium dependent signaling protein found in all eukaryotic cells. Binding calcium enables calmodulin to bind its targets: basic, amphipathic /spl infin/-helices. Such binding regulates the activities of many proteins. Because calmodulin wraps completely around the target helix upon binding, it is hypothesized that disorder of a target helix is an important feature of this process. We have used several sequence derived features of calmodulin binding targets, including intrinsic order/disorder predictions, to construct neural networks based on permutations of three or more of these features. The resulting networks demonstrate that the addition of intrinsic order/disorder information always increases the performance of a given neural network predictor. The best predictor generated has a performance of 87.8% true positive prediction and 87.2% true negative prediction.