Amino acid substitution matrices from an artificial neural network model

Ideal amino acid exchange forms for approximating substitution matrices

We have analyzed 29 published substitution matrices (SMs) and five statistical protein contact potentials (CPs) for comparison. We find that popular, 'classical' SMs obtained mainly from sequence alignments of globular proteins are mostly correlated by at least a value of 0.9. The BLOSUM62 is the central element of this group. A second group includes SMs derived from alignments of remote homologs or transmembrane proteins. These matrices correlate better with classical SMs (0.8) than among themselves (0.7). A third group consists of intermediate links between SMs and CPs - matrices and potentials that exhibit mutual correlations of at least 0.8. Next, we show that SMs can be approximated with a correlation of 0.9 by expressions c(0) + x(i)x(j) + y(i)y(j) + z(i)z(j), 1<or= i, j <or= 20, where c(0) is a constant and the vectors (x(i)), (y(i)), (z(i)) correlate highly with hydrophobicity, molecular volume and coil preferences of amino acids, respectively. The present paper is the continuation of our work (Pokarowski et al., Proteins 2005;59:49-57), where similar approximation were used to derive ideal amino acid interaction forms from CPs. Both approximations allow us to understand general trends in amino acid similarity and can help improve multiple sequence alignments using the fast Fourier transform (MAFFT), fast threading or another methods based on alignments of physicochemical profiles of protein sequences. The use of this approximation in sequence alignments instead of a classical SM yields results that differ by less than 5%. Intermediate links between SMs and CPs, new formulas for approximating these matrices, and the highly significant dependence of classical SMs on coil preferences are new findings.

A collection of amino acid replacement matrices derived from clusters of orthologs

Sequence divergence among orthologous proteins was characterized with 34 amino acid replacement matrices, sequence context analysis, and a phylogenetic tree. The model was trained on very large datasets of aligned protein sequences drawn from 15 organisms including protists, plants, Dictyostelium, fungi, and animals. Comparative tests with models currently used in phylogeny, i.e., with JTT+gamma+/-F and WAG+gamma+/-F, made on a test dataset of 380 multiple alignments containing protein sequences from all five of the major taxonomic groups mentioned, indicate that our model should be preferred over the JTT+gamma+/-F and WAG+gamma+/-F models on datasets similar to the test dataset. The strong performance of our model of orthologous protein sequence divergence can be attributed to its ability to better approximate amino acid equilibrium frequencies to compositions found in alignment columns.

Statistical Evaluation of Local Alignment Features Predicting Allergenicity Using Supervised Classification Algorithms

BACKGROUND

Recently, two promising alignment-based features predicting food allergenicity using the k nearest neighbor (kNN) classifier were reported. These features are the alignment score and alignment length of the best local alignment obtained in a database of known allergen sequences.

METHODS

In the work reported here a much more comprehensive statistical evaluation of the potential of these features was performed, this time for the prediction of allergenicity in general. The evaluation consisted of the following four key components. (1) A new high quality database consisting of 318 carefully selected, non-redundant allergens and 1,007 sequences carefully selected to be non-allergens. (2) Three different supervised algorithms: the kNN classifier, the Bayesian linear Gaussian classifier, and the Bayesian quadratic Gaussian classifier. (3) A large set of local alignment procedures defined using the FASTA3 alignment program by means of a wide range of different parameter settings. (4) Novel performance curves, alternative to conventional receiver-operating characteristic curves, to display not only average behaviors but also statistical variations due to small data sets.

RESULTS

The linear Gaussian classifier proved most useful among the tested supervised machine learning algorithms, closely followed by the quadratic Gaussian equivalent and kNN. The overall best classification results were obtained with a novel feature vector consisting of the combined alignment scores derived from local alignment procedures using different substitution matrices.

CONCLUSIONS

The models reported here should be useful as a part of an integrated assessment scheme for potential protein allergenicity and for future comparisons with alternative bioinformatic approaches.

Amino acid encoding schemes from protein structure alignments: multi-dimensional vectors to describe residue types

Bioinformatic software has used various numerical encoding schemes to describe amino acid sequences. Orthogonal encoding, employing 20 numbers to describe the amino acid type of one protein residue, is often used with artificial neural network (ANN) models. However, this can increase the model complexity, thus leading to difficulty in implementation and poor performance. Here, we use ANNs to derive encoding schemes for the amino acid types from protein three-dimensional structure alignments. Each of the 20 amino acid types is characterized with a few real numbers. Our schemes are tested on the simulation of amino acid substitution matrices. These simplified schemes outperform the orthogonal encoding on small data sets. Using one of these encoding schemes, we generate a colouring scheme for the amino acids in which comparable amino acids are in similar colours. We expect it to be useful for visual inspection and manual editing of protein multiple sequence alignments.