Protein secondary structure and homology by neural networks. The alpha-helices in rhodopsin

Artificial Intelligence Learns Protein Prediction

From AlphaGO over StableDiffusion to ChatGPT, the recent decade of exponential advances in artificial intelligence (AI) has been altering life. In parallel, advances in computational biology are beginning to decode the language of life: AlphaFold2 leaped forward in protein structure prediction, and protein language models (pLMs) replaced expertise and evolutionary information from multiple sequence alignments with information learned from reoccurring patterns in databases of billions of proteins without experimental annotations other than the amino acid sequences. None of those tools could have been developed 10 years ago; all will increase the wealth of experimental data and speed up the cycle from idea to proof. AI is affecting molecular and medical biology at giant steps, and the most important might be the leap toward more powerful protein design.

SignalP: The Evolution of a Web Server

SignalP ( https://services.healthtech.dtu.dk/services/SignalP-6.0/ ) is a very popular prediction method for signal peptides, the intrinsic signals that make proteins secretory. The SignalP web server has existed since 1995 and is now in its sixth major version. In this historical account, we (three authors who have taken part in the entire journey plus the first author of the latest version) describe the differences between the versions and discuss the various decisions taken along the way.

Deep Learning for Genomics: From Early Neural Nets to Modern Large Language Models

The data explosion driven by advancements in genomic research, such as high-throughput sequencing techniques, is constantly challenging conventional methods used in genomics. In parallel with the urgent demand for robust algorithms, deep learning has succeeded in various fields such as vision, speech, and text processing. Yet genomics entails unique challenges to deep learning, since we expect a superhuman intelligence that explores beyond our knowledge to interpret the genome from deep learning. A powerful deep learning model should rely on the insightful utilization of task-specific knowledge. In this paper, we briefly discuss the strengths of different deep learning models from a genomic perspective so as to fit each particular task with proper deep learning-based architecture, and we remark on practical considerations of developing deep learning architectures for genomics. We also provide a concise review of deep learning applications in various aspects of genomic research and point out current challenges and potential research directions for future genomics applications. We believe the collaborative use of ever-growing diverse data and the fast iteration of deep learning models will continue to contribute to the future of genomics.

Deep learning for protein secondary structure prediction: Pre and post-AlphaFold

This paper aims to provide a comprehensive review of the trends and challenges of deep neural networks for protein secondary structure prediction (PSSP). In recent years, deep neural networks have become the primary method for protein secondary structure prediction. Previous studies showed that deep neural networks had uplifted the accuracy of three-state secondary structure prediction to more than 80%. Favored deep learning methods, such as convolutional neural networks, recurrent neural networks, inception networks, and graph neural networks, have been implemented in protein secondary structure prediction. Methods adapted from natural language processing (NLP) and computer vision are also employed, including attention mechanism, ResNet, and U-shape networks. In the post-AlphaFold era, PSSP studies focus on different objectives, such as enhancing the quality of evolutionary information and exploiting protein language models as the PSSP input. The recent trend to utilize pre-trained language models as input features for secondary structure prediction provides a new direction for PSSP studies. Moreover, the state-of-the-art accuracy achieved by previous PSSP models is still below its theoretical limit. There are still rooms for improvement to be made in the field.

CSI-LSTM: a web server to predict protein secondary structure using bidirectional long short term memory and NMR chemical shifts

Protein secondary structure provides rich structural information, hence the description and understanding of protein structure relies heavily on it. Identification or prediction of secondary structures therefore plays an important role in protein research. In protein NMR studies, it is more convenient to predict secondary structures from chemical shifts as compared to the traditional determination methods based on inter-nuclear distances provided by NOESY experiment. In recent years, there was a significant improvement observed in deep neural networks, which had been applied in many research fields. Here we proposed a deep neural network based on bidirectional long short term memory (biLSTM) to predict protein 3-state secondary structure using NMR chemical shifts of backbone nuclei. While comparing with the existing methods the proposed method showed better prediction accuracy. Based on the proposed method, a web server has been built to provide protein secondary structure prediction service.

Can artificial neural replicators be useful for studying RNA replicators?

Here, I discuss the usefulness of the application of special artificial neural systems - neural replicators - to study viroids - small pathogens that are short replicating RNA sequences. Using special representations of nucleotide sequences in the form of two sequences with binary components - these two sequences are incomplete representations of the same nucleotide sequence - I show that these neural systems of different sizes are replicated in a special way on them. This allows us to extract some useful information about viroids and their structure, motifs, and relationships. This study is only the first attempt to use neural replicators to analyze genetic data.

Protein Secondary Structure Prediction: A Review of Progress and Directions

Background:

Over the last few decades, a search for the theory of protein folding has grown into a full-fledged research field at the intersection of biology, chemistry and informatics. Despite enormous effort, there are still open questions and challenges, like understanding the rules by which amino acid sequence determines protein secondary structure.

Objective:

In this review, we depict the progress of the prediction methods over the years and identify sources of improvement.

Methods:

The protein secondary structure prediction problem is described followed by the discussion on theoretical limitations, description of the commonly used data sets, features and a review of three generations of methods with the focus on the most recent advances. Additionally, methods with available online servers are assessed on the independent data set.

Results:

The state-of-the-art methods are currently reaching almost 88% for 3-class prediction and 76.5% for an 8-class prediction.

Conclusion:

This review summarizes recent advances and outlines further research directions.

Third generation antibody discovery methods: in silico rational design

Owing to their outstanding performances in molecular recognition, antibodies are extensively used in research and applications in molecular biology, biotechnology and medicine. Recent advances in experimental and computational methods are making it possible to complement well-established in vivo (first generation) and in vitro (second generation) methods of antibody discovery with novel in silico (third generation) approaches. Here we describe the principles of computational antibody design and review the state of the art in this field. We then present Modular, a method that implements the rational design of antibodies in a modular manner, and describe the opportunities offered by this approach.

Deep Learning in Biomedical Data Science

Since the 1980s, deep learning and biomedical data have been coevolving and feeding each other. The breadth, complexity, and rapidly expanding size of biomedical data have stimulated the development of novel deep learning methods, and application of these methods to biomedical data have led to scientific discoveries and practical solutions. This overview provides technical and historical pointers to the field, and surveys current applications of deep learning to biomedical data organized around five subareas, roughly of increasing spatial scale: chemoinformatics, proteomics, genomics and transcriptomics, biomedical imaging, and health care. The black box problem of deep learning methods is also briefly discussed.

Sixty-five years of the long march in protein secondary structure prediction: the final stretch?

Protein secondary structure prediction began in 1951 when Pauling and Corey predicted helical and sheet conformations for protein polypeptide backbone even before the first protein structure was determined. Sixty-five years later, powerful new methods breathe new life into this field. The highest three-state accuracy without relying on structure templates is now at 82-84%, a number unthinkable just a few years ago. These improvements came from increasingly larger databases of protein sequences and structures for training, the use of template secondary structure information and more powerful deep learning techniques. As we are approaching to the theoretical limit of three-state prediction (88-90%), alternative to secondary structure prediction (prediction of backbone torsion angles and Cα-atom-based angles and torsion angles) not only has more room for further improvement but also allows direct prediction of three-dimensional fragment structures with constantly improved accuracy. About 20% of all 40-residue fragments in a database of 1199 non-redundant proteins have <6 Å root-mean-squared distance from the native conformations by SPIDER2. More powerful deep learning methods with improved capability of capturing long-range interactions begin to emerge as the next generation of techniques for secondary structure prediction. The time has come to finish off the final stretch of the long march towards protein secondary structure prediction.

Mutations in OTOGL, encoding the inner ear protein otogelin-like, cause moderate sensorineural hearing loss

Hereditary hearing loss is characterized by a high degree of genetic heterogeneity. Here we present OTOGL mutations, a homozygous one base pair deletion (c.1430 delT) causing a frameshift (p.Val477Glufs(∗)25) in a large consanguineous family and two compound heterozygous mutations, c.547C>T (p.Arg183(∗)) and c.5238+5G>A, in a nonconsanguineous family with moderate nonsyndromic sensorineural hearing loss. OTOGL maps to the DFNB84 locus at 12q21.31 and encodes otogelin-like, which has structural similarities to the epithelial-secreted mucin protein family. We demonstrate that Otogl is expressed in the inner ear of vertebrates with a transcription level that is high in embryonic, lower in neonatal, and much lower in adult stages. Otogelin-like is localized to the acellular membranes of the cochlea and the vestibular system and to a variety of inner ear cells located underneath these membranes. Knocking down of otogl with morpholinos in zebrafish leads to sensorineural hearing loss and anatomical changes in the inner ear, supporting that otogelin-like is essential for normal inner ear function. We propose that OTOGL mutations affect the production and/or function of acellular structures of the inner ear, which ultimately leads to sensorineural hearing loss.

Induction of human immunodeficiency virus (HIV-1) envelope specific cell-mediated immunity by a non-homologous synthetic peptide

Background

Cell mediated immunity, including efficient CTL response, is required to prevent HIV-1 from cell-to-cell transmission. In previous investigations, we have shown that B1 peptide derived by Fourier transformation of HIV-1 primary structures and sharing no sequence homology with the parent proteins was able to generate antiserum which recognizes envelope and Tat proteins. Here we have investigated cellular immune response towards a novel non-homologous peptide, referred to as cA1 peptide.

Methodology/Principal Findings

The 20 amino acid sequence of cA1 peptide was predicted using the notion of peptide hydropathic properties; the peptide is encoded by the complementary anti-sense DNA strand to the sense strand of previously described non-homologous A1 peptide. In this report we demonstrate that the cA1 peptide can be a target for major histocompatibility complex (MHC) class I-restricted cytotoxic T lymphocytes in HIV-1-infected or envelope-immunized individuals. The cA1 peptide is recognized in association with different MHC class I allotypes and could prime in vitro CTLs, derived from gp160-immunized individuals capable to recognize virus variants.

Conclusions/Significance

For the first time a theoretically designed immunogen involved in broad-based cell-immune memory activation is described. Our findings may thus contribute to the advance in vaccine research by describing a novel strategy to develop a synthetic AIDS vaccine.

Application of expert networks for predicting proteins secondary structure

The present study utilizes expert neural networks for the prediction of proteins secondary structure. We use three independent networks, one for each structure (alpha, beta and coil) as the first-level processing unit; decision upon the chosen structure for each residue is carried out by a second-level, post-processing unit, which utilizes the Chou and Fasman frequency values Falpha and Fbeta in order to strengthen and/or deplete the probability of the specific structure under investigation. The highest prediction case was 76%. Our method requires primitive computational means and a relatively small training set, while still been comparable to previous work. It is not meant to be an alternative to the determination of secondary structure by means of free energy minimization, integration of dynamic equations of motion or crystallography, which are expensive, time-consuming and complicated, but to provide additional constrains, which might be considered and incorporated into larger computing setups in order to reduce the initial search space for the above methods.

ISIS: interaction sites identified from sequence

MOTIVATION

Large-scale experiments reveal pairs of interacting proteins but leave the residues involved in the interactions unknown. These interface residues are essential for understanding the mechanism of interaction and are often desired drug targets. Reliable identification of residues that reside in protein-protein interface typically requires analysis of protein structure. Therefore, for the vast majority of proteins, for which there is no high-resolution structure, there is no effective way of identifying interface residues.

RESULTS

Here we present a machine learning-based method that identifies interacting residues from sequence alone. Although the method is developed using transient protein-protein interfaces from complexes of experimentally known 3D structures, it never explicitly uses 3D information. Instead, we combine predicted structural features with evolutionary information. The strongest predictions of the method reached over 90% accuracy in a cross-validation experiment. Our results suggest that despite the significant diversity in the nature of protein-protein interactions, they all share common basic principles and that these principles are identifiable from sequence alone.

Hunting for "key residues" in the modeling of globular protein folding: an artificial neural network-based approach

An approach to modeling globular protein folding based on artificial neural networks (ANNs) is presented. This approach, that can be regarded as an inverse protein folding problem, investigates whether and when a protein fragment needs a specific residue in the center of its primary structure as a necessary condition to fold as observed. To perform this analysis, an ANN has been trained on a set of 55 proteins, searching for a relation between protein fragments modeled by 13alpha torsion angles and the residue corresponding to the central alpha torsion angle of the fragment. The results obtained show that only Asp, Gly, Pro, Ser and Val residues are often a necessary, even though not sufficient, condition to obtain a specific folded fragment structure, playing therefore, the role of "key residue" of this fragment.

PROFcon: novel prediction of long-range contacts

MOTIVATION

Despite the continuing advance in the experimental determination of protein structures, the gap between the number of known protein sequences and structures continues to increase. Prediction methods can bridge this sequence-structure gap only partially. Better predictions of non-local contacts between residues could improve comparative modeling, fold recognition and could assist in the experimental structure determination.

RESULTS

Here, we introduced PROFcon, a novel contact prediction method that combines information from alignments, from predictions of secondary structure and solvent accessibility, from the region between two residues and from the average properties of the entire protein. In contrast to some other methods, PROFcon predicted short and long proteins at similar levels of accuracy. As expected, PROFcon was clearly less accurate when tested on sparse evolutionary profiles, that is, on families with few homologs. Prediction accuracy was highest for proteins belonging to the SCOP alpha/beta class. PROFcon compared favorably with state-of-the-art prediction methods at the CASP6 meeting. While the performance may still be perceived as low, our method clearly pushed the mark higher. Furthermore, predictions are already accurate enough to seed predictions of global features of protein structure.

Applications of neural network prediction of conformational states for small peptides from spectra and of fold classes

Electronic structures of small peptides were calculated 'ab initio' with the help of Density Functional Theory (DFT) and molecular dynamics that rendered a set of conformational states of the peptides. For the structures of these states it was possible to derive atomic polar tensors that allowed us to construct vibrational spectra for each of the conformational states with low energy. From the spectra, neural networks could be trained to distinguish between the various states and thus be able to generate a larger set of relevant structures and their relation to secondary structures of the peptides. The calculations were done both with solvent atoms (up to ten water molecules) and without, and hence the neural networks could be used to monitor the influence of the solvent on hydrogen bond formation. The calculations at this stage only involved very short peptide fragments of a few alanine amino acids but already at this stage they could be compared with reasonable agreements to experiments. The neural networks are shown to be good in distinguishing the different conformers of the small alanine peptides, especially when in the gas phase. Also the task of predicting protein fold-classes, defined from line-geometry, seems promising.

Protein secondary structure: category assignment and predictability

In the last decade, the prediction of protein secondary structure has been optimized using essentially one and the same assignment scheme known as DSSP. We present here a different scheme, which is more predictable. This scheme predicts directly the hydrogen bonds, which stabilize the secondary structures. Single sequence prediction of the new three category assignment gives an overall prediction improvement of 3.1% and 5.1% compared to the DSSP assignment and schemes where the helix category consists of alpha-helix and 3(10)-helix, respectively. These results were achieved using a standard feed-forward neural network with one hidden layer on a data set identical to the one used in earlier work.

Neural-network analysis of the vibrational spectra of N-acetyl L-alanyl N'-methyl amide conformational states

Density-functional theory (DFT) calculations utilizing the Becke 3LYP hybrid functional have been carried out for N-acetyl L-alanine N'-methylamide and examined with respect to the effect of water on the structure, the vibrational frequencies, vibrational absorption (VA), vibrational circular dichroism (VCD), Raman spectra, and Raman optical activity (ROA) intensities. The large changes due to hydration in the structures, and the relative stability of the conformer, reflected in the VA, VCD, Raman spectra, and ROA spectra observed experimentally, are reproduced by the DFT calculations. A neural network has been constructed for reproducing the inverse scattering data (we infer the structural coordinates from spectroscopic data) that the DFT method could produce. The purpose of the network has also been to generate the large set of conformational states associated with each set of spectroscopic data for a given conformer of the molecule by interpolation. Finally the neural network performances are used to monitor a sensitivity analysis of the importance of secondary structures and the influence of the solvent. The neural network is shown to be good in distinguishing the different conformers of the small alanine peptide, especially in the gas phase.

Neural networks predict protein folding and structure: artificial intelligence faces biomolecular complexity

In the genomic era DNA sequencing is increasing our knowledge of the molecular structure of genetic codes from bacteria to man at a hyperbolic rate. Billions of nucleotides and millions of aminoacids are already filling the electronic files of the data bases presently available, which contain a tremendous amount of information on the most biologically relevant macromolecules, such as DNA, RNA and proteins. The most urgent problem originates from the need to single out the relevant information amidst a wealth of general features. Intelligent tools are therefore needed to optimise the search. Data mining for sequence analysis in biotechnology has been substantially aided by the development of new powerful methods borrowed from the machine learning approach. In this paper we discuss the application of artificial feedforward neural networks to deal with some fundamental problems tied with the folding process and the structure-function relationship in proteins.

Structure-based conformational preferences of amino acids

Proteins can be very tolerant to amino acid substitution, even within their core. Understanding the factors responsible for this behavior is of critical importance for protein engineering and design. Mutations in proteins have been quantified in terms of the changes in stability they induce. For example, guest residues in specific secondary structures have been used as probes of conformational preferences of amino acids, yielding propensity scales. Predicting these amino acid propensities would be a good test of any new potential energy functions used to mimic protein stability. We have recently developed a protein design procedure that optimizes whole sequences for a given target conformation based on the knowledge of the template backbone and on a semiempirical potential energy function. This energy function is purely physical, including steric interactions based on a Lennard-Jones potential, electrostatics based on a Coulomb potential, and hydrophobicity in the form of an environment free energy based on accessible surface area and interatomic contact areas. Sequences designed by this procedure for 10 different proteins were analyzed to extract conformational preferences for amino acids. The resulting structure-based propensity scales show significant agreements with experimental propensity scale values, both for alpha-helices and beta-sheets. These results indicate that amino acid conformational preferences are a natural consequence of the potential energy we use. This confirms the accuracy of our potential and indicates that such preferences should not be added as a design criterion.

Higher-order probabilistic perceptrons as Bayesian inference engines

An explicit structural connection is established between the Bayes optimal classifier operating on K binary input variables and a corresponding two-layer perceptron having normalized output activities and couplings from input to output units of all orders up to K. With suitable modification of connection weights and biases, such a higher-order probabilistic perceptron should in principle be able to learn the statistics of the classification problem and match the a posteriori probabilities given by Bayes optimal inference. Specific training algorithms are developed that allow this goal to be approximated in a controlled variational sense. An application to the task of discriminating between stable and unstable nuclides in nuclear physics yields network models with predictive performance comparable to the best that has been achieved with conventional multilayer perceptrons containing only pairwise connections.

Application of neural networks in structure-activity relationships

Methodology and application of artificial neural networks in structure-activity relationships are reviewed focusing on the most frequently used three-layer feedforward back-propagation procedure. Two applications of neural networks are presented and a comparison of the performance with those of CoMFA and a classical QSAR analysis is also discussed.

Investigations into the cross-reactivity of rabbit antibodies raised against nonhomologous pairs of synthetic peptides derived from HIV-1 gp120 proteins

The immunological cross-reactivity of several peptides with specific pattern-property characteristics related to the epitopes of human immunodeficiency virus type 1 (HIV-1) gp160/ 120 envelope proteins has been investigated. Proteins with similar primary structures can be expected to show functional or topographic similarities, such as specific epitopes which may cross-react with antibodies derived from the immunisation of animals with other members of the same protein family. These structure-function characteristics may be revealed as periodicities derived from presentations based on the discrete Fourier transformation of the distributions of various physico-chemical amino acid descriptors, constituting the polypeptide backbone and amino acid side-chains of the protein molecule. Such approaches, for example, have permitted prediction of periodicities corresponding to secondary structural motifs, including amphipathic alpha-helices and beta-sheets, within protein sequences, and have helped to clarify potential binding sites for ligands, substrates or cofactors with interacting macromolecules. Based on this approach, characteristic periodicities have been identified which represent common Fourier transform spectral properties of the envelope (ENV) gp160/120 glycoproteins from a range of HIV-1 isolates. In addition, similar periodicities have been detected as components of the discrete Fourier transform representation of the corresponding amino acid descriptors of the CD4 binding domain of gp120. Accordingly, we have synthesised several peptides having periodic characteristics in their discrete Fourier transform representations similar to these HIV-1 proteins. These nonhomologous synthetic peptides induced cross-reactive antibodies in New Zealand White rabbits. Polyclonal antibodies raised to one of these peptides reacted with HIV-1 ENV gp120-related proteins, as determined by enzyme-linked immunosorbent assay and Western blotting techniques. These findings provide further evidence for a role of immunological cross-reactivity and molecular biomimicry in the development of peptide-based vaccines directed against viral or bacterial pathogens.

NetOglyc: prediction of mucin type O-glycosylation sites based on sequence context and surface accessibility

The specificities of the UDP-GalNAc:polypeptide Nacetylgalactosaminyltransferases which link the carbohydrate GalNAc to the side-chain of certain serine and threonine residues in mucin type glycoproteins, are presently unknown. The specificity seems to be modulated by sequence context, secondary structure and surface accessibility. The sequence context of glycosylated threonines was found to differ from that of serine, and the sites were found to cluster. Non-clustered sites had a sequence context different from that of clustered sites. Charged residues were disfavoured at position -1 and +3. A jury of artificial neural networks was trained to recognize the sequence context and surface accessibility of 299 known and verified mucin type O-glycosylation sites extracted from O-GLYCBASE. The cross-validated NetOglyc network system correctly found 83% of the glycosylated and 90% of the non-glycosylated serine and threonine residues in independent test sets, thus proving more accurate than matrix statistics and vector projection methods. Predictions of O-glycosylation sites in the envelope glycoprotein gp120 from the primate lentiviruses HIV-1, HIV-2 and SIV are presented. The most conserved O-glycosylation signals in these evolutionary-related glycoproteins were found in their first hypervariable loop, V1. However, the strain variation for HIV-1 gp120 was significant. A computer server, available through WWW or E-mail, has been developed for prediction of mucin type O-glycosylation sites in proteins based on the amino acid sequence. The server addresses are http://www.cbs.dtu.dk/services/NetOGlyc/ and netOglyc@cbs.dtu.dk.

Artificial neural networks for molecular sequence analysis

Artificial neural networks provide a unique computing architecture whose potential has attracted interest from researchers across different disciplines. As a technique for computational analysis, neural network technology is very well suited for the analysis of molecular sequence data. It has been applied successfully to a variety of problems, ranging from gene identification, to protein structure prediction and sequence classification. This article provides an overview of major neural network paradigms, discusses design issues, and reviews current applications in DNA/RNA and protein sequence analysis.

Stabilization centers in proteins: identification, characterization and predictions

Methods are presented to locate residues, stabilization center elements, which are expected to stabilize protein structures by preventing their decay with their cooperative long range interactions. Artificial neural network-based algorithms were developed to predict these residues from the primary structure of single proteins and from the amino acid sequences of homologous proteins. The prediction accuracy using only single sequence information is 65%, but the incorporation of evolutionary information in the form of multiple alignments and conservation scores raises the efficiency by 3%. The composition, relative accessibility, number and type of interactions, conservation and the X-ray thermal factor of the identified stabilization center residues are different, not only from the whole data set but from the rest of the long range interacting residues as well. The most frequent stabilization center residues are usually found at buried positions and have a hydrophobic or aromatic side-chain, but some polar or charged residues also play an important role in the stabilization. The stabilization centers show significant difference in the composition and in the type of linked secondary structural elements compared with the rest of the residues. The performed structural and sequential conservation analysis showed the higher conservation of stabilization centers over protein families. The relation of the proposed stabilization centers to folding nuclei is also discussed.

Relationship between protein structure and geometrical constraints

We evaluate to what extent the structure of proteins can be deduced from incomplete knowledge of disulfide bridges, surface assignments, secondary structure assignments, and additional distance constraints. A cost function taking such constraints into account was used to obtain protein structures using a simple minimization algorithm. For small proteins, the approximate structure could be obtained using one additional distance constraint for each amino acid in the protein. We also studied the effect of using predicted secondary structure and surface assignments. The constraints used in this approach typically may be obtained from low-resolution experimental data. When using a cost function based on distances, half of the resulting structures will be mirrored, because the resulting structure and its mirror image will have the same cost. The secondary structure assignments were therefore divided into chirality constraints and distance constraints. Here we report that the problem of mirrored structures, in some cases, can be solved by using a chirality term in the cost function.

Biomolecular structure prediction at a low resolution using a neural network and the double-iterated Kalman filter technique

We report the application of an integrated computational approach for biomolecular structure determination at a low resolution. In particular, a neural network is trained to predict the spatial proximity of C-alpha atoms that are less than a given threshold apart, whereas a Kalman filter algorithm is employed to outline the biomolecular fold, with a constraints set that includes these pairwise atomic distances, and the distances and angles that define the structure as it is known from the protein's sequence. The results for Crambin demonstrate that this integrated approach is useful for molecular structure prediction at a low resolution and may also complement existing experimental distance data for a protein structure determination.

Protein structure and the sequential structure of mRNA: alpha-helix and beta-sheet signals at the nucleotide level

A direct comparison of experimentally determined protein structures and their corresponding protein coding mRNA sequences has been performed. We examine whether real world data support the hypothesis that clusters of rare codons correlate with the location of structural units in the resulting protein. The degeneracy of the genetic code allows for a biased selection of codons which may control the translational rate of the ribosome, and may thus in vivo have a catalyzing effect on the folding of the polypeptide chain. A complete search for GenBank nucleotide sequences coding for structural entries in the Brookhaven Protein Data Bank produced 719 protein chains with matching mRNA sequence, amino acid sequence, and secondary structure assignment. By neural network analysis, we found strong signals in mRNA sequence regions surrounding helices and sheets. These signals do not originate from the clustering of rare codons, but from the similarity of codons coding for very abundant amino acid residues at the N- and C-termini of helices and sheets. No correlation between the positioning of rare codons and the location of structural units was found. The mRNA signals were also compared with conserved nucleotide features of 16S-like ribosomal RNA sequences and related to mechanisms for maintaining the correct reading frame by the ribosome.

Self-organizing hierarchic networks for pattern recognition in protein sequence

We present a method based on hierarchical self-organizing maps (SOMs) for recognizing patterns in protein sequences. The method is fully automatic, does not require prealigned sequences, is insensitive to redundancy in the training set, and works surprisingly well even with small learning sets. Because it uses unsupervised neural networks, it is able to extract patterns that are not present in all of the unaligned sequences of the learning set. The identification of these patterns in sequence databases is sensitive and efficient. The procedure comprises three main training stages. In the first stage, one SOM is trained to extract common features from the set of unaligned learning sequences. A feature is a number of ungapped sequence segments (usually 4-16 residues long) that are similar to segments in most of the sequences of the learning set according to an initial similarity matrix. In the second training stage, the recognition of each individual feature is refined by selecting an optimal weighting matrix out of a variety of existing amino acid similarity matrices. In a third stage of the SOM procedure, the position of the features in the individual sequences is learned. This allows for variants with feature repeats and feature shuffling. The procedure has been successfully applied to a number of notoriously difficult cases with distinct recognition problems: helix-turn-helix motifs in DNA-binding proteins, the CUB domain of developmentally regulated proteins, and the superfamily of ribokinases. A comparison with the established database search procedure PROFILE (and with several others) led to the conclusion that the new automatic method performs satisfactorily.

Improving prediction of protein secondary structure using structured neural networks and multiple sequence alignments

The prediction of protein secondary structure by use of carefully structured neural networks and multiple sequence alignments has been investigated. Separate networks are used for predicting the three secondary structures alpha-helix, beta-strand, and coil. The networks are designed using a priori knowledge of amino acid properties with respect to the secondary structure and the characteristic periodicity in alpha-helices. Since these single-structure networks all have less than 600 adjustable weights, overfitting is avoided. To obtain a three-state prediction of alpha-helix, beta-strand, or coil, ensembles of single-structure networks are combined with another neural network. This method gives an overall prediction accuracy of 66.3% when using 7-fold cross-validation on a database of 126 nonhomologous globular proteins. Applying the method to multiple sequence alignments of homologous proteins increases the prediction accuracy significantly to 71.3% with corresponding Matthew's correlation coefficients C alpha = 0.59, C beta = 0.52, and Cc = 0.50. More than 72% of the residues in the database are predicted with an accuracy of 80%. It is shown that the network outputs can be interpreted as estimated probabilities of correct prediction, and, therefore, these numbers indicate which residues are predicted with high confidence.

Prediction of rib fracture injury outcome by an artificial neural network

Outcome-based therapy is becoming the standard for assessing patient care efficacy. This study examines the ability of an artificial neural network to predict rib fracture injury outcome based on 20 intake variables determined within 1 hour of admission. The data base contained 580 patient records with four outcome variables: Length of hospital stay (LOS), ICU days, Lived, and Died. A 522-patient training set and a 58-patient test set were randomly selected. Nine networks were set up in a feed-forward, back-propagating design with each trained under different initial conditions. These networks predicted the test set outcome variables with an accuracy as high as 98% at the 80% testing level. Internal weight matrix examination indicated that age, ventilatory support, and high trauma scores were strongly associated with both ICU days and mortality. Being female, injury severity, and injury type were associated with increased LOS. Smoking and rib fracture number were low-level predictors of the four outcome variables.

Progress of 1D protein structure prediction at last

Accuracy of predicting protein secondary structure and solvent accessibility from sequence information has been improved significantly by using information contained in multiple sequence alignments as input to a neural network system. For the Asilomar meeting, predictions for 13 proteins were generated automatically using the publicly available prediction method PHD. The results confirm the estimate of 72% three-state prediction accuracy. The fairly accurate predictions of secondary structure segments made the tool useful as a starting point for modeling of higher dimensional aspects of protein structure.

Prediction of O-glycosylation of mammalian proteins: specificity patterns of UDP-GalNAc:polypeptide N-acetylgalactosaminyltransferase

The specificity of the enzyme(s) catalysing the covalent link between the hydroxyl side chains of serine or threonine and the sugar moiety N-acetylgalactosamine (GalNAc) is unknown. Pattern recognition by artificial neural networks and weight matrix algorithms was performed to determine the exact position of in vivo O-linked GalNAc-glycosylated serine and threonine residues from the primary sequence exclusively. The acceptor sequence context for O-glycosylation of serine was found to differ from that of threonine and the two types were therefore treated separately. The context of the sites showed a high abundance of proline, serine and threonine extending far beyond the previously reported region covering positions -4 through +4 relative to the glycosylated residue. The O-glycosylation sites were found to cluster and to have a high abundance in the N-terminal part of the protein. The sites were also found to have an increased preference for three different classes of beta-turns. No simple consensus-like rule could be deduced for the complex glycosylation sequence acceptor patterns. The neural networks were trained on the hitherto largest data material consisting of 48 carefully examined mammalian glycoproteins comprising 264 O-glycosylation sites. For detection neural network algorithms were much more reliable than weight matrices. The networks correctly found 60-95% of the O-glycosylated serine/threonine residues and 88-97% of the non-glycosylated residues in two independent test sets of known glycoproteins. A computer server using E-mail for prediction of O-glycosylation sites has been implemented and made publicly available. The Internet address is NetOglyc@cbs.dtu.dk.

Feed-forward neural networks for secondary structure prediction

A feed-forward neural network has been employed for protein secondary structure prediction. Attempts were made to improve on previous prediction accuracies using a hierarchical mixture of experts (HME). In this method input data are clustered and used to train a series of different networks. Application of an HME to the prediction of protein secondary structure is shown to provide no advantages over a single network. We have also tried various new input representations, chosen to incorporate the effect of residues a long distance away in the one-dimensional amino acid chain. Prediction accuracy using these methods is comparable to that achieved by other neural networks.