On training targets for deep learning approaches to clean speech magnitude spectrum estimation.

Estimation of the clean speech short-time magnitude spectrum (MS) is key for speech enhancement and separation. Moreover, an automatic speech recognition (ASR) system that employs a front-end relies on clean speech MS estimation to remain robust. Training targets for deep learning approaches to clean speech MS estimation fall into three categories: computational auditory scene analysis (CASA), MS, and minimum mean square error (MMSE) estimator training targets. The choice of the training target can have a significant impact on speech enhancement/separation and robust ASR performance. Motivated by this, the training target that produces enhanced/separated speech at the highest quality and intelligibility and that which is best for an ASR front-end is found. Three different deep neural network (DNN) types and two datasets, which include real-world nonstationary and coloured noise sources at multiple signal-to-noise ratio (SNR) levels, were used for evaluation. Ten objective measures were employed, including the word error rate of the Deep Speech ASR system. It is found that training targets that estimate the a priori SNR for MMSE estimators produce the highest objective quality scores. Moreover, it is established that the gain of MMSE estimators and the ideal amplitude mask produce the highest objective intelligibility scores and are most suitable for an ASR front-end.

An objective measure of quality for time-scale modification of audio.

Objective evaluation of audio processed with time-scale modification (TSM) remains an open problem. Recently, a dataset of time-scaled audio with subjective quality labels was published and used to create an initial objective measure of quality (OMOQ). In this paper, an improved OMOQ for time-scaled audio is proposed. The measure uses handcrafted features and a fully connected network to predict subjective mean opinion scores (SMOS). Basic and advanced perceptual evaluation of audio quality features are used in addition to nine features specific to TSM artefacts. Six methods of alignment are explored with interpolation of the reference magnitude spectrum to the length of the test magnitude spectrum giving the best performance. The proposed measure achieves a mean root mean square error of 0.490 and a mean Pearson correlation of 0.864 to SMOS, equivalent to the 97th and 82nd percentiles of the subjective sessions, respectively. The proposed measure is used to evaluate TSM algorithms, finding that Elastique gives the highest objective quality for solo instrument and voice signals, whereas the identity phase-locking phase vocoder gives the highest objective quality for music signals and the best overall quality. The objective measure is available online at https://www.github.com/zygurt/TSM.

Spectral distortion level resulting in a just-noticeable difference between an a priori signal-to-noise ratio estimate and its instantaneous case.

Minimum mean-square error (MMSE) approaches to speech enhancement are widely used in the literature. The quality of enhanced speech produced by an MMSE approach is directly impacted by the accuracy of the employed a priori signal-to-noise ratio (SNR) estimator. In this paper, the a priori SNR estimate spectral distortion (SD) level that results in a just-noticeable difference (JND) in the perceived quality of MMSE approach enhanced speech is found. The JND SD level is indicative of the accuracy that an a priori SNR estimator must exceed to have no impact on the perceived quality of MMSE approach enhanced speech. To measure the JND SD level, listening tests are conducted across five SNR levels, five noise sources, and two MMSE approaches [the MMSE short-time spectral amplitude (MMSE-STSA) estimator and the Wiener filter]. A statistical analysis of the results indicates that the JND SD level increases with the SNR level, is higher for the MMSE-STSA estimator, and is not impacted by the type of background noise. Following the literature, a significant improvement in a priori SNR estimation accuracy is required to reach the JND SD level.

A time-scale modification dataset with subjective quality labels.

Time Scale Modification (TSM) is a well-researched field; however, no effective objective measure of quality exists. This paper details the creation, subjective evaluation, and analysis of a dataset for use in the development of an objective measure of quality for TSM. Comprised of two parts, the training component contains 88 source files processed using six TSM methods at 10 time scales, while the testing component contains 20 source files processed using three additional methods at four time scales. The source material contains speech, solo harmonic and percussive instruments, sound effects, and a range of music genres. Ratings (42 529) were collected from 633 sessions using laboratory and remote collection methods. Analysis of results shows no correlation between age and quality of rating; expert and non-expert listeners to be equivalent; minor differences between participants with and without hearing issues; and minimal differences between testing modalities. A comparison of published objective measures and subjective scores shows the objective measures to be poor indicators of subjective quality. Initial results for a retrained objective measure of quality are presented with results approaching average root mean squared error loss and Pearson correlation values of subjective sessions. The labeled dataset is available at http://ieee-dataport.org/1987.

Getting to Know Your Neighbor: Protein Structure Prediction Comes of Age with Contextual Machine Learning.

The folding of a protein structure is a process governed by both local and nonlocal interactions. While incorporating local dependencies into a machine learning algorithm for protein structure prediction is simple and has been exploited for some time, the modeling of long-range dependences which result from structurally-neighboring residues has only recently begun to be addressed. Structural properties designed to localize the prediction space from direct tertiary structure prediction, such as secondary structure, contact maps, and intrinsic disorder, among others, have begun to greatly benefit from machine learning models capable of modeling a widened, potentially global protein context. This has led to a direct enhancement of the quality of predicted tertiary structures through both the optimization of structural constraints and improved reliability of alignments to structural templates. These improvements have stemmed from the application of recurrent and convolutional neural network architectures effective not only at innate sequential context propagation but also deep feature extraction due to novel skip connections and normalization techniques allowing for greatly enhanced error back-propagation. The recent results from independent blind testing in Critical Assessment of protein Structure Prediction 13 have signaled the beginning of a new generation of protein structure prediction through the utilization of these contextual techniques. The ripples from advancements in the determination of one-dimensional and two-dimensional structural properties have us moving ever closer to the solution of the protein structure prediction problem.

SPOT-Disorder2: Improved Protein Intrinsic Disorder Prediction by Ensembled Deep Learning.

Intrinsically disordered or unstructured proteins (or regions in proteins) have been found to be important in a wide range of biological functions and implicated in many diseases. Due to the high cost and low efficiency of experimental determination of intrinsic disorder and the exponential increase of unannotated protein sequences, developing complementary computational prediction methods has been an active area of research for several decades. Here, we employed an ensemble of deep Squeeze-and-Excitation residual inception and long short-term memory (LSTM) networks for predicting protein intrinsic disorder with input from evolutionary information and predicted one-dimensional structural properties. The method, called SPOT-Disorder2, offers substantial and consistent improvement not only over our previous technique based on LSTM networks alone, but also over other state-of-the-art techniques in three independent tests with different ratios of disordered to ordered amino acid residues, and for sequences with either rich or limited evolutionary information. More importantly, semi-disordered regions predicted in SPOT-Disorder2 are more accurate in identifying molecular recognition features (MoRFs) than methods directly designed for MoRFs prediction. SPOT-Disorder2 is available as a web server and as a standalone program at https://sparks-lab.org/server/spot-disorder2/.

Protein fold recognition using HMM-HMM alignment and dynamic programming.

Detecting three dimensional structures of protein sequences is a challenging task in biological sciences. For this purpose, protein fold recognition has been utilized as an intermediate step which helps in classifying a novel protein sequence into one of its folds. The process of protein fold recognition encompasses feature extraction of protein sequences and feature identification through suitable classifiers. Several feature extractors are developed to retrieve useful information from protein sequences. These features are generally extracted by constituting protein's sequential, physicochemical and evolutionary properties. The performance in terms of recognition accuracy has also been gradually improved over the last decade. However, it is yet to reach a well reasonable and accepted level. In this work, we first applied HMM-HMM alignment of protein sequence from HHblits to extract profile HMM (PHMM) matrix. Then we computed the distance between respective PHMM matrices using kernalized dynamic programming. We have recorded significant improvement in fold recognition over the state-of-the-art feature extractors. The improvement of recognition accuracy is in the range of 2.7-11.6% when experimented on three benchmark datasets from Structural Classification of Proteins.

A mixture of physicochemical and evolutionary-based feature extraction approaches for protein fold recognition.

Recent advancement in the pattern recognition field stimulates enormous interest in Protein Fold Recognition (PFR). PFR is considered as a crucial step towards protein structure prediction and drug design. Despite all the recent achievements, the PFR still remains as an unsolved issue in biological science and its prediction accuracy still remains unsatisfactory. Furthermore, the impact of using a wide range of physicochemical-based attributes on the PFR has not been adequately explored. In this study, we propose a novel mixture of physicochemical and evolutionary-based feature extraction methods based on the concepts of segmented distribution and density. We also explore the impact of 55 different physicochemical-based attributes on the PFR. Our results show that by providing more local discriminatory information as well as obtaining benefit from both physicochemical and evolutionary-based features simultaneously, we can enhance the protein fold prediction accuracy up to 5% better than previously reported results found in the literature.

Probabilistic expression of spatially varied amino acid dimers into general form of Chou׳s pseudo amino acid composition for protein fold recognition.

BACKGROUND: Identification of the tertiary structure (3D structure) of a protein is a fundamental problem in biology which helps in identifying its functions. Predicting a protein׳s fold is considered to be an intermediate step for identifying the tertiary structure of a protein. Computational methods have been applied to determine a protein׳s fold by assembling information from its structural, physicochemical and/or evolutionary properties. METHODS: In this study, we propose a scheme in which a feature extraction technique that extracts probabilistic expressions of amino acid dimers, which have varying degree of spatial separation in the primary sequences of proteins, from the Position Specific Scoring Matrix (PSSM). SVM classifier is used to create a model from extracted features for fold recognition. RESULTS: The performance of the proposed scheme is evaluated against three benchmarked datasets, namely the Ding and Dubchak, Extended Ding and Dubchak, and Taguchi and Gromiha datasets. CONCLUSIONS: The proposed scheme performed well in the experiments conducted, providing improvements over previously published results in literature.

Improving protein fold recognition using the amalgamation of evolutionary-based and structural based information.

Deciphering three dimensional structure of a protein sequence is a challenging task in biological science. Protein fold recognition and protein secondary structure prediction are transitional steps in identifying the three dimensional structure of a protein. For protein fold recognition, evolutionary-based information of amino acid sequences from the position specific scoring matrix (PSSM) has been recently applied with improved results. On the other hand, the SPINE-X predictor has been developed and applied for protein secondary structure prediction. Several reported methods for protein fold recognition have only limited accuracy. In this paper, we have developed a strategy of combining evolutionary-based information (from PSSM) and predicted secondary structure using SPINE-X to improve protein fold recognition. The strategy is based on finding the probabilities of amino acid pairs (AAP). The proposed method has been tested on several protein benchmark datasets and an improvement of 8.9% recognition accuracy has been achieved. We have achieved, for the first time over 90% and 75% prediction accuracies for sequence similarity values below 40% and 25%, respectively. We also obtain 90.6% and 77.0% prediction accuracies, respectively, for the Extended Ding and Dubchak and Taguchi and Gromiha benchmark protein fold recognition datasets widely used for in the literature.

Protein fold recognition by alignment of amino acid residues using kernelized dynamic time warping.

In protein fold recognition, a protein is classified into one of its folds. The recognition of a protein fold can be done by employing feature extraction methods to extract relevant information from protein sequences and then by using a classifier to accurately recognize novel protein sequences. In the past, several feature extraction methods have been developed but with limited recognition accuracy only. Protein sequences of varying lengths share the same fold and therefore they are very similar (in a fold) if aligned properly. To this, we develop an amino acid alignment method to extract important features from protein sequences by computing dissimilarity distances between proteins. This is done by measuring distance between two respective position specific scoring matrices of protein sequences which is used in a support vector machine framework. We demonstrated the effectiveness of the proposed method on several benchmark datasets. The method shows significant improvement in the fold recognition performance which is in the range of 4.3-7.6% compared to several other existing feature extraction methods.

A tri-gram based feature extraction technique using linear probabilities of position specific scoring matrix for protein fold recognition.

In biological sciences, the deciphering of a three dimensional structure of a protein sequence is considered to be an important and challenging task. The identification of protein folds from primary protein sequences is an intermediate step in discovering the three dimensional structure of a protein. This can be done by utilizing feature extraction technique to accurately extract all the relevant information followed by employing a suitable classifier to label an unknown protein. In the past, several feature extraction techniques have been developed but with limited recognition accuracy only. In this study, we have developed a feature extraction technique based on tri-grams computed directly from Position Specific Scoring Matrices. The effectiveness of the feature extraction technique has been shown on two benchmark datasets. The proposed technique exhibits up to 4.4% improvement in protein fold recognition accuracy compared to the state-of-the-art feature extraction techniques.

A strategy to select suitable physicochemical attributes of amino acids for protein fold recognition.

BACKGROUND: Assigning a protein into one of its folds is a transitional step for discovering three dimensional protein structure, which is a challenging task in bimolecular (biological) science. The present research focuses on: 1) the development of classifiers, and 2) the development of feature extraction techniques based on syntactic and/or physicochemical properties. RESULTS: Apart from the above two main categories of research, we have shown that the selection of physicochemical attributes of the amino acids is an important step in protein fold recognition and has not been explored adequately. We have presented a multi-dimensional successive feature selection (MD-SFS) approach to systematically select attributes. The proposed method is applied on protein sequence data and an improvement of around 24% in fold recognition has been noted when selecting attributes appropriately. CONCLUSION: The MD-SFS has been applied successfully in selecting physicochemical attributes of the amino acids. The selected attributes show improved protein fold recognition performance.

A feature extraction technique using bi-gram probabilities of position specific scoring matrix for protein fold recognition.

Discovering a three dimensional structure of a protein is a challenging task in biological science. Classifying a protein into one of its folds is an intermediate step for deciphering the three dimensional protein structure. The protein fold recognition can be done by developing feature extraction techniques to accurately extract all the relevant information from a protein sequence and then by employing a suitable classifier to label an unknown protein. Several feature extraction techniques have been developed in the past but with limited recognition accuracy only. In this work, we have developed a feature extraction technique which is based on bi-grams computed directly from Position Specific Scoring Matrices and demonstrated its effectiveness on a benchmark dataset. The proposed technique exhibits an absolute improvement of around 10% compared with existing feature extraction techniques.