Stacked ensemble learning based on deep convolutional neural networks for pediatric pneumonia diagnosis using chest X-ray images.

Pneumonia is an acute respiratory infection caused by bacteria, viruses, or fungi and has become very common in children ranging from 1 to 5 years of age. Common symptoms of pneumonia include difficulty breathing due to inflamed or pus and fluid-filled alveoli. The United Nations Children's Fund reports nearly 800,000 deaths in children due to pneumonia. Delayed diagnosis and overpriced tests are the prime reason for the high mortality rate, especially in underdeveloped countries. A time and cost-efficient diagnosis tool: Chest X-rays, was thus accepted as the standard diagnostic test for pediatric pneumonia. However, the lower radiation levels for diagnosis in children make the task much more onerous and time-consuming. The mentioned challenges initiate the need for a computer-aided detection model that is instantaneous and accurate. Our work proposes a stacked ensemble learning of deep learning-based features for pediatric pneumonia classification. The extracted features from the global average pooling layer of the fine-tuned Xception model pretrained on ImageNet weights are sent to the Kernel Principal Component Analysis for dimensionality reduction. The dimensionally reduced features are further trained and validated on the stacking classifier. The stacking classifier consists of two stages; the first stage uses the Random-Forest classifier, K-Nearest Neighbors, Logistic Regression, XGB classifier, Support Vector Classifier (SVC), Nu-SVC, and MLP classifier. The second stage operates on Logistic Regression using the first stage predictions for the final classification with Stratified K-fold cross-validation to prevent overfitting. The model was tested on the publicly available pediatric pneumonia dataset, achieving an accuracy of 98.3%, precision of 99.29%, recall of 98.36%, F1-score of 98.83%, and an AUC score of 98.24%. The performance shows its reliability for real-time deployment in assisting radiologists and physicians.

Pediatric pneumonia diagnosis using stacked ensemble learning on multi-model deep CNN architectures.

Pediatric pneumonia has drawn immense awareness due to the high mortality rates over recent years. The acute respiratory infection caused by bacteria, viruses, or fungi infects the lung region and hinders oxygen transport, making breathing difficult due to inflamed or pus and fluid-filled alveoli. Being non-invasive and painless, chest X-rays are the most common modality for pediatric pneumonia diagnosis. However, the low radiation levels for diagnosis in children make accurate detection challenging. This challenge initiates the need for an unerring computer-aided diagnosis model. Our work proposes Contrast Limited Adaptive Histogram Equalization for image enhancement and a stacking classifier based on the fusion of deep learning-based features for pediatric pneumonia diagnosis. The extracted features from the global average pooling layers of the fine-tuned MobileNet, DenseNet121, DenseNet169, and DenseNet201 are concatenated for the final classification using a stacked ensemble classifier. The stacking classifier uses Support Vector Classifier, Nu-SVC, Logistic Regression, K-Nearest Neighbor, Random Forest Classifier, Gaussian Naïve Bayes, AdaBoost classifier, Bagging Classifier, and Extra-trees Classifier for the first stage, and Nu-SVC as the meta-classifier. The stacking classifier validated using Stratified K-Fold cross-validation achieves an accuracy of 98.62%, precision of 98.99%, recall of 99.53%, F1 score of 99.26%, and an AUC score of 93.17% on the publicly available pediatric pneumonia dataset. We expect this model to greatly help the real-time diagnosis of pediatric pneumonia.

CNN based classification of motor imaginary using variational mode decomposed EEG-spectrum image.

A novel approach of preprocessing EEG signals by generating spectrum image for effective Convolutional Neural Network (CNN) based classification for Motor Imaginary (MI) recognition is proposed. The approach involves extracting the Variational Mode Decomposition (VMD) modes of EEG signals, from which the Short Time Fourier Transform (STFT) of all the modes are arranged to form EEG spectrum images. The EEG spectrum images generated are provided as input image to CNN. The two generic CNN architectures for MI classification (EEGNet and DeepConvNet) and the architectures for pattern recognition (AlexNet and LeNet) are used in this study. Among the four architectures, EEGNet provides average accuracies of 91.37%, 94.41%, 85.67% and 90.21% for the four datasets used to validate the proposed approach. Consistently better results in comparison with results in recent literature demonstrate that the EEG spectrum image generation using VMD-STFT is a promising method for the time frequency analysis of EEG signals.

A hybrid method for fundamental heart sound segmentation using group-sparsity denoising and variational mode decomposition.

Segmentation of fundamental heart sounds-S1 and S2 is important for automated monitoring of cardiac activity including diagnosis of the heart diseases. This pa-per proposes a novel hybrid method for S1 and S2 heart sound segmentation using group sparsity denoising and variation mode decomposition (VMD) technique. In the proposed method, the measured phonocardiogram (PCG) signals are denoised using group sparsity algorithm by exploiting the group sparse (GS) property of PCG signals. The denoised GS-PCG signals are then decomposed into subsequent modes with specific spectral characteristics using VMD algorithm. The appropriate mode for further processing is selected based on mode central frequencies and mode energy. It is then followed by the extraction of Hilbert envelope (HEnv) and a thresholding on the selected mode to segment S1 and S2 heart sounds. The performance advantage of the proposed method is verified using PCG signals from benchmark databases namely eGeneralMedical, Littmann, Washington, and Michigan. The proposed hybrid algorithm has achieved a sensitivity of 100%, positive predictivity of 98%, accuracy of 98% and detection error rate of 1.5%. The promising results obtained suggest that proposed approach can be considered for automated heart sound segmentation.

Determining the three-dimensional fold of a protein from approximate constraints: a simulation study.

We propose a new approach for calculating the three-dimensional (3D) structure of a protein from distance and dihedral angle constraints derived from experimental data. We suggest that such constraints can be obtained from experiments such as tritium planigraphy, chemical or enzymatic cleavage of the polypeptide chain, paramagnetic perturbation of nuclear magnetic resonance (NMR) spectra, measurement of hydrogen-exchange rates, mutational studies, mass spectrometry, and electron paramagnetic resonance. These can be supplemented with constraints from theoretical prediction of secondary structures and of buried/exposed residues. We report here distance geometry calculations to generate the structures of a test protein Staphylococcal nuclease (STN), and the HIV-1 rev protein (REV) of unknown structure. From the available 3D atomic coordinates of STN, we set up simulated data sets consisting of varying number and quality of constraints, and used our group's Self Correcting Distance Geometry (SECODG) program DIAMOD to generate structures. We could generate the correct tertiary fold from qualitative (approximate) as well as precise distance constraints. The root mean square deviations of backbone atoms from the native structure were in the range of 2.0 A to 8.3 A, depending on the number of constraints used. We could also generate the correct fold starting from a subset of atoms that are on the surface and those that are buried. When we used data sets containing a small fraction of incorrect distance constraints, the SECODG technique was able to detect and correct them. In the case of REV, we used a combination of constraints obtained from mutagenic data and structure predictions. DIAMOD generated helix-loop-helix models, which, after four self-correcting cycles, populated one family exclusively. The features of the energy-minimized model are consistent with the available data on REV-RNA interaction. Our method could thus be an attractive alternative for calculating protein 3D structures, especially in cases where the traditional methods of X-ray crystallography and multidimensional NMR spectroscopy have been unsuccessful.

Homology modeling and characterization of IgE binding epitopes of mountain cedar allergen Jun a 3.

The Jun a 3 protein from mountain cedar (Juniperus ashei) pollen, a member of group 5 of the family of plant pathogenesis-related proteins (PR-proteins), reacts with serum IgE from patients with cedar hypersensitivity. We used the crystal structures of two other proteins of this group, thaumatin and an antifungal protein from tobacco, both approximately 50% identical in sequence to Jun a 3, as templates to build homology models for the allergen. The in-house programs EXDIS and FANTOM were used to extract distance and dihedral angle constraints from the Protein Data Bank files and determine energy-minimized structures. The mean backbone deviations for the energy-refined model structures from either of the templates is <1 A, their conformational energies are low, and their stereochemical properties (determined with PROCHECK) are acceptable. The circular dichroism spectrum of Jun a 3 is consistent with the postulated beta-sheet core. Tryptic fragments of Jun a 3 that reacted with IgE from allergic patients all mapped to one helical/loop surface of the models. The Jun a 3 models have features common to aerosol allergens from completely different protein families, suggesting that tertiary structural elements may mediate the triggering of an allergic response.

Mutational analysis of apolipoprotein B mRNA editing enzyme (APOBEC1). structure-function relationships of RNA editing and dimerization.

APOBEC1 is the catalytic subunit of an enzyme complex that mediates apolipoprotein (apo) B mRNA editing. It dimerizes in vitro and requires complementation factor(s) for its editing activity. We have performed a systematic analysis of the structure-functional relationship of APOBEC1 by targeted mutagenesis of various sequence motifs within the protein. Using in vitro RNA editing assay, we found that basic amino acid clusters at the amino-terminal region R15R16R17 and R33K34, are essential for apoB mRNA editing. Mutation of R15R16R17 to K15K16K17 and mutation of R33K34 simultaneously to A33A34 almost completely abolished in vitro editing activity. The carboxy-terminal region of APOBEC1 contains a leucine-rich motif. Deletion analysis of this region indicates that residues 181 to 210 are important for in vitro apoB mRNA editing. Single amino acid substitutions demonstrate that L182, I185, and L189 are important residues required for normal editing function. Furthermore, the double mutant P190A/P191A also lost >90% of editing activity which suggests that a beta turn in this region of the molecule may be essential for proper functioning of APOBEC1. It was suggested that dimerization of APOBEC1 creates an active structure for deamination of apoB mRNA. When we examined the dimerization potential of truncated APOBEC1s using both amino and carboxy termini deletion mutants, we found that amino-terminal deletions up to residue A117 did not impair dimerization activity whereas carboxy-terminal deletions showed diminished dimerization. The systematic and extensive mutagenesis experiments in this study provide information on the role of various sequence motifs identified in APOBEC1 in enzyme catalysis and dimerization.

Conformationally variable Rab protein surface regions mapped by limited proteolysis and homology modelling.

Tryptic proteolysis of the small GTPases Rab4 and Rab5 is a multi-step, nucleotide-dependent process. Using N-terminal peptide sequencing, matrix-assisted laser desorption ionization-time-of-flight MS and molecular modelling, we identified the three initial sites of proteolysis in Rab5 as Arg-4, Arg-81 and Arg-197. Arg-4 and Arg-81 lie within regions previously implicated in Rab5 endocytic function, and Arg-197 lies in a region involved in membrane targeting. Topologically, Arg-81 lies within the conformationally variable Switch II region shown to be important for protein-protein interactions of other GTPases. Homology modelling studies on Rab5 indicate that the Arg-81 side chain is buried in the Rab5 GTP conformation, but is solvent-accessible in the GDP conformation, explaining the dependence of proteolysis on nucleotides. Peptide mapping of Rab4 was performed to take advantage of additional scissile bonds within Switch II to determine more precisely the limits of the nucleotide-dependent protease-accessible region. The Rab4 cleavage sites corresponded to Arg-81 and Pro-87 of Rab5, and taken together with the finding that Rab5 was not cleaved at Arg-91 this analysis defines an eight-residue surface-exposed conformationally variable region lying in the centre of Switch II. A sequence comparison of Rab proteins shows these eight residues to have a loosely conserved motif that we term Switch II(v) for its relative variability. C-terminal to Switch II(v) is a highly conserved Rab-specific YYRGA motif that we term Switch II(c) for its constant sequence. N-terminal to Switch II(v) is a sequence-invariant G-domain involved in nucleotide binding and hydrolysis. We propose that the Rab Switch II(v) region imparts specificity to nucleotide-dependent protein-protein interactions.

Template-based docking of a prolactin receptor proline-rich motif octapeptide to FKBP12: implications for cytokine receptor signaling.

A conserved proline-rich motif (PRM) in the cytoplasmic domain of cytokine receptors has been suggested to be a signaling switch regulated by the action of the FK506 binding protein (FKBP) family of peptidylprolyl isomerases (O'Neal KD, Yu-Lee LY, Shearer WT, 1995, Ann NY Acad Sci 766:282-284). We have docked the prolactin receptor PRM (Ile1-Phe2-Pro3-Pro4-Val5-Pro6-Gly7-Pro8) to the ligand binding site of FKBP12. The procedure involved conformational search restricted by NMR restraints (O'Neal KD et al., 1996, Biochem J 315:833-844), energy minimization of the octapeptide conformers so obtained, template-based docking of a selected conformer to FKBP12, and energy refinement of the resulting complex. The template used was the crystal structure of a cyclic FK506-peptide hybrid bound to FKBP12. Val5-Pro6 of the PRM was taken to be the biologically relevant Xaa-Pro bond. The docked conformer is stabilized by two intramolecular hydrogen bonds, N7H7-->O4 and N2H2-->O8, and two intermolecular ones, Ile56; N-H-->O = C:Pro6 and Tyr82:O-H-->O = C:Gly7. This conformer features a Type I beta-turn and has extensive hydrophobic contacts with the FKBP12 binding surface. The observed interactions support the hypothesis that FKBP12 catalyzes cis-trans isomerization in the PRM when it is part of the longer cytoplasmic domain of a cytokine receptor, and suggest a significant role for the PRM in signal transduction.

Structure-activity relationship of quaternary ammonium ions at the external tetraethylammonium binding site of cloned potassium channels.

Changes in the chemical structure of the tetraethylammonium (TEA) ion reduce binding affinity at the external TEA receptor of outwardly rectifying potassium channels. To study the mechanism of selective binding, we applied a variety of hydrophilic quaternary ammonium (QA) ions to the noninactivating mutant of Shaker B T449Y, to Kv3.1, and to Kv3.1 mutants, expressed in Xenopus oocytes. In outside-out patches, QA ions in which ethyl groups of TEA were replaced by methyl groups had a lower affinity than TEA, whereas changes in binding affinity were minor when propyl groups were substituted for ethyl groups. All channels tested showed this pattern. Changes in free energy of binding correlated well with changes in the computed free energy of hydration of the TEA derivatives that we used. The affinity for TEA derivatives was reduced in Kv3.1Y407T, which is in support of the hypothesis that cation pi-electron interaction is involved. Binding affinities of QA ions were higher in Kv3.1 Y407F than in the wild-type, suggesting that the hydroxyl groups of the tyrosines reduce QA binding. The rank order of potency of the QA ions toward the different channels studied was the same. These results indicate that external QA ions bind strongly to hydrophobic pi-electron-rich functions. The selectivity, however, is determined by the physical properties of the QA ion.

The selectivity of different external binding sites for quaternary ammonium ions in cloned potassium channels.

Tetraethylammonium (TEA) is thought to be the most effective quaternary ammonium (QA) ion blocker at the external site of K+ channels, and small changes to the TEA ion reduce its potency. To examine the properties of the external QA receptor, we applied a variety of QA ions to excised patches from human embryonic kidney cells or Xenopus oocytes transfected with the delayed rectifying K+ channels Kv 2.1 and Kv 3.1. In outside-out patches of Kv 3.1, the relative potencies were TEA > tetrapropylammonium (TPA) > tetrabutylammonium (TBA). In contrast to Kv 3.1, the relative potencies in Kv 2.1 were TBA > TEA > TPA. In Kv 3.1 and Kv 2.1, external tetrapentylammonium (TPeA) blocked K+ currents in a fast, reversible and, in contrast to TEA, time-dependent manner. The external binding of TPeA appeared to be voltage independent, unlike the effects of TPeA applied to inside-out patches. External n-alkyl-triethylammonium compounds (C8, C10 chain length) had a lower affinity than TEA in Kv 3.1, but a higher affinity than TEA in Kv 2.1. In Kv 3.1, the decrease in QA affinity was large when one or two methyl groups were substituted for ethyl groups in TEA, but minor when propyl groups replaced ethyl groups. Changes in the free energy of binding could be correlated to changes in the free energy of hydration of TEA derivatives calculated by continuum methodology. These results reveal a substantial hydrophobic component of external QA ion binding to Kv 2.1, and to a lesser degree to Kv 3.1, in addition to the generally accepted electrostatic interactions. The chain length of hydrophobic TEA derivatives affects the affinity for the hydrophobic binding site, whereas the hydropathy of QA ions determines the electrostatic interaction energy.

Secondary structure prediction of the H5 pore of potassium channels.

The 'H5' segment located between the putative fifth and sixth transmembrane helices is the most highly conserved region in voltage-gated potassium channels and it is believed to constitute a major part of the ion conduction path (pore). Here we present a two-step procedure, comprising secondary structure prediction and hydrophobic moment profiling, to predict the structure of this important region. Combined results from the application of the procedure to the H5 region of four voltage-gated and five other K+ channel sequences lead to the prediction of a beta-strand-turn-beta-strand structure for H5. The reasons for the application of these soluble protein methods to parts of membrane proteins are: (i) that pore-lining residues are accessible to water and (ii) that a large enough database of high-resolution membrane protein structures does not yet exist. The results are compared with experimental results, in particular spectroscopic studies of two peptides based on the H5 sequence of SHAKER potassium channel. The procedure developed here may be applicable to water-accessible regions of other membrane proteins.

Molecular dynamics analysis of a ribonuclease C-peptide analogue.

We have carried out a nanosecond molecular dynamics simulation of an analogue of the ribonuclease C-peptide in water. The overall conformation has an extended region for the first three amino acids connected to an alpha-helix for residues 4-13, and this basic structure is preserved throughout the simulation, with helical hydrogen bonds present 87% of the time, on average. The final helical hydrogen bond is spontaneously broken and re-formed several times, providing a detailed picture of such winding/unwinding events. The simulation was used to estimate the effects of internal motion on proton nuclear Overhauser effect spectroscopy (NOESY) intensities for several classes of important cross peaks. Within the helical regions, the effects of internal motion vary only a little from one residue to another for backbone-backbone cross peaks, and the relevant correlation functions reach plateau values within about 50 ps. The spectral simulations show, however, that it may be difficult to establish a close quantitative connection between NOESY cross-peak volumes and measures of helical content.

Conformational studies on beta-bend containing a cis peptide unit.

Conformational studies have been carried out on the X-cis-Pro tripeptide system (a system of three linked peptide units, in the trans-cis-trans configuration) using energy minimization techniques. For X, residues Gly, L-Ala, D-Ala and L-Pro have been used. The energy minima have been classified into different groups based upon the conformational similarity. There are 15, 20, 18 and 6 minima that are possible for the four cases respectively and these fall into 11 different groups. A study of these minima shows that, (i) some minima contain hydrogen bonds--either 4-->1 or 1-->2 type, (ii) the low energy minima qualify themselves as bend conformations, (iii) cis' and trans' conformations are possible for the prolyl residue as also the C gamma-endo and C gamma-exo puckerings, and (iv) for Pro-cis-Pro, cis' at the first prolyl residue is ruled out, due to the high energy. The available crystal structure data on proteins and peptides, containing cis-Pro segment have been examined with a view to find the minima that occur in solid state. The data from protein show that they fall under two groups. The conformation at X in X-cis-Pro is near extended when it is a non-glycyl residue. In both peptides and proteins there exists a preference for trans' conformation at prolyl residue over cis' when X is a non-glycyl residue. The minima obtained can be useful in modelling studies.

Unfolding of an alpha-helix in water.

We describe a 1 ns molecular dynamics simulation of an 18-residue peptide (corresponding to a portion of the H helix of myoglobin) in water. The initial helical conformation progressively frays to a more disordered structure, with the loss of internal secondary structure generally proceeding from the C-terminus toward the N-terminus. Although a variety of mechanisms are involved in the breaking of helical hydrogen bonds, the formation of transient turn structures, with i----i + 3 hydrogen bonds, and bifurcated hydrogen-bond structures intermediate between alpha and turn or 3(10) structures is a common motif. In some cases a single water molecule is inserted into an internal hydrogen bond, but it is also common to have several water molecules involved in transient intermediates. Overall, the results provide new information about the detailed mechanisms by which helices are made and broken in aqueous solution.

Electrical potentials in trypsin isozymes.

Cow and rat trypsin differ in net charge by 12.5 units yet have the same enzymatic mechanism. The role of electrical potentials in the catalytic mechanism of these trypsin isozymes is investigated by using the finite difference Poisson-Boltzmann method. The calculations reveal that the active sites are effectively shielded from surface charge, thus making it possible for the two enzymes to have essentially identical potentials in their catalytically important regions. The potentials in both active sites are dominated by local interactions arising both from partial charges and from the negative charge on Asp-102. The latter is found to stabilize the transition state by about 4 kcal/mol, a value that is consistent with the extent of reduced catalytic activity in the variant Asn-102 trypsin, in which the negative charge is absent. The calculations predict that Asp-102 is ionized and that His-57 is neutral in the resting state of the enzyme. In contrast to their negligible effect on catalytic activity, the cumulative effect of surface charges is found to raise the pK of the N-terminal alpha-amino group of Ile-16 in the rat enzyme by about 1.5 units relative to that of cow trypsin. This charged amino acid forms an ionic bond with Asp-194, which stabilizes the active conformation of the enzyme. An increase in pK of Ile-16 thus provides a possible explanation for the retention of activity of rat trypsin at high pH. The results of this study could not have been obtained from an electrostatic model based on Coulombic potentials.

Three-dimensional solution structure of a single zinc finger DNA-binding domain.

The three-dimensional solution structure of a zinc finger nucleic acid binding motif has been determined by nuclear magnetic resonance (NMR) spectroscopy. Spectra of a synthetic peptide corresponding to a single zinc finger from the Xenopus protein Xfin yielded distance and dihedral angle constraints that were used to generate structures from distance geometry and restrained molecular dynamics calculations. The zinc finger is an independently folded domain with a compact globular structure in which the zinc atom is bound by two cysteine and two histidine ligands. The polypeptide backbone fold consists of a well-defined helix, starting as alpha and ending as 3(10) helix, packed against two beta strands that are arranged in a hairpin structure. A high density of basic and polar amino acid side chains on the exposed face of the helix are probably involved in DNA binding.

A novel supersecondary structure in globular proteins comprising the collagen-like helix and beta-turn.

A structure consisting of the polyproline-II or collagen-like helix immediately succeeded by a beta-turn is seen in several synthetic peptides and has been suggested to be the conformational requirement for proline hydroxylation in nascent procollagen. Using a simple algorithm for detecting secondary structures, we have analysed crystal structure data on 40 globular proteins and have found eight examples of the collagen-helix + beta-turn supersecondary structure in 15 proteins that contain the collagen-like helical segments.