Quasi-supervised Strategies for Compound-protein Interaction Prediction.

In-silico compound-protein interaction prediction addresses prioritization of drug candidates for experimental biochemical validation because the wet-lab experiments are time-consuming, laborious and costly. Most machine learning methods proposed to that end approach this problem with supervised learning strategies in which known interactions are labeled as positive and the rest are labeled as negative. However, treating all unknown interactions as negative instances may lead to inaccuracies in real practice since some of the unknown interactions are bound to be positive interactions waiting to be identified as such. In this study, we propose to address this problem using the Quasi-Supervised Learning (QSL) algorithm. In this framework, potential interactions are predicted by estimating the overlap between a true positive dataset of compound-protein pairs with known interactions and an unknown dataset of all the remaining compound-protein pairs. The potential interactions are then identified as those in the unknown dataset that overlap with the interacting pairs in the true positive dataset in terms of the associated similarity structure. We also address the class-imbalance problem by modifying the conventional cost function of the QSL algorithm. Experimental results on GPCR and Nuclear Receptor datasets show that the proposed method can identify actual interactions from all possible combinations.

On the characterization of cognitive tasks using activity-specific short-lived synchronization between electroencephalography channels.

Accurate characterization of brain activity during a cognitive task is challenging due to the dynamically changing and the complex nature of the brain. The majority of the proposed approaches assume stationarity in brain activity and disregard the systematic timing organization among brain regions during cognitive tasks. In this study, we propose a novel cognitive activity recognition method that captures the activity-specific timing parameters from training data that elicits maximal average short-lived pairwise synchronization between electroencephalography signals. We evaluated the characterization power of the activity-specific timing parameter triplets in a motor imagery activity recognition framework. The activity-specific timing parameter triplets consist of latency of the maximally synchronized signal segments from activity onset Δt, the time lag between maximally synchronized signal segments τ, and the duration of the maximally synchronized signal segments w. We used cosine-based similarity, wavelet bi-coherence, phase-locking value, phase coherence value, linearized mutual information, and cross-correntropy to calculate the channel synchronizations at the specific timing parameters. Recognition performances as well as statistical analyses on both BCI Competition-III dataset IVa and PhysioNet Motor Movement/Imagery dataset, indicate that the inter-channel short-lived synchronization calculated using activity-specific timing parameter triplets elicit significantly distinct synchronization profiles for different motor imagery tasks and can thus reliably be used for cognitive task recognition purposes.

Evaluation of synchronization measures for capturing the lagged synchronization between EEG channels: A cognitive task recognition approach.

During cognitive, perceptual and sensory tasks, connectivity profile changes across different regions of the brain. Variations of such connectivity patterns between different cognitive tasks can be evaluated using pairwise synchronization measures applied to electrophysiological signals, such as electroencephalography (EEG). However, connectivity-based task recognition approaches achieving viable recognition performance have been lacking from the literature. By using several synchronization measures, we identify time lags between channel pairs during different cognitive tasks. We employed mutual information, cross correntropy, cross correlation, phase locking value, cosine similarity and nonlinear interdependence measures. In the training phase, for each type of cognitive task, we identify the time lags that maximize the average synchronization between channel pairs. These lags are used to calculate pairwise synchronization values with which we construct the train and test feature vectors for recognition of the cognitive task carried out using Fisher's linear discriminant (FLD) analysis. We tested our framework in a motor imagery activity recognition scenario on PhysioNet Motor Movement/Imagery and BCI Competition-III Ⅳa datasets. For PhysioNet dataset, average performance results ranging between % 51 and % 61 across 20 subjects. For BCI Competition-Ⅲ dataset, we achieve an average recognition performance of % 76 which is above the minimum reliable communication rate (% 70). We achieved an average accuracy over the minimum reliable communication rate on the BCI Competition-Ⅲ dataset. Performance levels were lower on the PhysioNet dataset. These results indicate that a viable task recognition system is achievable using pairwise synchronization measures evaluated at the proper task specific lags.

Taste dysfunction in multiple sclerosis.

Empirical studies of taste function in multiple sclerosis (MS) are rare. Moreover, a detailed assessment of whether quantitative measures of taste function correlate with the punctate and patchy myelin-related lesions found throughout the CNS of MS patients has not been made. We administered a 96-trial test of sweet (sucrose), sour (citric acid), bitter (caffeine) and salty (NaCl) taste perception to the left and right anterior (CN VII) and posterior (CN IX) tongue regions of 73 MS patients and 73 matched controls. The number and volume of lesions were assessed using quantitative MRI in 52 brain regions of 63 of the MS patients. Taste identification scores were significantly lower in the MS patients for sucrose (p = 0.0002), citric acid (p = 0.0001), caffeine (p = 0.0372) and NaCl (p = 0.0004) and were present in both anterior and posterior tongue regions. The percent of MS patients with identification scores falling below the 5th percentile of controls was 15.07 % for caffeine, 21.9 % for citric acid, 24.66 % for sucrose, and 31.50 % for NaCl. Such scores were inversely correlated with lesion volumes in the temporal, medial frontal, and superior frontal lobes, and with the number of lesions in the left and right superior frontal lobes, right anterior cingulate gyrus, and left parietal operculum. Regardless of the subject group, women outperformed men on the taste measures. These findings indicate that a sizable number of MS patients exhibit taste deficits that are associated with MS-related lesions throughout the brain.

Automated labeling of cancer textures in larynx histopathology slides using quasi-supervised learning.

OBJECTIVE: To evaluate the performance of a quasi-supervised statistical learning algorithm, operating on datasets having normal and neoplastic tissues, to identify larynx squamous cell carcinomas. Furthermore, cancer texture separability measures against normal tissues are to be developed and compared either for colorectal or larynx tissues. STUDY DESIGN: Light microscopic digital images from histopathological sections were obtained from laryngectomy materials including squamous cell carcinoma and nonneoplastic regions. The texture features were calculated by using co-occurrence matrices and local histograms. The texture features were input to the quasi-supervised learning algorithm. RESULTS: Larynx regions containing squamous cell carcinomas were accurately identified, having false and true positive rates up to 21% and 87%, respectively. CONCLUSION: Larynx squamous cell carcinoma versus normal tissue texture separability measures were higher than colorectal adenocarcinoma versus normal textures for the colorectal database. Furthermore, the resultant labeling performances for all larynx datasets are higher than or equal to that of colorectal datasets. The results in larynx datasets, in comparison with the former colorectal study, suggested that quasi-supervised texture classification is to be a helpful method in histopathological image classification and analysis.

Automatic identification of highly conserved family regions and relationships in genome wide datasets including remote protein sequences.

Identifying shared sequence segments along amino acid sequences generally requires a collection of closely related proteins, most often curated manually from the sequence datasets to suit the purpose at hand. Currently developed statistical methods are strained, however, when the collection contains remote sequences with poor alignment to the rest, or sequences containing multiple domains. In this paper, we propose a completely unsupervised and automated method to identify the shared sequence segments observed in a diverse collection of protein sequences including those present in a smaller fraction of the sequences in the collection, using a combination of sequence alignment, residue conservation scoring and graph-theoretical approaches. Since shared sequence fragments often imply conserved functional or structural attributes, the method produces a table of associations between the sequences and the identified conserved regions that can reveal previously unknown protein families as well as new members to existing ones. We evaluated the biological relevance of the method by clustering the proteins in gold standard datasets and assessing the clustering performance in comparison with previous methods from the literature. We have then applied the proposed method to a genome wide dataset of 17793 human proteins and generated a global association map to each of the 4753 identified conserved regions. Investigations on the major conserved regions revealed that they corresponded strongly to annotated structural domains. This suggests that the method can be useful in predicting novel domains on protein sequences.

Automated labelling of cancer textures in colorectal histopathology slides using quasi-supervised learning.

Quasi-supervised learning is a statistical learning algorithm that contrasts two datasets by computing estimate for the posterior probability of each sample in either dataset. This method has not been applied to histopathological images before. The purpose of this study is to evaluate the performance of the method to identify colorectal tissues with or without adenocarcinoma. Light microscopic digital images from histopathological sections were obtained from 30 colorectal radical surgery materials including adenocarcinoma and non-neoplastic regions. The texture features were extracted by using local histograms and co-occurrence matrices. The quasi-supervised learning algorithm operates on two datasets, one containing samples of normal tissues labelled only indirectly, and the other containing an unlabeled collection of samples of both normal and cancer tissues. As such, the algorithm eliminates the need for manually labelled samples of normal and cancer tissues for conventional supervised learning and significantly reduces the expert intervention. Several texture feature vector datasets corresponding to different extraction parameters were tested within the proposed framework. The Independent Component Analysis dimensionality reduction approach was also identified as the one improving the labelling performance evaluated in this series. In this series, the proposed method was applied to the dataset of 22,080 vectors with reduced dimensionality 119 from 132. Regions containing cancer tissue could be identified accurately having false and true positive rates up to 19% and 88% respectively without using manually labelled ground-truth datasets in a quasi-supervised strategy. The resulting labelling performances were compared to that of a conventional powerful supervised classifier using manually labelled ground-truth data. The supervised classifier results were calculated as 3.5% and 95% for the same case. The results in this series in comparison with the benchmark classifier, suggest that quasi-supervised image texture labelling may be a useful method in the analysis and classification of pathological slides but further study is required to improve the results.

The prognostic value of tumor-stroma proportion in laryngeal squamous cell carcinoma.

OBJECTIVE: Tumor-stroma proportion of tumor has been presented as a prognostic factor in some types of adenocarcinomas, but there is no information about squamous cell carcinomas and laryngeal carcinomas. MATERIAL AND METHOD: Five digital images of the tumor sections were obtained from 85 laryngeal carcinomas. Proportion of epithelial tumor component and stroma were measured by a software tool, allowing the pathologists to mark 205.6 µm2 blocks on areas as carcinomatous/stromal, by clicking at the image. Totally, 3.451 mm2 tumor areas have been marked to 16.785 small square blocks for each case. RESULTS: Median follow up was 48 months (range 3-194). The mean tumor-stroma proportion was 48.63+18.18. There was no difference for tumor-stroma proportion when tumor location, grade, stage and perinodal invasion were considered. Although the following results were statistically insignificant, the mean tumor-stroma proportion was the lowest (37.46±12.49) for subglottic carcinomas, and it was 52.41±37.47, 50.86+19.84 and 44.56±16.91 for supraglottic, transglottic and glottic cases. The tumor-stroma proportion was lowest in cases with perinodal invasion and the highest in cases without lymph node metastasis (44.72±20.23, 47.77±17.37, 50.05±17.34). Tumor-stroma proportion was higher in the basaloid subtype compared with the classical squamous cell carcinoma (53.76±14.70 and 48.63±18.38 respectively). The overall and disease-free survival analysis did not reveal significance for tumor-stroma proportion (p=0.08, p=0.38). Only pathological stage was an independent factor for overall survival (p=0.008). CONCLUSION: This is the first series investigating tumor-stroma proportion as a prognostic marker in laryngeal carcinomas proposing a new method, but the findings do not support tumor-stroma proportion as a prognostic marker.

Hierarchical motif vectors for prediction of functional sites in amino acid sequences using quasi-supervised learning.

We propose hierarchical motif vectors to represent local amino acid sequence configurations for predicting the functional attributes of amino acid sites on a global scale in a quasi-supervised learning framework. The motif vectors are constructed via wavelet decomposition on the variations of physico-chemical amino acid properties along the sequences. We then formulate a prediction scheme for the functional attributes of amino acid sites in terms of the respective motif vectors using the quasi-supervised learning algorithm that carries out predictions for all sites in consideration using only the experimentally verified sites. We have carried out comparative performance evaluation of the proposed method on the prediction of N-glycosylation of 55,184 sites possessing the consensus N-glycosylation sequon identified over 15,104 human proteins, out of which only 1,939 were experimentally verified N-glycosylation sites. In the experiments, the proposed method achieved better predictive performance than the alternative strategies from the literature. In addition, the predicted N-glycosylation sites showed good agreement with existing potential annotations, while the novel predictions belonged to proteins known to be modified by glycosylation.

Prediction potential of candidate biomarker sets identified and validated on gene expression data from multiple datasets.

BACKGROUND: Independently derived expression profiles of the same biological condition often have few genes in common. In this study, we created populations of expression profiles from publicly available microarray datasets of cancer (breast, lymphoma and renal) samples linked to clinical information with an iterative machine learning algorithm. ROC curves were used to assess the prediction error of each profile for classification. We compared the prediction error of profiles correlated with molecular phenotype against profiles correlated with relapse-free status. Prediction error of profiles identified with supervised univariate feature selection algorithms were compared to profiles selected randomly from a) all genes on the microarray platform and b) a list of known disease-related genes (a priori selection). We also determined the relevance of expression profiles on test arrays from independent datasets, measured on either the same or different microarray platforms. RESULTS: Highly discriminative expression profiles were produced on both simulated gene expression data and expression data from breast cancer and lymphoma datasets on the basis of ER and BCL-6 expression, respectively. Use of relapse-free status to identify profiles for prognosis prediction resulted in poorly discriminative decision rules. Supervised feature selection resulted in more accurate classifications than random or a priori selection, however, the difference in prediction error decreased as the number of features increased. These results held when decision rules were applied across-datasets to samples profiled on the same microarray platform. CONCLUSION: Our results show that many gene sets predict molecular phenotypes accurately. Given this, expression profiles identified using different training datasets should be expected to show little agreement. In addition, we demonstrate the difficulty in predicting relapse directly from microarray data using supervised machine learning approaches. These findings are relevant to the use of molecular profiling for the identification of candidate biomarker panels.

Automated recognition of cell phenotypes in histology images based on membrane- and nuclei-targeting biomarkers.

BACKGROUND: Three-dimensional in vitro culture of cancer cells are used to predict the effects of prospective anti-cancer drugs in vivo. In this study, we present an automated image analysis protocol for detailed morphological protein marker profiling of tumoroid cross section images. METHODS: Histologic cross sections of breast tumoroids developed in co-culture suspensions of breast cancer cell lines, stained for E-cadherin and progesterone receptor, were digitized and pixels in these images were classified into five categories using k-means clustering. Automated segmentation was used to identify image regions composed of cells expressing a given biomarker. Synthesized images were created to check the accuracy of the image processing system. RESULTS: Accuracy of automated segmentation was over 95% in identifying regions of interest in synthesized images. Image analysis of adjacent histology slides stained, respectively, for Ecad and PR, accurately predicted regions of different cell phenotypes. Image analysis of tumoroid cross sections from different tumoroids obtained under the same co-culture conditions indicated the variation of cellular composition from one tumoroid to another. Variations in the compositions of cross sections obtained from the same tumoroid were established by parallel analysis of Ecad and PR-stained cross section images. CONCLUSION: Proposed image analysis methods offer standardized high throughput profiling of molecular anatomy of tumoroids based on both membrane and nuclei markers that is suitable to rapid large scale investigations of anti-cancer compounds for drug development.

Automated detection of regions of interest for tissue microarray experiments: an image texture analysis.

BACKGROUND: Recent research with tissue microarrays led to a rapid progress toward quantifying the expressions of large sets of biomarkers in normal and diseased tissue. However, standard procedures for sampling tissue for molecular profiling have not yet been established. METHODS: This study presents a high throughput analysis of texture heterogeneity on breast tissue images for the purpose of identifying regions of interest in the tissue for molecular profiling via tissue microarray technology. Image texture of breast histology slides was described in terms of three parameters: the percentage of area occupied in an image block by chromatin (B), percentage occupied by stroma-like regions (P), and a statistical heterogeneity index H commonly used in image analysis. Texture parameters were defined and computed for each of the thousands of image blocks in our dataset using both the gray scale and color segmentation. The image blocks were then classified into three categories using the texture feature parameters in a novel statistical learning algorithm. These categories are as follows: image blocks specific to normal breast tissue, blocks specific to cancerous tissue, and those image blocks that are non-specific to normal and disease states. RESULTS: Gray scale and color segmentation techniques led to identification of same regions in histology slides as cancer-specific. Moreover the image blocks identified as cancer-specific belonged to those cell crowded regions in whole section image slides that were marked by two pathologists as regions of interest for further histological studies. CONCLUSION: These results indicate the high efficiency of our automated method for identifying pathologic regions of interest on histology slides. Automation of critical region identification will help minimize the inter-rater variability among different raters (pathologists) as hundreds of tumors that are used to develop an array have typically been evaluated (graded) by different pathologists. The region of interest information gathered from the whole section images will guide the excision of tissue for constructing tissue microarrays and for high throughput profiling of global gene expression.

Simulating deformations of MR brain images for validation of atlas-based segmentation and registration algorithms.

Simulated deformations and images can act as the gold standard for evaluating various template-based image segmentation and registration algorithms. Traditional deformable simulation methods, such as the use of analytic deformation fields or the displacement of landmarks followed by some form of interpolation, are often unable to construct rich (complex) and/or realistic deformations of anatomical organs. This paper presents new methods aiming to automatically simulate realistic inter- and intra-individual deformations. The paper first describes a statistical approach to capturing inter-individual variability of high-deformation fields from a number of examples (training samples). In this approach, Wavelet-Packet Transform (WPT) of the training deformations and their Jacobians, in conjunction with a Markov random field (MRF) spatial regularization, are used to capture both coarse and fine characteristics of the training deformations in a statistical fashion. Simulated deformations can then be constructed by randomly sampling the resultant statistical distribution in an unconstrained or a landmark-constrained fashion. The paper also describes a model for generating tissue atrophy or growth in order to simulate intra-individual brain deformations. Several sets of simulated deformation fields and respective images are generated, which can be used in the future for systematic and extensive validation studies of automated atlas-based segmentation and deformable registration methods. The code and simulated data are available through our Web site.

Simulation of tissue atrophy using a topology preserving transformation model.

We propose a method to simulate atrophy and other similar volumetric change effects on medical images. Given a desired level of atrophy, we find a dense warping deformation that produces the corresponding levels of volumetric loss on the labeled tissue using an energy minimization strategy. Simulated results on a real brain image indicate that the method generates realistic images of tissue loss. The method does not make assumptions regarding the mechanics of tissue deformation, and provides a framework where a pre-specified pattern of atrophy can readily be simulated. Furthermore, it provides exact correspondences between images prior and posterior to the atrophy that can be used to evaluate provisional image registration and atrophy quantification algorithms.

Statistical representation and simulation of high-dimensional deformations: application to synthesizing brain deformations.

This paper proposes an approach to effectively representing the statistics of high-dimensional deformations, when relatively few training samples are available, and conventional methods, like PCA, fail due to insufficient training. Based on previous work on scale-space decomposition of deformation fields, herein we represent the space of "valid deformations" as the intersection of three subspaces: one that satisfies constraints on deformations themselves, one that satisfies constraints on Jacobian determinants of deformations, and one that represents smooth deformations via a Markov Random Field (MRF). The first two are extensions of PCA-based statistical shape models. They are based on a wavelet packet basis decomposition that allows for more accurate estimation of the covariance structure of deformation or Jacobian fields, and they are used jointly due to their complementary strengths and limitations. The third is a nested MRF regularization aiming at eliminating potential discontinuities introduced by assumptions in the statistical models. A randomly sampled deformation field is projected onto the space of valid deformations via iterative projections on each of these subspaces until convergence, i.e. all three constraints are met. A deformation field simulator uses this process to generate random samples of deformation fields that are not only realistic but also representative of the full range of anatomical variability. These simulated deformations can be used for validation of deformable registration methods. Other potential uses of this approach include representation of shape priors in statistical shape models as well as various estimation and hypothesis testing paradigms in the general fields of computational anatomy and pattern recognition.

Estimating topology preserving and smooth displacement fields.

We propose a method for enforcing topology preservation and smoothness onto a given displacement field. We first analyze the conditions for topology preservation on two- and three-dimensional displacement fields over a discrete rectangular grid. We then pose the problem of finding the closest topology preserving displacement field in terms of its complete set of gradients, which we later solve using a cyclic projections framework. Adaptive smoothing of a displacement field is then formulated as an extension of topology preservation, via constraints imposed on the Jacobian of the displacement field. The simulation results indicate that this technique is a fast and reliable method to estimate a topology preserving displacement field from a noisy observation that does not necessarily preserve topology. They also show that the proposed smoothing method can render morphometric analysis methods that are based on displacement field of shape transformations more robust to noise without removing important morphologic characteristics.

Morphological classification of brains via high-dimensional shape transformations and machine learning methods.

A high-dimensional shape transformation posed in a mass-preserving framework is used as a morphological signature of a brain image. Population differences with complex spatial patterns are then determined by applying a nonlinear support vector machine (SVM) pattern classification method to the morphological signatures. Significant reduction of the dimensionality of the morphological signatures is achieved via wavelet decomposition and feature reduction methods. Applying the method to MR images with simulated atrophy shows that the method can correctly detect subtle and spatially complex atrophy, even when the simulated atrophy represents only a 5% variation from the original image. Applying this method to actual MR images shows that brains can be correctly determined to be male or female with a successful classification rate of 97%, using the leave-one-out method. This proposed method also shows a high classification rate for old adults' age classification, even under difficult test scenarios. The main characteristic of the proposed methodology is that, by applying multivariate pattern classification methods, it can detect subtle and spatially complex patterns of morphological group differences which are often not detectable by voxel-based morphometric methods, because these methods analyze morphological measurements voxel-by-voxel and do not consider the entirety of the data simultaneously.

Topology preservation and regularity in estimated deformation fields.

A general formalism to impose topology preserving regularity on a given irregular deformation field is presented. The topology preservation conditions are derived with regard to the discrete approximations to the deformation field Jacobian in a two-dimensional image registration problem. The problem of enforcing topology preservation onto a given deformation field is formulated in terms of the deformation gradients, and solved using a cyclic projections approach. The generalization of the developed algorithm leads to a deformation field regularity control by limiting the per voxel volumetric change within a prescribed interval. Extension of the topology preservation conditions onto a three-dimensional registration problem is also presented, together with a comparative analysis of the proposed algorithm with respect to a Gaussian regularizer that enforces the same topology preservation conditions.