Konstantinos P. Exarchos

Machine learning applications in cancer prognosis and prediction.

Cancer has been characterized as a heterogeneous disease consisting of many different subtypes. The early diagnosis and prognosis of a cancer type have become a necessity in cancer research, as it can facilitate the subsequent clinical management of patients. The importance of classifying cancer patients into high or low risk groups has led many research teams, from the biomedical and the bioinformatics field, to study the application of machine learning (ML) methods. Therefore, these techniques have been utilized as an aim to model the progression and treatment of cancerous conditions. In addition, the ability of ML tools to detect key features from complex datasets reveals their importance. A variety of these techniques, including Artificial Neural Networks (ANNs), Bayesian Networks (BNs), Support Vector Machines (SVMs) and Decision Trees (DTs) have been widely applied in cancer research for the development of predictive models, resulting in effective and accurate decision making. Even though it is evident that the use of ML methods can improve our understanding of cancer progression, an appropriate level of validation is needed in order for these methods to be considered in the everyday clinical practice. In this work, we present a review of recent ML approaches employed in the modeling of cancer progression. The predictive models discussed here are based on various supervised ML techniques as well as on different input features and data samples. Given the growing trend on the application of ML methods in cancer research, we present here the most recent publications that employ these techniques as an aim to model cancer risk or patient outcomes.

A multiscale and multiparametric approach for modeling the progression of oral cancer

BackgroundIn this work, we propose a multilevel and multiparametric approach in order to model the growth and progression of oral squamous cell carcinoma (OSCC) after remission. OSCC constitutes the major neoplasm of the head and neck region, exhibiting a quite aggressive nature, often leading to unfavorable prognosis.MethodsWe formulate a Decision Support System assembling a multitude of heterogeneous data sources (clinical, imaging tissue and blood genomic), aiming to capture all manifestations of the disease. Our primary aim is to identify the factors that dictate OSCC progression and subsequently predict potential relapses of the disease. The discrimination potential of each source of data is initially explored separately, and afterwards the individual predictions are combined to yield a consensus decision achieving complete discrimination between patients with and without a disease relapse. Moreover, we collect and analyze gene expression data from circulating blood cells throughout the follow-up period in consecutive time-slices, in order to model the temporal dimension of the disease. For this purpose a Dynamic Bayesian Network (DBN) is employed which is able to capture in a transparent manner the underlying mechanism dictating the disease evolvement, and employ it for monitoring the status and prognosis of the patients after remission.ResultsBy feeding as input to the DBN data from the baseline visit we achieve accuracy of 86%, which is further improved to complete discrimination when data from the first follow-up visit are also employed.ConclusionsKnowing in advance the progression of the disease, i.e. identifying groups of patients with higher/lower risk of reoccurrence, we are able to determine the subsequent treatment protocol in a more personalized manner.

Prediction of cis/trans isomerization using feature selection and support vector machines

In protein structures the peptide bond is found to be in trans conformation in the majority of the cases. Only a small fraction of peptide bonds in proteins is reported to be in cis conformation. Most of these instances (>90%) occur when the peptide bond is an imide (X-Pro) rather than an amide bond (X-nonPro). Due to the implication of cis/trans isomerization in many biologically significant processes, the accurate prediction of the peptide bond conformation is of high interest. In this study, we evaluate the effect of a wide range of features, towards the reliable prediction of both proline and non-proline cis/trans isomerization. We use evolutionary profiles, secondary structure information, real-valued solvent accessibility predictions for each amino acid and the physicochemical properties of the surrounding residues. We also explore the predictive impact of a modified feature vector, which consists of condensed position-specific scoring matrices (PSSMX), secondary structure and solvent accessibility. The best discriminating ability is achieved using the first feature vector combined with a wrapper feature selection algorithm and a support vector machine (SVM). The proposed method results in 70% accuracy, 75% sensitivity and 71% positive predictive value (PPV) in the prediction of the peptide bond conformation between any two amino acids. The output of the feature selection stage is investigated in order to identify discriminatory features as well as the contribution of each neighboring residue in the formation of the peptide bond, thus, advancing our knowledge towards cis/trans isomerization.

Multiparametric Decision Support System for the Prediction of Oral Cancer Reoccurrence

Oral squamous cell carcinoma (OSCC) constitutes the predominant neoplasm of the head and neck region, featuring particularly aggressive nature, associated with quite unfavorable prognosis. In this paper, we formulate a decision support system that integrates a multitude of heterogeneous data (clinical, imaging, and genomic), thus, framing all manifestations of the disease. Our primary aim is to identify the factors that dictate OSCC progression and subsequently predict potential relapses (local or metastatic) of the disease. The discrimination potential of each source of data is initially explored separately, and afterward the individual predictions are combined to yield a consensus decision achieving complete discrimination between patients with and without a disease relapse.

Prediction of coronary atherosclerosis progression using dynamic Bayesian networks

In this paper we propose a methodology for predicting the progression of atherosclerosis in coronary arteries using dynamic Bayesian networks. The methodology takes into account patient data collected at the baseline study and the same data collected in the follow-up study. Our aim is to analyze all the different sources of information (Demographic, Clinical, Biochemical profile, Inflammatory markers, Treatment characteristics) in order to predict possible manifestations of the disease; subsequently, our purpose is twofold: i) to identify the key factors that dictate the progression of atherosclerosis and ii) based on these factors to build a model which is able to predict the progression of atherosclerosis for a specific patient, providing at the same time information about the underlying mechanism of the disease.

Towards building a Dynamic Bayesian Network for monitoring oral cancer progression using time-course gene expression data

In this work we present a methodology for modeling and monitoring the evolvement of oral cancer in remittent patients during the post-treatment follow-up period. Our primary aim is to calculate the probability that a patient will develop a relapse but also to identify the approximate time-frame that this relapse is prone to appear. To this end, we start off by analyzing a broad set of time-course gene expression data in order to identify a set of genes that are mostly differentially expressed between patients with and without relapse and are therefore discriminatory and indicative of a disease reoccurrence evolvement. Next, we employ the maintained genes coupled with a patient-specific risk indicator in order to build upon them a Dynamic Bayesian Network (DBN) able to stratify patients based on their probability for a disease reoccurrence, but also pinpoint an approximate time-frame that the relapse might appear.

PBOND: web server for the prediction of proline and non-proline cis/trans isomerization.

PBOND is a web server that predicts the conformation of the peptide bond between any two amino acids. PBOND classifies the peptide bonds into one out of four classes, namely cis imide (cis-Pro), cis amide (cis-nonPro), trans imide (trans-Pro) and trans amide (trans-nonPro). Moreover, for every prediction a reliability index is computed. The underlying structure of the server consists of three stages: (1) feature extraction, (2) feature selection and (3) peptide bond classification. PBOND can handle both single sequences as well as multiple sequences for batch processing. The predictions can either be directly downloaded from the web site or returned via e-mail. The PBOND web server is freely available at http://195.251.198.21/pbond.html.

Extraction of consensus protein patterns in regions containing non-proline cis peptide bonds and their functional assessment

BackgroundIn peptides and proteins, only a small percentile of peptide bonds adopts the cis configuration. Especially in the case of amide peptide bonds, the amount of cis conformations is quite limited thus hampering systematic studies, until recently. However, lately the emerging population of databases with more 3D structures of proteins has produced a considerable number of sequences containing non-proline cis formations (cis-nonPro).ResultsIn our work, we extract regular expression-type patterns that are descriptive of regions surrounding the cis-nonPro formations. For this purpose, three types of pattern discovery are performed: i) exact pattern discovery, ii) pattern discovery using a chemical equivalency set, and iii) pattern discovery using a structural equivalency set. Afterwards, using each pattern as predicate, we search the Eukaryotic Linear Motif (ELM) resource to identify potential functional implications of regions with cis-nonPro peptide bonds. The patterns extracted from each type of pattern discovery are further employed, in order to formulate a pattern-based classifier, which is used to discriminate between cis-nonPro and trans-nonPro formations.ConclusionsIn terms of functional implications, we observe a significant association of cis-nonPro peptide bonds towards ligand/binding functionalities. As for the pattern-based classification scheme, the highest results were obtained using the structural equivalency set, which yielded 70% accuracy, 77% sensitivity and 63% specificity.

Modelling of Oral Cancer Progression Using Dynamic Bayesian Networks

This work addresses the problem of finding the mortality distribution for lung cancer in Mexican districts, through clustering patterns discovery. A data mining system was developed which consists of a pattern generator and a visualization subsystem. Such an approach may contribute to biomarker discovery by means of identifying risk regions for a given cancer type and further reduce the cost and time spend in conducting cancer studies. The k-means algorithm was used for the generation of patterns, which permits expressing patterns as groups of districts with affinity in their location and mortality rate attributes. The source data were obtained from Mexican official institutions. As a result, a set of grouping patterns reflecting the mortality distribution of lung cancer in Mexico was generated. Two interesting patterns in northeastern and northwestern Mexico with high mortality rate were detected. We consider that patterns generated by the data mining system, can be useful for identifying high risk cancer areas and biomarkers discovery.

Prediction of time dependent survival in HF patients after VAD implantation using pre- and post-operative data

Heart failure is one of the most common diseases worldwide. In recent years, Ventricular Assist Devices (VADs) have become a valuable option for patients with advanced HF. Although it has been shown that VADs improve patient survival rates, several complications persist during left VAD (LVAD) support. The stratification scores currently employed are mostly generic, i.e. not specifically built for LVAD patients, and are based on pre-implantation patient data. In this work we apply data mining approaches for the prediction of time dependent survival in patients after LVAD implantation. Moreover, the predictions acquired with the use of pre-implantation data are enriched by employing post-implantation data, i.e. follow-up data. Different clinical scenarios have been depicted and the subsequent conditions are tested in order to identify the optimal set of pre- and post-implant features, as well as the most suitable algorithms for feature selection and prediction. The proposed approach is applied to a real dataset of 71 patients, reporting an accuracy of 84.5%, sensitivity of 87% and specificity of 82%. Based on the reported results, expert cardio-surgeons can be supported in planning the treatment of VAD patients.