Machine learning approaches in bioinformatics and genomic analysis

Machine Learning Method in Bioinformatics

Bioinformatics involves an integrated approach involving the use of information technology, computer science to biology and medicine as professional and knowledge fields. It encompasses the knowledge associated with information systems, artificial intelligence, databases, and algorithms, soft computing, software engineering, image processing, modeling and simulation, data mining, signal processing, computation theory and information, system an d control theory, discrete mathematics, statistics and circuit theory. On the other hand, machine learning entails a sub-division of artificial intelligence and operates with technical skills to permit computers to adapt to certain responses and initiate actions (Zhang et al., 2009). Machine learning entails a range technical knowledge that looks at the scientific application of search engines, natural language processing, bioinformatics, medical diagnosis and cheminformatics, analysis of the stock market, game playing and computer vision.

The development of machine learning has been as a matter of necessity given the fact that, current knowledge that needs some levels of sophistication and technological advancement has been on the consistent growth. These are included in the revolution in the genomic field that entails amino acid sequencing and nucleotide sequences. To accomplish the possibility of storing these essential informati0on, machine learning has been imperative has led t-o the building of several sophisticated interfaces that researchers can manipulate to establish access to the available databases. In general, it is evident that the abilities of computers in learning these large numbers of application has not only provided solutions to a great deal of technological problems, but has also provided a prolific ground for knowledge acquisition (Fasconi & Nato, 2003).

Machine Learning Approaches

The process of machine learning involves the adoption of certain approaches that assist in .performing separate objectives or functions. These involve two most applied learning scenarios with their distinct criterion functions (Zhang et al., 2009). These two approaches are commonly referred to as supervised and unsupervised selection approaches/criteria.

Supervised approach

This may also be intimated to as discrimination or prediction classification. In this approach, algorithms are developed to levels priori-defined. The construction of algorithm takes place in the dataset training followed by comprehensive tests on independent data set to examine the algorithmic accuracy and efficiency. In the process of regression and classification, a group of support vectors that are related to methods of supervised learning. Such related vector machines include among others linear classification, which develops a straight line providing a distinct boundary between two dimensions (zhang et al., 2009). These lines may also be referred to as hyper lines which have replaced the use of the dot product for reasons of fitting in the maximum-margin. A decision tree structure may also be applies whereby classifications are represented by the leaves while feature conjunctions that direct to the classification are represented by the branches. Decision tree algorithm may be efficiently changed into a paradigm of rules of production. The supervised appr5oach also entails the use of artificial Neural Networks, a group of nodes that are interconnected that process information through the use of computational model. The information that flows through the network whether external or internal may change ANN's structure. The relationship that exists between inputs and outputs can be modeled by the use of ANN. Multi-Layer Perception (MLP) and the Radial basis function (RBF) are the most used algorithms of the ANN.

Unsupervised classification

It involves two distinct ways applicable in designing of selection criteria. They are identified on their metric of performance illustrated as classification driven criterio0n and fidelity driven criterion. Fidelity driven criterion is dependent upon the bulk of the, original information stored or discarded after the reduction of the feature dimension. Unsupervised approach operates on the basis of cluster analysis in which the method of clustering separates objects, into a number of predetermined groups assuming a pattern that increases a specific functionality.

Random forests

The term random forest is used to describe an ensemble of learning that comprises a bagging of a decision tree that has not been modified or not pruned that also exhibit a randomized identification of features in every split. Decision tree refer to combined individual learners and is popularly adopted in the exploration of data. An example of decision tree is known as CART (classification and regression tree). The random forest integrates the idea of feature selection and bagging which is used in the construction of a group of decision trees with a supervised variation. The random forest is useful in that it builds a learning algorithm that is precise and accurate hence giving a classifier which is accurate, when used on very large or expansive databases it passes over efficiently. A large number of input variables may be used with no deletion of any variable, provides an estimation of the type of variables that may be used in the classification. It also has the ability to initiate internal indiscriminate estimation of the generalization error during the continuation of forest building.it is easy to estimate the data which is missing and retains accuracy should there be a large chunk of missing data, in class population unbalanced data sets, random forests provide effective methods of error balancing. The relationship between classification and the variables are possible to identify easily due to the ability to compute prototypes for the sake of deriving such information. It is easy to cluster, locate outliers and provide vital data information due to the possibilities of proximity computation between pairs of cases. Random forest gives an experimental method that can be used in the identification of interactions between variables.

On the other hand, random forests sometimes over fit in certain datasets that shows regression functionality and noisy classification. It is also clear that humans experience a lot of difficulties in the interpretation of the random forests classification. Of importance to note is the fact that a data set of 200 random forests has been developed so that an intuitive visualization of a model space could be developed.

Data sampling

It refers to selection of a particular group or population for a particular study. A process of experiments, direct observation, and/or surveys aids sampling. This is at least when it is applied in the normal cases of research. In machine learning, data sampling helps in allocating the most appropriate data set before or during mapping. During data sampling, a sampling plan is developed to help in specifying the method inn which a particular sample would be extracted from a given set. The following planning methods may be used; simple random sampling, cluster sampling or stratified random sampling (Maindonald et al., 2010).

One of the biggest problems in handling data sampling is when one is presented with data sets that are not balanced. An inductive learner may project a black box which assumes the roles of binary labeling for instance. This is also complimented by the fact that most of the algorithms for learning initiate categorical classifiers increasing the extent of problems when imbalanced class distributions come into play. Hence, unsatisfactory classifiers emanate from such data. Using the random forests also create big problems to humans. This is because; they may produce data that is not easy to interpret by researchers. The interpretation is very essential in ensuring complete information extraction. However, one can make attempts of rebalancing the imbalanced data artificially in what is known as 'up-sampling'; when cases are replicated from the minority and down sampling if the cases are ignored from the majority. Nevertheless, artificial class rebalancing has little influence on the predictive functionality of learned classifiers.