Introduction to Feature selection for bioinformaticians using R, correlation matrix filters, PCA & backward selection

Bioinformatics is becoming more and more a Data Mining field. Every passing day, Genomics and Proteomics yield bucketloads of multivariate data (genes, proteins, DNA, identified peptides, structures), and every one of these biological data units is described by a number of features: length, physicochemical properties, scores, etc. Careful consideration of which features to select when trying to reduce the dimensionality of a specific dataset is, therefore, critical if one wishes to analyze and understand their impact on a model, or to identify what attributes produce a specific biological effect.

For instance, considering a predictive model C1A1 + C2A2 + C3A3 … CnAn = S, where Ci are constants, Ai are features or attributes and S is the predictor output (retention time, toxicity, score, etc). It is essential to identify which of those features (A1, A2 and A3…An) are most relevant to the model and to understand how they correlate with S, as working with such a subset will enable the researcher to discard a lot of irrelevant and redundant information.

Why R for Mass Spectrometrist and Computational Proteomics

Why R:

Actually, It is a common practice the integration of the statistical analysis of the resulted data and in silico predictions of the data generated in your manuscript and your daily research. Mass spectrometrist, biologist and bioinformaticians commonly use programs like excel, calc or other office tools to generate their charts and statistical analysis. In recent years many computational biologists especially those from the Genomics field, regard R and Bioconductor as fundamental tools for their research.

R is a modern, functional programming language that allows for rapid development of ideas; it is a language and environment for statistical computing and graphics.The rich set of inbuilt functions makes it ideal for high-volume analysis or statistical studies.

Perl Proteomics & InSilicoSpectro

In contrast with genomics, bioinformaticians in proteomics don’t have a "big" and "complete" perl library for proteomics data analysis. It could be related with the "heterogeneity" in proteomics. A lot of different instruments, protocols, properties. Also genomic have a huge community (bioinformaticians) and standardize tools (instruments and software’s). In 2006 Collinge and Masselot published an open-source perl library named InSilicoSpectro. The aim was provide a set of recurrent functions that are necessary for proteomics data analysis.

Some of the Illustrative functions are: mz list file format conversions, protein sequence digestion, theoretical peptide and fragment mass computations, graphical display, matching with experimental data, isoelectric point estimation (with different methods), and peptide retention time prediction.