Feature Selection Extraction

Spectra can be broken up into constituent features, which give us information about underlying signals. In NMR data, a signal gives rise to a Lorentzian 'peak' about a particular frequency. We call these peaks 'features', as they represent distinct, though often connected, pieces of information about the constituent signals in a spectrum.

We use various scripts to assist with quantification:

The maximum intensity of a peak is a convenient approximation for the signal intensity:

Peakpick1D

Some have found SRV to be useful for feature extraction (not in our toolbox): Statistical Recoupling of Variables (SRV). Ref. Blaise, B. et al (2009) Statistical recoupling prior to significance testing in nuclear magnetic resonance based metabonomics, Anal. Chem., 81(15), 6242-6251.

More robust quantification requires integration of peaks, as peak shape with width can change. Determining the boundaries of peaks is a difficult task. We utilize several tools for this:

our Manual Binning Workflow (defining rectangular bucket boundaries with the mouse)

automated bucketing using opt_bucket (from Laboratory for Theoretical and Applied Chemometrics Institute of Chemistry - University of Campinas)

For known peak boundaries, you can use IntegralPeak_roi

For CIVM data, we utilize an approach called ridge tracing/ridge tracking:

For 2D:

segment2D

bin2D