How many points needed for the smooth?
Posted: Thu Aug 28, 2008 4:01 pm
by hover
How to decide how many points needed for the smooth process? I run the LC/MS/MS experiments to quantitate analytes. In Masslynx, it seems to select the points number equal to half peak width. Is it the best choice for others too?
Posted: Fri Aug 29, 2008 1:30 pm
by lmh
It depends on the smoothing algorithm. For instance, Gauss and Box-car smoothing broaden peaks horrendously, and can ruin a well-shaped peak; use them with only a very few points. If you have 15 points across a peak, I wouldn't dare go beyond 3 (or 5 in an emergency) with these.
Savitsky-Golay smoothing, on the other hand, is an extremely clever way to fit a polynomial through nearby points (with computational costs similar to Gauss and Box-car), and make the new point a point from this polynomial. It can match the curve at the start of a chromatographic peak without messing it up, so it can be used with quite substantial numbers of points, without any damage to peak-shape. For the 15-point peak-width, depending on the polynomial order that your algorithm uses, you may be able to use 11-15 point smoothing without trouble.
Frankly, I don't know why anyone offers Gaussian or boxcar smoothing, except for backward-compatibility to old methods, since both are thoroughly brutish. Gaussian is also open to a dangerous misuse: If you have a scan speed so slow that something runs as a peak only 1-2 points wide, a Gaussian smooth can make it into a beautiful Gaussian peak.