Skip to main content
. Author manuscript; available in PMC: 2012 Jun 8.
Published in final edited form as: Structure. 2011 Jun 8;19(6):844–858. doi: 10.1016/j.str.2011.03.019

Figure 6. Rotameric χ mean estimates calculated with several methods and smoothing effects of query-adaptive kernel regressions.

Figure 6

Non-overlapping 10°×10°-bin average (A), 2002 Bayesian (B), non-adaptive kernel regression (C) and query-adaptive kernel regression (D, E, F) estimates are shown for μ(χ | ϕ, ψ, r = g+, aa = Cys). The 10°×10°-bin average has only the bins with at least five g+ rotamers per bin. The non-adaptive kernel regression (C) has a fixed bandwidth (κ = 54, bandwidth radius, R = 8°), the same as for (D). The query-adaptive kernel regression estimates with widening geometric-mean kernel bandwidth are ordered from (D) to (F). The maximum log-likelihood (κ = 54, R = 8°), 5%-stepdown (κ = 29, R = 11°) and 20%-stepdown (κ = 17, R = 14°) bandwidths are in (D), (E) and (F) respectively.