On Thu, Feb 14, 2013 at 7:50 AM, Pavel Afonine
The algorithm implemented in Phenix is fast: it should take from a few seconds for small structures to a few minutes for large ones. I do not understand why it should take long time to run (as pointed out in that Acta D paper).
I suspect that's because they're running a much different algorithm. The Phenix implementation doesn't reproduce the difference densities they display, for what it's worth, but since neither the code or even the binaries for the ENIGMA program are available (!), it's hard to know exactly what they're doing differently.
I see that phenix.maximum_entropy_map is now a command in Phenix. Some quick questions: Where is this likely to be must useful and does it take ridiculously long to run? From the Nishibori 2008 paper in Acta D it seems like this would mainly be useful for very high resolution structures that you would normally call complete - and that it would take a very long time to compute.
I must say, I find that paper very misleading - the conventional maps from Phenix are sufficient to identify the alternate conformation of Tyr33 in Figure 2, for instance. The published structure doesn't have *any* alternate conformations, which at 1.3Å resolution is absurd, so it's very easy to produce an improved model without doing anything fancy. In Figure 5 they compare a conventional omit map with the MEM version, but they're using much different grid spacings, so of course they look different! Maximum entropy tends to be used most frequently by small-molecule crystallographers looking at charge densities, which is partly what the Nishibori paper is doing. For proteins, this is what Nicholas Glykos (author of GraphEnt) told me about its use: "For well behaved and complete data the maps look very similar. But in other cases the ability of the maxent map to alleviate the problems arising from series termination errors made a difference. We had one case of [redacted] that diffracted to ~0.8A. Four passes were made to measure both high and low resolution data. Unfortunately, for one of the passes the time-per-frame was completely wrong and we ended-up with a data set missing all terms between ~4 and ~3 Angstrom. The conventional FFT had numerous peaks arising from the series termination errors, the maxent map was significantly better. In other cases, we use maxent to artificially sharpen maps (by reducing the esd's) while avoiding the noise introduced by normal (E-value-based) sharpening." -Nat