RE: Basics of MD

From: Lennart Nilsson (
Date: Tue Sep 08 2020 - 13:05:28 CDT

I fully agree with JC and Peter.
This is not an unexpected observation, and already in 1992, with the somewhat more limited resources of the time, it was reported {J. Mol. Biol. 233(4): 766-780] that “The results suggest that performing ten 100 ps simulations spans the conformation space better than one 1 ns simulation” (Quote from the Abstract). My humble nature forbids me to mention but one of the authors of the study – Arne Elofsson, who went on to a very successful career in structural bioinformatics. I fondly remember several interesting discussions with Peter Kollman on this topic.

Lennart Nilsson, Senior Professor
Karolinska Institutet
From: [] On Behalf Of JC Gumbart
Sent: Tuesday, September 8, 2020 6:46 PM
To:; Peter Freddolino <>
Cc: Raman Preet Singh <>
Subject: Re: namd-l: Basics of MD

Here is an excellent paper that addresses the second question:>

Their conclusion is "On the basis of our data, it appears that a good rule of thumb is to perform a minimum of five to 10 replicas.” In practice, admittedly we usually only run 2-3 replicas in my lab, which is still a lot better than 1.


On Sep 8, 2020, at 11:17 AM, Peter Freddolino <<>> wrote:

Dear Raman,
These are good questions, but they are also questions that are addressed in any textbook or course on molecular simulation, and so I would highly suggest that you seek out such information. Books like "Understanding molecular simulation" (Frenkel/Smit) or "Molecular Modelling: Principles and Applications" (Leach) are good starting points, as well as recordings of one of the MMBioS computational biophysics workshops (which you can find on youtube), are good starting points.

To give very brief answers to your two questions (which are absolutely not a substitute for doing a lot of your own reading and study):

On Tue, Sep 8, 2020 at 7:02 AM Raman Preet Singh <<>> wrote:

1. MD simulations are done on the scale of tens to hundreds of ns. Does this truly represent the timescales (ns) at which the observed phenomena occur in real-life situations (within a reasonable margin of error)? Has any validation been done in this regard? If the MD timescales do not really transform to real-life situations then what does ns really represent?\

ns are ns, to a reasonable approximation (there is certainly some fuzz, eg many water models have different diffusive properties from real water, etc, but these aren't order-of-magnitude differences). That means you have to be either asking carefully posed questions that can be answered by looking at ns timescale dynamics (or longer if you can run your simulations longer), or you need to be doing clever things like enhanced sampling methods, free energy calculations, etc. NAMD is good at integrating equations of motions for a bunch of charged vdW spheres connected by specific bonded potentials. Everything beyond that is up to you.

2. MD simulations are performed either as a single long simulation (upto several hundreds of ns) or multiple smaller simulations (tens of ns). Here, I would like to know what are the determinants for selecting a single long versus multiple small simulations. I would like to clarify here that by multiple small simulations I don't mean using restart files from a previous run and then continuing; rather, I mean multiple simulations all starting from the same initial configuration. Under what conditions should a single long versus multiple small simulations be used?

It is almost always going to be best to run multiple replicate simulations, each long enough to observe whatever phenomenon that you're interested in, and ensure consistency between them or use them to figure out your inferential uncertainty (just like you would not typically want to do an experiment only once). In principle, if your system is at equilibrium and you're interested in only equilibrium properties and you are sampling sufficiently, the results of a single long simulation ought to be the same as many shorter simulations, but that is a high-sampling limit that would have to be validated.

Thank you!


När du skickar e-post till Karolinska Institutet (KI) innebär detta att KI kommer att behandla dina personuppgifter. Här finns information om hur KI behandlar personuppgifter<>.

Sending email to Karolinska Institutet (KI) will result in KI processing your personal data. You can read more about KI’s processing of personal data here<>.

This archive was generated by hypermail 2.1.6 : Thu Dec 31 2020 - 23:17:14 CST