Statistical Analysis Method for Peak force and total work in Steered molecular dynamics

User discussions
1

GROMACS version:
GROMACS modification: Yes/No
Here post your question

Is a two-sample t-test reasonable for comparing SMD results between groups?

If not T-test then what other statistical method would be appropriate and why?

I have four groups (protein-protein), 10 replicates for each protein-protein group. I want to plot the graph and perform statistical analysis for peak force and total work across 4 groups.

To create the replicates within each SMD group, I used

  • Same pressure-equilibrated .gro starting structure

  • Different velocity seeds for each replicate

  • Same pulling protocol for all groups

Would you treat the replicates within each group as an independent enough for pairwise t-tests?

1 Like
2

A few checks before applying a t-test:

Do the peaks (max force value in each simulation) assume a normal distribution in your 4 systems?

Are you equilibrating every replica briefly before starting the pulling so that the pulling is not starting from exactly the same coordinates?

3

Thank you for your response.

Each SMD replicate was initiated using different velocity seeds from the same pressure-equilibrated starting structure(.gro) under identical pulling conditions.

Does this make the replicates partially dependent? If so, how can I make them truly independent or given my simulation conditions, what other statistical test would be more appropriate? I am so confuse on the correct approach in my case.

Also the smd code for velocity generation that I used is this:



gen_vel         = yes        
gen_temp        = 310       ; temperature for Maxwell distribution
gen_seed        = 102        ; generate a random seed, for replication 2

I am not assuming normal distribution for force. Rather my force profile look like this- multimodal distribution. But this distribution does has a biological reason.

image
image1211×673 68.5 KB

4

Apologies for the very late reply. Regarding the t-test, my question was regarding the distribution of the maximum value for each type of system simulated, meaning that if you take the max value for every one of the 10 replicas and histogram them, are they approximating a normal dist? If yes, then using a t-test can be better justified.

Since you mentioned that the multi-modal dist has a biological reason, you can do two t-tests for every pair of systems, one for the local maximum (first peak) and second for the global maximum (second peak). Although you need to be wary of type I error when you do lots of t-tests at alpha=0.05, as you may get a significance by chance simply because you did too many t-tests. Consider doing a one-way ANOVA followed by Tukey HSD post-hoc test to avoid this pitfall.

Regarding independent simulations, starting from exactly the same .gro file with no further equilibration (even if for half a nanosecond) means any residual strain from equilibration can introduce a systematic bias.