How to (not) perform a Molecular Dynamics simulation study

At this week’s group meeting I presented my recently published paper:

Knapp B, Dunbar J, Deane CM. Large scale characterization of the LC13 TCR and HLA-B8 structural landscape in reaction to 172 altered peptide ligands: a molecular dynamics simulation study. PLoS Comput Biol. 2014 Aug 7;10(8):e1003748. doi: 10.1371/journal.pcbi.1003748. 2014.

This paper was presented on the front page of PLoS Computational Biology:

cover

The paper describes Molecular Dynamics simulations (MD) of T-cell receptor (TCR) / peptide / MHC complexes.

MD simulation calculate the movement of atoms of a given system over time. The zoomed-in movements of the TCR regions (CDRs) recognizing a peptide/MHC are shown here:
simulation

Often scientists perform MD simulations of two highly similar complexes with different immunological outcome. An example is the Epstein-Barr Virus peptide FLRGRAYGL which leads to induction of an immune response when presented by HLA-B*08:01 to the LC 13 TCR. If position 7 of the peptide is mutated to A then the peptide does not induce an immune reaction any-more.
In both simulations the TCR and MHC are identical only the peptide differs in one amino acid. Once the simulations of those two complexes are done one could for example investigate the root mean square fluctuation of both peptides to investigate if one of them is more flexible. And indeed this is the case:

1v1RMSF

Another type of analysis would be the solvent accessible surface area of the peptide or, as shown below, the CDR regions of the TCR beta chain:

1v1SASA

… and again it is obvious that those two simulations strongly differ from each other.

Is it then a good idea to publish an article about these findings and state something like “Our results indicate that peptide flexibility and CDR solvent accessible surface area play an important in the activation of a T cell” ?

As we know from our statistics courses a 1 vs 1 comparison is an anecdotic example but those often do not hold true for larger sample sizes.

So, let’s try it with many more peptides and simulations (performing these simulations took almost 6 months on the Oxford Advanced Research Computing facility). We split the simulations in two categories, those which are more immunogenic and those which are less immunogenic:

compareAll

Let’s see what happens to the difference in RMSF that we found before:

allRMSF

So much for “More immunogenic peptides are more/less flexible as less immunogenic ones” …

How about the SASA that we “found” before:

allSASA

… again almost perfectly identical for larger sample sizes.

And yes, we have found some minor differences between more and less immunogenic peptides that hold true for larger sample sizes but non of them was striking enough to predict peptide immunogenicity based on MD simulations.

What one really learns from this study is that you should not compare 1 MD simulation against 1 other MD simulation as this will almost certainly lead to false positive findings. Exactly the same applies for experimental data such as x-ray structures because this is a statistical problem rather than a prediction based on. If you want to make sustainable claims about something that you found then you will need a large sample size and a proper statistical sample size estimation and power analysis (as done in clinical studies) before you run any simulations. Comparing structures is always massive multiple testing and therefore high sample sizes are needed to draw valid conclusions.

link pdf

Leave a Reply