Determining best MCL inflation factor

Question

[Anand Rao]

I am trying to cluster orthologs (and paralogs) at the protein level. I seem to be getting groups that have very disparate proteins, because they are of very different lengths and their alignments returned by MAFFT are extremely gappy. So I am considering playing around with the Inflation factor of MCL.

Some info about that is at http://micans.org/mcl/man/mcl.html. "A good set of starting values is 1.4, 2, 4, and 6." While I understand, in theory, the effect changing inflation factor will have on the coarseness of clustering, how can I practically determine the best inflation factor for my dataset if I do not have any extensive information on it a priori? Any thoughts? Thank you!

[h.mon]

You may set more stringent blast thresholds as well.

[Andrzej Zielezinski]

OrthoMCL uses an inflation of around 1.5 to balance sensitivity and selectivity based on grouping of enzymes and their E.C. numbers.

[Aba.j1987]

Hi Anand, Could you find a solid method to identify the best inflation rate for your MCL clustering? I used BMGE for trimming and it somehow made the MSA file better and removed many gaps. But still, I'm missing many sequences within my orthogroups.

Answer 1

[Jean-Karim Heriche]

Increasing inflation will increase granularity, that is it will produce smaller clusters. So you need to use higher values than what you've used so far to try and break up the clusters into smaller, more homogeneous ones. Also you seem to have information to use to assess clustering quality since you can tell that your current clustering is not satisfactory.

[Anand Rao]

Thanks for your response but Nope, I said "I do not have any extensive information on it a priori". How do I then practically assess what the best inflation factor is. And if I should check more values....Hope that clarifies it.

[Jean-Karim Heriche]

What I meant is that you somehow can determine the clustering quality since you find that what you get is not good enough. If you could quantify this clustering quality then you could measure it for different values of inflation. Alternatively, depending on the cluster structure you're trying to extract, other clustering algorithms may be worth considering. In my hands, MCL tends to produce very unbalanced clusters so if that's also a problem for you, you should consider another algorithm.