evcomplex scoring model

[cross12tamu]

I have ran different jobs using two primary scoring models, skewnormal and evcomplex. My question is regarding results from evcomplex jobs, which is, what does it mean to have a probability > 1. We have numerous pairs resulting in scores > 1, and want to verify with y'all the meaning.

@jrr-cpt

[jrr-cpt]

Jobs identical in all parameters but the scoring_model (using skewnormal) do not yield probabilities greater than 1. Given that we are analyzing two proteins, the evcomplex algorithm seemed the appropriate choice. I was surprised to see that the scoring model used on the webserver for couplings jobs is skewnormal. Since every output we have generated locally using evcomplex has probabilities greater than one, in addition to the above question, is evcomplex not the best scoring model for this type of analysis?

[cross12tamu]

Hey @thomashopf , I hope you are doing well!

I'm just wondering if you have had any time to take a look at this?

Thanks as always for your time and brain energy 😄

[thomashopf]

@aggreen Could you maybe have a look at this?

[cross12tamu]

Let me know if you need any more information @aggreen

[aggreen]

Hi all, The EVcomplex scoring model (See Hopf and Scharfe, Elife, 2014) is the published state of the art for evcomplex runs (ie, runs between interacting proteins). This scoring model can have values greater than 1 - a more appropriate name for the "probability" column would be "model_score" or something like that, I can see why having "probability" values greater than 1 is alarming. The skewnormal score is not an appropriate score for evcomplex runs based on my studies - it is not robust to the failure modes that happen when complexes have low numbers of sequences or other oddities. Stay tuned for a new complex scoring model to replace the old evcomplex model, which you can check out in Green and Elhabashy 2019, currently on Biorxiv. Anna

…

On Mon, Apr 27, 2020 at 11:14 AM cross12tamu ***@***.***> wrote: Let me know if you need any more information @aggreen <https://github.com/aggreen> — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#234 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB4KMXUNIYP25K6LFI7QAUDROWONFANCNFSM4MK3MBCQ> .

-- Anna G. Green, PhD Postdoctoral Fellow, Farhat Lab Department of Biomedical Informatics Harvard Medical School 10 Shattuck Street, Suite 514 Boston, MA 02115

[cross12tamu]

Thanks @aggreen

For the "probability (model_score)" column, is there a standard or simply a method, on understanding the values (specifically, the range of values)?

The contact map generation confuses me a bit, with default settings being 0.95/0.99 cutoffs, and me thinking of that cutoff being "probability" driven and not "model_score" driven. Although, those values are what it uses (right?) when selecting the plot.

For context this column (probability) is of interest to us, due to the resultant cn values not "jumping out". But, ec pairs from the probability/model_score column that rank high, are biologically intriguing, although again, with low cn scores.

Let me know if any of my questions are confusing, and I can reiterate/clarify. Thanks as always.

ASIDE I saw the evcomplex2 PR. 👍