[maker-devel] Predictions without evidence

[Carson Holt]

Gene predictors have to be trained on each organism to generate a matched HMM. If they are not trained, they will not work well. MAKER also sends hints to the predictor based on the evidence alignments to further alter probabilities used by the predictor to better match the evidence. Evidence is also used in final filtering. All models without evidence will have an AED of 1, which means no support.

Not using evidence will result in very poor models especially if you don’t have an HMM built exactly for the organism. The main problem will be over prediction. Note the behavior of SNAP alone in the MAKER2 paper. The result is tens of thousands of false positive gene models.

If you only run multiple gene predictors without evidence, the final model will be whatever model has the best consensus structure for the set. If the set consists of two models, then there is no consensus and the longest one is kept.

—Carson



> On Nov 14, 2016, at 1:50 AM, Loïc <lmeunier at ulg.ac.be> wrote:
> 
> Hello,
> 
> I am a Ph. D. student, and I am using MAKER to automate gene prediction for many genomes as part of a genome mining work, so I don't include evidence for its use.
> If I understood well, when exploiting multiple gene predictor softwares, AED is used to define the prediction which matches the best the evidence.
> 
> So, as I don't use evidence, is there a choice made by MAKER when working with multiple gene predictors? If yes, how does it work?
> Also, I have not well understood, if the selection of the gene predictor to use is made for every gene?
> 
> Sorry to asking if the answer is obvious, but after reading your papers and looking on the archived posts, I have not found the answer.
> 
> By the way, I have also a question about your paper on MAKER2 (Holt and Yandell, 2011). It is said many times that gene predictors used in MAKER pipeline give better results than when used alone, but I have not understand why. Can you explain this fact?
> 
> Best regards,
> 
> Loïc Meunier
> 
> 
> 
> 
> _______________________________________________
> maker-devel mailing list
> maker-devel at box290.bluehost.com
> http://box290.bluehost.com/mailman/listinfo/maker-devel_yandell-lab.org