Open zacmon opened 5 years ago
Hi @zacmontague ,
What exactly do you mean in the last sentence? It's unclear. Do you mean, perhaps, that the commands I've written below would allow for model_parms.txt and the like to be produced for a new species?
This is indeed what I mean: provided you want to create for new species (e.g a platypus) a recombination model with the same topology than an existing model (say e.g human heavy chain), the commands you have used will create the correct model_parms and marginals files with the custom (platypus) genomic templates you have provided
As for your second question about inference parameters:
N_iter
controls the number of iterations in the Expectation Maximization algorithm, each new iteration will increase the final likelihood of the model. From my experience this likelihood quickly plateaus after a few iterations. Again from my personal experience (which you should take with agrain of salt and check by yourself the evolution of the likelihood and model parameters) 10 iterations is more than enough to obtain robust results. This is most likely not the parameter you should tune to solve your situation with many NaN in your Pgen (or sequence likelihood?)
L_thresh
is indeed a better candidate for your problem, especially if your are dealing with highly mutated sequences (e.g ones that underwent somatic hypermutations). This parameter exists because you might want to rule out some bad quality sequences (e.g ones with many errors introduced during PCR). You should set it such that most of your sequences are assigned a likelihood starting from the first iteration (which you can monitor looking at the inference_logs.txt
file)
However in as explained in #28 the most likely reason for your NaNs come from alignment quality and I strongly encourage you to read my answer for that issue.
Hi Quentin,
In your docs, you stated in the third paragraph of Advanced Usage:
What exactly do you mean in the last sentence? It's unclear. Do you mean, perhaps, that the commands I've written below would allow for model_parms.txt and the like to be produced for a new species?
Example bash script template and IGoR commands
We first set up a convenient variable for setting all the genomic options.
We then run IGoR employing these options.
We initialize a model for the new species from another (predefined) species, here the human model in IGoR.
Recommendations for inference options
For a new species, what do you recommend be the options for
N_iter
andL_thresh
in the inference stage? Currently,N_iter
defaults to five and my Pgen values are mostly NaNs. Would increaseN_iter
improve this as well as manipulatingL_thresh
?