shahcompbio / MultiModalMuSig.jl

A Julia package for extracting mutation signatures using topic models
Other
18 stars 4 forks source link

How does MultiModalMuSig Infer Mutational Signatures and Signature Exposures? #5

Closed WuyangFF95 closed 4 years ago

WuyangFF95 commented 4 years ago

Dear @funnell,

MultiModalMuSig.jl package uses a very ingenious generative model to design the dependency among all latent variables. But I'm curious about the strategy to estimate the latent variables and the mutational signatures and exposures. Do you use Expectation Maximization (EM), Markov Chain Monte Carlo (MCMC), Simulated Annealing (SA), or other strategies?

Thanks much! Yang

funnell commented 4 years ago

Hi Yang,

I’m using variational inference for all the models in this package.

Cheers!

WuyangFF95 commented 4 years ago

Thank you Tyler!

Actually, I found in your paper that you are using EM algorithm for Variational Inference.

For LDA and ILDA, parameters were inferred using mean-field Variational Bayes. For CTM, MMCTM, ICTM and IMMCTM, parameter inference was performed using mean-field variational EM. The MMCTM updates and derivations can be found in Salomatin et al. [22]. See S1 Text for detailed descriptions of the topic models.

funnell commented 4 years ago

By the way @WuyangFF95, if you post a preprint of your work at some point, I'd be interested to see the results of your comparisons!

WuyangFF95 commented 4 years ago

Sure! I'll send the link to you once it's on a preprint archive.