How to get interaction matrix between taxa/modules (while BF > 100)

[linlin0026]

Hello, dear Professor. While studying mdsine2, I am confused about how to get the glv module's interaction with bf>100 between taxa/modules.
1)According to MDSINE2's and mdsine2_paper's GitHub, there is “Computation and visualization of forward-simulation of learned gLV model for each OTU.” after“MCMC inference using MDSINE2's model“, which is not shown in jupyter. Do we need to use md2.model.gLVDynamicsSingleClustering to get the interaction?
2）while in the fixed-clustering model， I also got the same problem with networkx version 2.6.1 not working when using the codes. And I wonder whether the output edges here are BF>100?

_mcmc = md2.BaseMCMC.load("fixed_clustering/mcmc.pkl")
md2.write_fixed_clustering_as_json(
    mcmc=mcmc,
    output_path="fixed_module_interactions.json"
)
(mdsine2_new) C:\Users\Lin\MDSINE2_Paper>conda list networkx
# packages in environment at D:\anaconda\envs\mdsine2_new:
#
# Name                    Version                   Build  Channel
networkx                  2.3                      pypi_0    pypi_

3. Can mdsine2 directly print fixed_model/full model's interaction matrix (BF>100)? For the non-clustering method, I did not find a function like md2.write_fixed_clustering_as_json, so I used the following method. However, the values in this matrix are very close to 0, which makes me wonder if my thinking needs to be corrected.

_seed = 456
mcmc = chains[seed]
study = healthy
INTERACTIONS = mcmc.graph[STRNAMES.INTERACTIONS_OBJ]
INTERACTIONS_max = md2.summary(INTERACTIONS)['mean']_

Thanks for reading and helping; I'm sorry my questions seem superficial, as I'm just beginning my journey in learning Python.

[yk23]

Hello -- this software has been under development for some time in our lab, while we were preparing our official methods publication. (I have been contributing to the software periodically after our main code author departed our group a few years ago).

I don't know if you have since stopped using MDSINE2, but for you or anyone else who wants to do something like this, we have updated our software.

Now, what you should do is invoke:
mdsine2 extract-posterior -i mcmc.pkl -o your_posterior_dir
which extracts the necessary posterior ingredients into binary .npy-formatted arrays. In particular, it should output an MxNxN array (M = # of MCMC samples, N = # of taxa) which represents the M different samples of NxN interaction matrices. Each NxN matrix is either a float or a np.nan. The np.nan indicates that the interaction was turned off for that sample.

So, to compute the posterior probability between modules (e.g. in the fixed-module setting), what you need to do is the following:

cluster_reps = [41, 3, 94, ...]  # an array of arbitrarily-chosen representative taxa indices for each module
interactions = np.load("interactions.npy")  # an NxN array
interactions_p = np.mean(~np.isnan(interactions), axis=0)
module_interactions_p = interactions_p[np.ix_(cluster_reps, cluster_reps)]

and from this, you should be able to calculate the Bayes factor (or work directly with these probabilities however you wish).

[yk23]

It's been a while since this issue was addressed, but I think a good TODO from our end would be to add a small tutorial on the math formulas (specifically the posterior probabilities and bayes factor calculations). It's on our radar.