Songbird parameter

Hello, I'm interested in Songbird to check the correlation between microbial relative abundance and ferritin level. However, I can not get the best model because Pseudo-Q-squared is very low (negative value = -0.008)

using songbird code =>

qiime songbird multinomial
--i-table filtered_table_fea_final.qza
--m-metadata-file metadata.tsv
--p-formula "ferritinvalue"
--p-epochs 20000
--p-batch-size 20
--p-differential-prior 0.5
--p-summary-interval 1
--o-differentials differentials.qza
--o-regression-stats regression-stats.qza
--o-regression-biplot regression-biplot.qza

I tried to adjust parameters in many times but I can not get pseudo-Q-squared in positive. Could you please suggest me how to adjust them?
Thanks in advance.

Hi @cocoa, this probably suggests that you are hovering close to the noise floor. Do you see a significant difference in beta diversity?


Dear @mortonjt, I used weighted unifrac in beta diversity analysis with four groups in low, medium, high ferritin group and normal group and test permanona. I did not get the significant difference.

ok, then that explains your problem. If there is no significant difference then songbird won’t give you meaningful results.

Hello, I have a similar problem. I want to use songbird for detecting differentially abundant microbes between two categorical groups, but the model fit is really bad, the loss is huge and I don't know why.. For weighted UniFrac the same variable is sign. explaining variation in an adonis test (R²=0.04, p=0.004) and I found differentially abundant ASVs between the two groups using Ancom.

I played around with the parameters a lot, increased the Nr of samples to include in Test & Train groups, specify the samples for Train & Test groups for equal distribution, applied more or less ASV-filtering, but could not really improve the model fit yet. Do you have any idea what might be the problem? Is songbird only useful with strong effects? Or can a too diverse microbiome be a problem (I currently use a biom table with ASVs present in at least 30% of my samples)?

qiime songbird multinomial
--i-table biom.qza
--m-metadata-file metadata.txt
--p-formula "C(variable1)"
--p-num-random-test-examples 15
--p-epochs 10000
--p-batch-size 15
--p-differential-prior 0.5
--p-summary-interval 1
--o-differentials results/differentials.qza
--o-regression-stats results/regression-stats.qza
--o-regression-biplot results/regression-biplot.qza

Many thanks in advance!

Hi @rfleischer the loss is always going to be large, that isn’t something to worry about. I would focus on getting null statistics generated to see if your model can outperform the null model.