Applying the model from one sample to another

[camelest]

Hi, thank you so much for the wonderful tool.

I'm trying to integrate several samples together following the discussion of #33 but still gets some sample-specific factors that are a bit hard to be interpreted.

Do you think it makes sense to apply the model obtained from sampleA to sampleB?
I was hoping one of the below would work:

1. Project sampleA-fit model to sampleB by ficture transfrom

ficture transform --input ${sampleB}_gene.tsv.gz --output_pref ${sampleB}/analysis/nF40.d_15/nF40.d_15.prj_15.r_4.0 \
   --model ${sampleA}/analysis/nF40.d_15/nF40.d_15.model.p --key Count --major_axis X --hex_width 15 --n_move 3 \
   --min_ct_per_unit 20 --mu_scale 1.0 --thread 20 --precision 2

2. Use the model file from sampleA as a pseudo-bulk count by init_model_from_pseudobulk followed by ficture choose_color etc.
   (Will this make any difference from option#1?)

ficture init_model_from_pseudobulk --input ${sampleB}/hexagon.d_15.tsv.gz  --output ${output} \
    --feature ${sampleB}_gene.tsv.gz --epoch 0 --scale_model_rel -1 \
    --reorder_factors --model ${sampleA}.model_matrix.tsv.gz \
    --key Count --thread 8 --min_ct_per_feature 20

Thank you very much in advance!

Best,

[Yichen-Si]

With option 1, you will project the model learned from sample A directly; with option 2 and --epoch 0 it should be the same as option 1, but if you specify --epoch 1 or larger the input model is used as a (strong) prior and will be updated with the data from sample B.
Unfortunately we don't have a great idea for how to deal with batch effect yet, but I'll be interested in hearing what you find.

[camelest]

Thank you very much for your prompt responses.

I will try both options and get back here.
Sorry for further bothering you. I'm a bit lost which file I should use as ${model} in both ficture transform and ficture slda_decode after option#2.
Is my understanding as the below correct?

For epoch0,

ficture init_model_from_pseudobulk --input ${sampleB}/hexagon.d_15.tsv.gz  --output ${output} \
    --feature ${sampleB}_gene.tsv.gz --epoch 0 --scale_model_rel -1 \
    --reorder_factors --model ${sampleA}.model_matrix.tsv.gz \
    --key Count --thread 8 --min_ct_per_feature 20

gave me ${output}.fit_result.tsv.gz and ${output}.posterior.count.tsv.gz. Then, following the documentation and the discussion #48, do we use this posterior.count as a model.p file or...?

cp ${output}.posterior.count.tsv.gz ${output}.model_matrix.tsv.gz

# use this posterior. count as a model.p file or...?
ficture transform --input ${sampleB}_gene.tsv.gz --output_pref ${output} \
    --model ${output}.model_matrix.tsv.gz --key Count --major_axis X --hex_width 15 --n_move 3 \
    --min_ct_per_unit 20 --mu_scale 1.0 --thread 20 --precision 2

For epoch1 or larger,

ficture init_model_from_pseudobulk --input ${sampleB}/hexagon.d_15.tsv.gz  --output ${output}_epoch1 \
    --feature ${sampleB}_gene.tsv.gz --epoch 1 --scale_model_rel -1 \
    --reorder_factors --model ${sampleA}.model_matrix.tsv.gz \
    --key Count --thread 8 --min_ct_per_feature 20

gave me ${output}_epoch1.model.p as well. So should I use this file in the next?

ficture transform --input ${sampleB}_gene.tsv.gz --output_pref ${output}_epoch1 \
    --model ${output}_epoch1.model.p --key Count --major_axis X --hex_width 15 --n_move 3 \
    --min_ct_per_unit 20 --mu_scale 1.0 --thread 20 --precision 2

Thank you again for your kind help.

[Yichen-Si]

Yes, you can use posterior.count.tsv.gz in the first case and .model.p in the second.