reproducing result for pbmc

[fanying2015]

Thank you for sharing the script and the input pbmc_data.h5ad. I was able to reproduce the good performance with the script TAPE_realbulk.ipynb. However, when I extracted the input training and testing data from pbmc_data.h5ad, and re-ran it with the most-updated scTAPE, I was not able to reproduce the good performance.

Below is the script:
bulkdata = "data/sdy67_pbmc_T.txt"
SignatureMatrix, CellFractionPrediction =
Deconvolution(data8k,
bulkdata,
datatype='counts',
mode='overall',
adaptive=True,
genelenfile='data/GeneLength.txt',
variance_threshold=0.98,
save_model_name=None,
batch_size=128, epochs=128, seed=1)

Here's the output I got:
Overall ccc score is 0.2681609832003653
Overall L1 error is 0.09960681947553503
Monocytes's ccc score is 0.016779699438862313 l1 error is 0.16584189928137005
Unknown's ccc score is 0.018452949504123843 l1 error is 0.04238455120367027
CD4Tcells's ccc score is 0.01811648028004676 l1 error is 0.11177192915514465
Bcells's ccc score is 0.012134334312653563 l1 error is 0.03713114535164252
NK's ccc score is 0.0049701328499787535 l1 error is 0.05943935639606579
CD8Tcells's ccc score is 0.007200250637215528 l1 error is 0.1810720354653168

Any suggestion on reproducing the result? Thank you!

[poseidonchan]

Hi,

Sorry for the inconvenience. I realized there is some issue with the package but did not put effort on maintenance. Probably the first issue is the usage of "adaptive" parameter. For prediction on cell type proportions, set it as False would help reproduce the results. I will try to spare time to fix other issues. Thanks for pointing it out.

Regards,
Yanshuo

[fanying2015]

Thank you for the quick response! Turning of 'adaptive' did improve the results a lot. Looking forward to the fixed version of the package where we can do both expression prediction and fraction prediction.