expDeconv, improved performance with marker gene selection, loss updates, while keeping lite-weighted and fast, achieving improved performance with default setting on the 32 benchmark datasets by Li et al..
expVeloImp, post-processing for imputed Spliced and Unspliced count matrices, the issue reported in Fig. 6 of VISTA paper is now fixed with updated experiment result in notebook transDeconv.ipynb. The fix was to sparsify small non-zero values (at 1e-6 magnitude) to zero, sincescv.tl.proportionscacluate non-zero gene counts to compute proportions, and recalibrate the proportion of spliced/unspliced imputations towards reference distributions.expTransImp, added batch training options to metigate OOM issues raised in VISTA paper.expTransImp, ran 45 benchmark datasets by Li et al.. The performances of TransImpLR and TransImpSpa are summarized as following:
| Group | Datasets | Ref Cells | Spots | Genes | Peak RSS (GB) |
|---|---|---|---|---|---|
| 0 | D1–D6 | 19K–34K | 175–8,425 | 4,650–31,298 | 19.47 |
| 1 | D7–D12 | 14K–48K | 645–6,963 | 7,239–15,927 | 15.11 |
| 2 | D13–D18 | 8.9K–34K | 982–11,426 | 1,296–14,248 | 10.07 |
| 3 | D19–D23 | 15K–21K | 995–4,895 | 4,815–6,177 | 3.62 |
| 4 | D24–D28 | 16K–27K | 198–2,425 | 1,910–6,177 | 3.67 |
| 5 | D29–D33 | 14K–21K | 1,835–3,794 | 3,511–13,599 | 6.62 |
| 6 | D34–D39 | 12K–25K | 369–9,852 | 3,498–8,345 | 4.49 |
| 7 | D40–D45 | 17K–28K | 603–19,522 | 980–48,163 | 24.04 |
| Summary | |||||
| --------- | -------- | ||||
| Average per group | 10.89 GB | ||||
| Median per group | 8.35 GB | ||||
| Max (group 7, D40–D45) | 24.04 GB | ||||
| Min (group 3, D19–D23) | 3.62 GB |
| Group | Datasets | Ref Cells | Spots | Genes | Peak RSS (GB) |
|---|---|---|---|---|---|
| 0 | 1,2,3,4 | 10K/10K/3.8K/6.9K | 1,000 each | 33K/20K/18K/17K | 5.73 |
| 1 | 5,6,7,8 | 10K each | 1,000 each | 21K/19K/18K/18K | 3.28 |
| 2 | 9,10,11,12 | 10K/10K/10K/2.3K | 1,000 each | 21K/16K/18K/16K | 3.29 |
| 3 | 13,14,15,16 | 1K/0.9K/1.4K/6.9K | 1,000 each | 16K/18K/13K/16K | 2.47 |
| 4 | 1r,2r,3r,4r | 10K/10K/10K/8.8K | 1,000 each | 33K/20K/18K/17K | 4.23 |
| 5 | 5r,6r,7r,8r | 10K each | 1,000 each | 21K/19K/18K/18K | 2.60 |
| 6 | 9r,10r,11r,12r | 10K/10K/10K/7.9K | 1,000 each | 21K/16K/18K/16K | 2.49 |
| 7 | 13r,14r,15r,16r | 8.5K/2.3K/1.8K/7.1K | 1,000 each | 16K/18K/13K/16K | 2.50 |
| Summary | |||||
| --------- | -------- | ||||
| Average per group | 3.32 GB | ||||
| Median per group | 2.94 GB | ||||
| Max (group 0, datasets 1–4) | 5.73 GB | ||||
| Min (group 3, datasets 13–16) | 2.47 GB |
This tool implements Translation-based imputation methods (TransImp) and translation based cell type deconvolution (TransDeconv). Experiments reported in the manuscript are displayed in jupyter notebooks under repo TranSpaAnalysis, they were run with transpa V0.1.1. Report for TransImp can be accessed from biorxiv with the latest title
Reliable imputation of spatial transcriptome with uncertainty estimation and spatial regularization
Three demo notebooks are also available under the demo folder.
- Different configurations of TransImp applied to SeqFISH dataset dataset
- Exploration for unprobed genes with SeqFISH ST dataset
- Cell type deconvolution with TransDeconv and ST Velocity estimation
- Cell type deconvolution with advanced setting
TransImp is available through PyPI. To install, type the following command line and add -U for updates:
pip install -U transpa (not yet updated to pypi, so please install from the repo)
Or, install from the repo:
pip install -U git+https://github.com/qiaochen/tranSpa
Data used for running the demo notebooks can be downloaed from Zenodo
- seqfish.ipynb and seqfish_unprobed_genes.ipynb requires input data in seqfish.tar.gz
- transDeconv.ipynb requires input data in Mouse_brain.tar.gz
Please visit TransImp website for more details.