huggingbahl21 commited on
Commit
b7e3f56
·
verified ·
1 Parent(s): a022bb6

Simplify CausalBench sources, preprocessing, and verification notes

Browse files
Files changed (1) hide show
  1. real/causalbench/README.md +16 -25
real/causalbench/README.md CHANGED
@@ -1,38 +1,29 @@
1
  # CausalBench K562 and RPE1
2
 
3
- Filtered single-cell CRISPRi perturbation data from [Replogle et al. (2022)](https://doi.org/10.25452/figshare.plus.20029387.v1), prepared following [CausalBench (Chevalley et al., 2025)](https://doi.org/10.1038/s42003-025-07764-y).
4
 
5
- | Dataset | Cells | Gene columns | File |
6
  | --- | ---: | ---: | --- |
7
- | K562 | 162,751 | 622 | `data/causalbench_k562.tsv.gz` |
8
- | RPE1 | 162,733 | 383 | `data/causalbench_rpe1.tsv.gz` |
9
 
10
- ## Format
11
 
12
- Files are gzip-compressed, tab-separated tables with a header and no index column. Rows are cells. Gene columns use Ensembl IDs and contain normalized, natural-log1p-transformed expression values. An additional `intervention` column contains the targeted gene ID, `non-targeting` for controls, or `excluded` for targets not retained by preprocessing. `excluded` rows are not controls.
13
 
14
- No complete ground-truth causal graph or causal-discovery expert constraints are supplied.
15
 
16
- ## Sources and preparation
17
 
18
- The [source repository](https://huggingface.co/datasets/huggingbahl21/causalbench) contains the raw inputs, preparation scripts, environment, and prepared outputs. Original inputs are [K562 counts](https://ndownloader.figshare.com/files/35773219), [RPE1 counts](https://ndownloader.figshare.com/files/35775606), and the [perturbation summary workbook](https://ars.els-cdn.com/content/image/1-s2.0-S0092867422005979-mmc2.xlsx).
 
 
 
19
 
20
- Preparation follows the [CausalBench preprocessing function at commit 1a2143c](https://github.com/causalbench/causalbench/blob/1a2143cffdc85f835b41ce8d52034be1bf903e71/causalscbench/data_access/utils/preprocessing.py):
21
 
22
- 1. Retain controls and strong perturbations: more than 50 differentially expressed genes, knockdown at most -0.3, and more than 25 filtered cells in the summary.
23
- 2. For measured targets, retain perturbed cells whose target expression is at or below the control 10th percentile.
24
- 3. Remove zero-total cells, normalize totals to the retained-cell median using all original genes, and apply natural log1p.
25
- 4. Keep measured genes targeted in more than 100 retained cells. Preserve gene order and relabel unsupported targets `excluded`, retaining their rows.
26
 
27
- K562 used the original preprocessing implementation; RPE1 used a memory-saving implementation of the same rules. See the source repository for reproduction instructions. These tables mix control and perturbed cells. The target-expression filter affects observed knockdown strength.
28
 
29
- ## Attribution and license
30
-
31
- Data are distributed under **CC BY 4.0**, credited to Replogle et al., as documented in the source repository. Cite the original data and the CausalBench paper when using these prepared matrices. Preparation code is available in the source repository with its original attribution and licensing.
32
-
33
- ## SHA-256
34
-
35
- ```text
36
- 455024efc9f40b877f7f9c1d55f9eaaf957c7995c6289ffbcd65251fc856fee5 causalbench_k562.tsv.gz
37
- 6ddd100eff9dbcab65ff94b662a87e9eb50de6401b5242c0f0108692c92cc408 causalbench_rpe1.tsv.gz
38
- ```
 
1
  # CausalBench K562 and RPE1
2
 
3
+ Filtered single-cell CRISPRi data from [Replogle et al. (2022)](https://doi.org/10.25452/figshare.plus.20029387.v1), prepared following [CausalBench](https://doi.org/10.1038/s42003-025-07764-y). Data license: CC BY 4.0, credited to Replogle et al.
4
 
5
+ | Dataset | Cells | Genes | Prepared table |
6
  | --- | ---: | ---: | --- |
7
+ | K562 | 162,751 | 622 | [causalbench_k562.tsv.gz](data/causalbench_k562.tsv.gz) |
8
+ | RPE1 | 162,733 | 383 | [causalbench_rpe1.tsv.gz](data/causalbench_rpe1.tsv.gz) |
9
 
10
+ Each gzip TSV contains gene-expression columns (Ensembl IDs) and an `intervention` column: a targeted gene ID, `non-targeting` for controls, or `excluded` for targets not retained by preprocessing. `excluded` cells are not controls.
11
 
12
+ ## Sources and preprocessing
13
 
14
+ Inputs are the original [K562 counts](https://ndownloader.figshare.com/files/35773219), [RPE1 counts](https://ndownloader.figshare.com/files/35775606), and [perturbation summary workbook](https://ars.els-cdn.com/content/image/1-s2.0-S0092867422005979-mmc2.xlsx), using sheets `TabB_K562_day6_summary_stat` and `TabC_RPE1_summary_statistic`.
15
 
16
+ The reference is the authors' [preprocessing implementation at commit 1a2143c](https://github.com/causalbench/causalbench/blob/1a2143cffdc85f835b41ce8d52034be1bf903e71/causalscbench/data_access/utils/preprocessing.py), with strong-perturbation filtering enabled:
17
 
18
+ 1. Keep controls and perturbations with >50 differentially expressed genes, knockdown <= -0.3, and >25 filtered cells in the summary.
19
+ 2. For measured targets, keep perturbed cells whose target expression is at or below the control 10th percentile.
20
+ 3. Remove zero-total cells, normalize cell totals to the retained-cell median using all original genes, and apply natural `log1p`.
21
+ 4. Keep measured genes targeted in >100 retained cells; relabel unsupported targets `excluded` without dropping their rows.
22
 
23
+ K562 used the original function. RPE1 used a memory-saving implementation of the same rules. [Raw inputs, preparation code, and environment](https://huggingface.co/datasets/huggingbahl21/causalbench) are available for reproduction.
24
 
25
+ ## Verification
 
 
 
26
 
27
+ Saved checks found that both raw count files matched the publisher's MD5 checksums. Full reconstruction from raw inputs using chunked code matched every saved NPZ expression value and the ordered gene and intervention labels. TSV export preserved those labels; its maximum absolute numerical difference from the NPZ was `2.3841e-7` for both datasets.
28
 
29
+ A small fixture also matched the unmodified authors' function. The full datasets were not rerun through that function during verification; these checks establish consistency with the implemented preprocessing, not an independent comparison against author-distributed processed matrices. The summary workbook was read locally because its remote verification request returned HTTP 403.