Virtual Cell Perturbation Transcriptomics Dataset ELEM-1200RS

• Note on Data Release Schedule:

Upon approval of the application, the complete sample information table can be downloaded. Additionally, sequencing data for 34 wild-type (WT) cells are scheduled to be uploaded by October 23, 2026.

Dataset Summary

The ELEM Biotech Virtual Cell Perturbation Transcriptomics Dataset "ELEM-1200RS" [V1.0], comprises approximately 1,200 bulk RNA-seq samples derived from 63 wild-type cell lines, with knockdown experiments targeting 914 unique genes. Gene expression data are provided as TPM-normalized expression matrices in .h5ad format, making the dataset directly suitable for downstream bioinformatics analyses and machine learning workflows (e.g., cell perturbation prediction and gene regulatory network inference), while maintaining the data structure adopted by the Virtual Cell Challenge.

Data Structure

File1: 'ELEM_1200RS_P1.h5ad'

File2: 'ELEM_1200RS_P2.h5ad'

File3: 'ELEM_1200RS_P3.h5ad'

adata.X TPM-normalized expression matrix (Samples × Genes) 'adata.X' contains the TPM-normalized gene expression matrix corresponding to the same type of RNA-seq samples. The TPM values are derived from the raw RNA-seq expression measurements using transcript/gene abundance normalization.

adata.obs 'sample_id': Unique identifier for each RNA-seq sample.

adata.var 'gene_id': Ensembl Gene Identifier (e.g., 'ENSG00000000457.15'). 'gene_symbol': HGNC gene symbol (e.g., 'SCYL3'). 'gene_type': Gene feature type (e.g., 'protein_coding', 'lncRNA').

File4: batch0908_0911.result.tar.gz

File5: batch0917_0922.result.tar.gz

File6: batch0928.result.tar.gz

The dataset is organized hierarchically into sample directories following the pattern //. Each directory contains expression quantification tables and quality control (QC) metrics:

1. Main Expression Results .gene.TPM.xls Gene-level TPM expression table for this sample. Columns: Gene (Ensembl ID), Symbol (Gene Symbol), Chr (Chromosome), GeneType (Biotype), TPM. .gene.reads.xls Gene-level raw read count table for this sample. Columns: Gene, Symbol, Chr, Start-End, Strand, GeneLength, ExonLength, Depth, Coverage, Reads. _expression_profile.txt Raw gene expression profile. 17 columns, including TPM, SenseReads, AntisenseReads, IntronReads, Depth, Coverage, and other metrics.

2. Quality Control (QC) Reports QC.clean.tsv: Read cleaning QC. QC.mapping.tsv: Alignment QC. QC.expr.tsv: Expression QC. _randCheck_gene.{txt,pdf,png}: Uniformity of gene 5'→3' read distribution. _randCheck_mRNA.{txt,pdf,png}: Uniformity of mRNA 5'→3' read distribution. _randCheck_region.{txt,pdf,png}: Uniformity of read distribution across 5'UTR/CDS/3'UTR. _fastp.html: Interactive fastp QC report (read cleaning). _hisat2_stat.txt: HISAT2 alignment summary.

File7: 1171_sampleID_Cell_name_gene_name_update_20261006.xlsx

This Excel file provides annotation metadata and correction records for the 1,171 samples included in the dataset.

Sheet1 contains annotation metadata with the following columns: 'Sample ID': Unique identifier of each sample. 'KO/KD Target Gene Name': Name of the gene targeted for knockout or knockdown. 'UniProt ID': UniProt ID of the KO/KD target gene. 'WT Cell Name': Cell line names. 'Sample Type': Sample type, including KO Pool, KO Cell Line, and WT Cell. 'Extraction/QC Time': Time of RNA extraction and quality control. 'RNA A260/A280': A260/A280 ratio of the extracted RNA. 'Library Preparation Batches and Chips': Library preparation batch and sequencing chip information. 'Note': Revision.

Sheet2 contains an error list documenting the identified annotation errors and the corresponding correction information.

Data Processing Pipeline

1. Experimental Setup: Human cells were subjected to targeted perturbations alongside non-targeting controls.
2. Sequencing: RNA was extracted and sequenced on GeneMind platforms (150 bp paired-end).
3. Bioinformatics & Quantification: Reads were aligned to the human reference genome (GRCh38 v49) using 'HISAT2'. Matrices were assembled into an 'AnnData' object with standardized 'obs' and 'var' annotations.

Quick Start & Usage

You can load the '.h5ad' file directly using 'anndata' or 'scanpy' in Python.

I. Updates to Sample Annotations

1. Issue: Discrepancy between target gene name and UniProt ID. For sample LM2303019-78, the gene name is listed as CTBP1, but the UniProt ID is P56545 (CTBP2). Expression data shows CTBP2 downregulation and unchanged CTBP1 levels; additionally, the adjacent sample LM2303019-79 involves CtBP1 knockdown. Please confirm whether the target gene for this sample is CTBP2. Response: Confirmed as CTBP2; the record has been updated. Note: The UniProt ID represents the most accurate information.

2. Issue: Isoform-specific targets. The UniProt IDs for the following samples include isoform suffixes. Please confirm whether these represent isoform-specific knockdowns (i.e., the siRNA/sgRNA targets only that specific isoform) or simply reflect annotation conventions: FGFR3 P22607-2: LM02-1484-065 (T24), LM02-1484-168/169/170 (UMUC-3)，PTK7 Q13308-6: LM02-1484-049，TMEM230 Q96A57-2: LM02-1484-118，FOXA2 Q9Y261-2: LM02-1484-130 Response: None of the above are isoform-specific knockdowns (sgRNA designs for the gene knockout strategy targeted shared transcript regions); the entries have been updated to use UniProt IDs without suffixes.

3. Issue: TE1-wt naming note. The metadata indicates that the original sequencing file labeled "TW1-wt" should be "TE1-wt." Please confirm whether this has been consistently corrected in the h5ad and tar packages, and whether there are any other similar sample renaming issues. Response: The h5ad file has been updated. However, as the error originated from the sequencing sample coding, the raw data tar package remains unchanged. The other samples have been thoroughly reviewed, and no name changes were made.

II. Experimental Design

4. Question: Perturbation method: The metadata header lists "KO/KD," while the README specifies "knockdown." Please clarify the perturbation technique used for each sample (e.g., siRNA, shRNA, CRISPR KO, CRISPRi) and whether the method is consistent across all parts; if there are differences, please provide a field specifying the perturbation method for each sample. Response: All 1,171 samples in the ELEM-1200RS dataset were generated using CRISPR KO. There are three cell types: WT cells (unedited), KO Pools (knockout efficiency >70%), and KO Cell Lines (knockout efficiency = 100%).

5. Question: Definition of controls: In the metadata, control samples are labeled as 0 (P2), / (P3), or null (HAP1-WT in P1), and most sample names end in "-WT." Please confirm whether these are untreated wild-type cells or negative controls transfected with non-targeting siRNA/sgRNA. Response: The control samples consist of wild-type cells that have not undergone any editing.

6. Question: Insufficient control coverage: P1 contains 543 HAP1 perturbation samples but only one control (HAP1-WT); the 280 HAP1 samples in P2 lack batch-matched controls. Are there any unpublished HAP1 control samples available? Response: Due to project timelines, no data for new HAP1 control samples were available prior to October 12. Data for new HAP1 wild-type cells are expected to be obtained and uploaded by October 23.

7. Question: The following 13 cell lines lack any control samples: HCT116, RD, Patu8988, SNUC5, BJ-5ta, LX-2, HaCaT, Caco-2, PANC-1, RBE, NCI-H460, T24, and MRC-5. Is there corresponding wild-type data available to supplement these? Response: Due to project scheduling constraints, no new data for control samples will be available before October 12. We will organize RNA-seq sequencing for the aforementioned wild-type cell samples (totaling 34 samples: 13 × 2 = 26 samples, plus HAP-1 × 6 and A549 × 2) during the first week following the National Day holiday. We anticipate obtaining and uploading the new wild-type cell data by October 23.

8. Question: Biological replicates: For 833 out of 908 target genes, only one sample exists. Are there biological replicates? Do multiple samples for the same target gene in the same cell line (e.g., GFPT2, FGFR3, ACOD1) represent biological replicates, or do they involve different siRNAs/sgRNAs? If the latter, could you provide the guide/siRNA IDs? Additionally, what is the meaning of the "-1" suffix in "LM01-1484-0084-1"? Does it indicate a re-sequenced sample? Response: There are two types of biological replicates in this dataset: 1. For the same target in the same cell line, there are 2–3 separate assays representing distinct biological samples (i.e., 2–3 homozygous clones); 2. For the same target across different cell lines, the same sgRNA sequence was selected to edit that specific target.

9. Question: Sampling time points and conditions: Were the post-perturbation sampling times (e.g., 48h, 72h, 7d) and culture conditions standardized? If not, could you provide sample-specific information? Response: The samples originated from projects accumulated by the Liuman team over the past 2–3 years. Sampling times were relatively consistent for samples with the same cellular background (e.g., HAP-1) but varied across different cellular backgrounds; generally, samples were taken directly for testing after the cells had been thawed from cryopreservation.

10. Question: Knockdown efficiency validation: Based on TPM calculations, approximately 38% of the evaluable samples showed a reduction in target gene expression of less than 1.4-fold (log2FC > −0.5). Have knockdown efficiency validations (such as qPCR or Western blot) been performed? Could you provide the results or the QC standards used for the knockdown efficiency threshold? Response: The knockdown efficiency for samples in this dataset was confirmed to be >70% via PCR-Sanger sequencing, or verified using Western blot/mass spectrometry. Previous experience indicates that RNA-seq data does not always fully reflect changes at the gene or protein level.

III. Consistency between data and metadata

11. Question: Regarding the statistical basis for counts: The README lists 914 target genes and 63 cell lines. Deduplicating based on original metadata names (case-insensitive) yields exactly 914 entries; however, this includes aliases for the same gene (e.g., APE1/APEX1, MB21D1/CGAS, EPCR/PROCR, HADHSC/HADH, STING/STING1, MHC class I/HLA-A), resulting in 908 unique genes after merging. The cell lines normalize to 60 (there were 77 original variations, such as THP1/THP-1, H1299/NCI-H1299, U87MG/U87-MG, hela/HeLa). Please confirm: Do the aforementioned aliases indeed refer to the same target? Response: UniProt IDs provide the most consistent and accurate reference. Wild-type cell line names have been revised to a single, standardized name.

12. Question: Regarding the count of 63 cell lines: Is a distinction made between HEK293 and HEK293T, or between the T and S sub-strains of Patu8988? Response: A clear distinction is made between HEK293 and HEK293T. No distinction is made for Patu8988.

13. Question: Protein names in P1: Approximately 80 targets in P1 are listed by protein name rather than gene symbol (e.g., "14-3-3 epsilon," "MHC class I," "PI3-kinase p110 subunit beta"). We mapped the entries to gene symbols using UniProt IDs; specifically, "MHC class I" corresponds to the merged entry P16189 (now P04439 / HLA-A). Please confirm that the target gene for the MHC class I sample is HLA-A, rather than HLA-B/C or B2M. Response: Confirmed; the target gene for the MHC class I sample is HLA-A. Some protein names have been updated to gene names.

14. Question: Regarding the October 1st update: The P2 h5ad file and the batch0917_0922 tar archive were re-uploaded on 2026-10-01 (file sizes remain unchanged). Please specify the changes made in this update (e.g., numerical corrections, sample replacements, or metadata modifications) and explain the meaning of "opti" in the P2 metadata filename. Response: "opti" indicates only that the sample name "TW1-wt" was changed to "TE1-wt"; no other content was modified.

15. Question: Inconsistent gene sets: The gene counts differ across the three h5ad files (58,989 / 64,866 / 62,148), likely due to the exclusion of all-zero genes. Please confirm if this is the reason and whether the same annotation version (GRCh38 GENCODE v49) was used for all three. Response: Yes, the variation in gene counts is due to the exclusion of all-zero genes from the set. In the gene expression file for each sample, genes that are not detected are omitted; the gene set file represents the union of results from all samples in the batch, so all-zero genes do not appear. We confirm that the same annotation version (GRCh38 GENCODE v49) was used for all three batches.

IV. Quality Control

16. Question: Low-quality samples: The following samples exhibit low alignment rates (< 90%) or gene detection counts falling within the bottom 1% of the entire dataset. Please confirm whether these samples fall into the category of known problematic samples and if their exclusion is recommended. Additionally, the overall alignment rate for batch0917_Chip1 (median 88.6%) is significantly lower than that of other chips (92–98%); is there a known issue with this chip? The samples in question are: LM01-1484-0005, -0012, -0031, -0033, -0047, -0058, -0067 (batch0917_Chip1), LM01-1484-0084-1, and LM02-1484-029, -030, -053, -126. Response: It is currently impossible to confirm whether the issues with the aforementioned samples stem from the samples themselves. Please handle the samples (i.e., decide whether to exclude or retain them) based on your own quality control criteria.

17. Question: Batch information—could you provide details such as the library preparation batch, RNA extraction date, and RIN values ​​for the samples so that batch correction can be performed alongside the sequencing chip batch information? Response: The basic information mentioned above has already been provided (please refer to the latest sample information table: 1171_sampleID_Cell_name_gene_name_update_20261006).

License and Data Use Terms

Allowed Scope of Use

Permission is granted to use this dataset for the training, fine-tuning, and academic research of AI models.

Attribution Requirement

Every paper, preprint, abstract, poster, presentation, public model card, dataset card, benchmark page, repository page, press release, customer-facing technical material, competition disclosure, and public or regulatory submission that makes Material Use of the Data Release shall identify the exact dataset version and include the following acknowledgment, with only venue-required formatting changes: “This work used the ELEM Biotech Virtual Cell Perturbation Transcriptomics Dataset, Release [ELEM-1200RS V1.0], provided under the ELEM 2026 Priority Access Program.”

Derivative Works Limitation

Any derivative datasets generated based on this dataset must be released and open-sourced under the same license agreement.

Contact

For questions regarding the dataset, please open an issue in this Hugging Face repository or contact us by email at yongbo_cheng@elem-bio.cn.

Disclaimer

This dataset is provided for research purposes. The dataset authors make no guarantees regarding the completeness, accuracy, or suitability of the data for any particular application.

Users are responsible for verifying the provenance, preprocessing status, and licensing conditions of the data before using it in downstream research.