rewirebio.iobenchmarks
Source

HyCell: a Compact Embedding Model for Single-Cell Transcriptomes (Hyra Team, 2026-11-01)

Dated 2026-11-01 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (9 pages) · Project source

Original source

What this source contributed

Records in this release that cite this source. A record can cite several sources.

Results

Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.

Model or configurationMetricValueDatasetEvaluation
HyCell fine-tuned annotator (HyCell report)Macro F1 (Macro-F1 over the cell types of the held-out batch after fine-tuning on the other batches, averaged over held-out batches)0.877 fractionscIB human immune atlas, integer-count part (HyCell report)Fine-tuned on annotation-finetuned, immune (HyCell report)
HyCell fine-tuned annotator (HyCell report)Macro F1 (Macro-F1 over the cell types of the held-out batch after fine-tuning on the other batches, averaged over held-out batches)0.794 fractionscIB human lung atlas, integer-count part (HyCell report)Fine-tuned on annotation-finetuned, lung (HyCell report)
HyCell fine-tuned annotator (HyCell report)Macro F1 (Macro-F1 over the cell types of the held-out batch after fine-tuning on the other batches, averaged over held-out batches)0.907 fractionscIB human pancreas atlas, integer-count part (HyCell report)Fine-tuned on annotation-finetuned, pancreas (HyCell report)
HyCell fine-tuned annotator (HyCell report)Macro F1 (Macro-F1 over the cell types of the held-out batch after fine-tuning on the other batches, averaged over held-out batches)0.966 fractionPBMC from two 10x Genomics runs (HyCell report)Fine-tuned on annotation-finetuned, pbmc (HyCell report)
Geneformer V2 (HyCell report)Macro F1 (Macro-F1 of a 15-nearest-neighbour classifier trained on the other batches, over the cell types of the held-out batch, averaged over held-out batches; per atlas or averaged over the seven atlases)0.782 fractionSeven held-out atlases of the HyCell reportGeneformer V2 on annotation-transfer-knn, seven-held-out-atlases (HyCell report)
Geneformer V2 (HyCell report)Macro F1 (Macro-F1 of a logistic-regression classifier trained on the other batches, over the cell types of the held-out batch, averaged over held-out batches; per atlas or averaged over the seven atlases)0.834 fractionSeven held-out atlases of the HyCell reportGeneformer V2 on annotation-transfer-lr, seven-held-out-atlases (HyCell report)
Geneformer V2 (HyCell report)Scib batch correction score (Mean of five scIB batch-correction metrics (batch silhouette, iLISI, kBET, graph connectivity and principal-component regression), per atlas or averaged over the seven atlases)0.465 unitlessSeven held-out atlases of the HyCell reportGeneformer V2 on batch-correction, seven-held-out-atlases (HyCell report)
Geneformer V2 (HyCell report)Scib bio conservation score (Mean of seven scIB biological-conservation metrics (isolated-label score, NMI and ARI of k-means and of Leiden clusters against the cell types, cell-type silhouette and cLISI), per atlas or averaged over the seven atlases)0.606 unitlessSeven held-out atlases of the HyCell reportGeneformer V2 on bio-conservation, seven-held-out-atlases (HyCell report)
Geneformer V2 (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.630 unitlessAsian Immune Diversity Atlas, sample (HyCell report)Geneformer V2 on composite-score, aida (HyCell report)
Geneformer V2 (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.756 unitlessMultiVINE brain vasculature, healthy donors (HyCell report)Geneformer V2 on composite-score, brain (HyCell report)
Geneformer V2 (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.634 unitlessCross-tissue immune atlas, stratified sample (HyCell report)Geneformer V2 on composite-score, cross-tissue-immune (HyCell report)
Geneformer V2 (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.683 unitlessscIB human immune atlas, integer-count part (HyCell report)Geneformer V2 on composite-score, immune (HyCell report)
Geneformer V2 (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.653 unitlessscIB human lung atlas, integer-count part (HyCell report)Geneformer V2 on composite-score, lung (HyCell report)
Geneformer V2 (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.579 unitlessscIB human pancreas atlas, integer-count part (HyCell report)Geneformer V2 on composite-score, pancreas (HyCell report)
Geneformer V2 (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.816 unitlessPBMC from two 10x Genomics runs (HyCell report)Geneformer V2 on composite-score, pbmc (HyCell report)
Geneformer V2 (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.679 unitlessSeven held-out atlases of the HyCell reportGeneformer V2 on composite-score, seven-held-out-atlases (HyCell report)
HyCell (HyCell report)Macro F1 (Macro-F1 of a 15-nearest-neighbour classifier trained on the other batches, over the cell types of the held-out batch, averaged over held-out batches; per atlas or averaged over the seven atlases)0.873 fractionscIB human immune atlas, integer-count part (HyCell report)HyCell on annotation-transfer-knn, immune (HyCell report)
HyCell (HyCell report)Macro F1 (Macro-F1 of a 15-nearest-neighbour classifier trained on the other batches, over the cell types of the held-out batch, averaged over held-out batches; per atlas or averaged over the seven atlases)0.758 fractionscIB human lung atlas, integer-count part (HyCell report)HyCell on annotation-transfer-knn, lung (HyCell report)
HyCell (HyCell report)Macro F1 (Macro-F1 of a 15-nearest-neighbour classifier trained on the other batches, over the cell types of the held-out batch, averaged over held-out batches; per atlas or averaged over the seven atlases)0.838 fractionscIB human pancreas atlas, integer-count part (HyCell report)HyCell on annotation-transfer-knn, pancreas (HyCell report)
HyCell (HyCell report)Macro F1 (Macro-F1 of a 15-nearest-neighbour classifier trained on the other batches, over the cell types of the held-out batch, averaged over held-out batches; per atlas or averaged over the seven atlases)0.936 fractionPBMC from two 10x Genomics runs (HyCell report)HyCell on annotation-transfer-knn, pbmc (HyCell report)
HyCell (HyCell report)Macro F1 (Macro-F1 of a 15-nearest-neighbour classifier trained on the other batches, over the cell types of the held-out batch, averaged over held-out batches; per atlas or averaged over the seven atlases)0.833 fractionSeven held-out atlases of the HyCell reportHyCell on annotation-transfer-knn, seven-held-out-atlases (HyCell report)
HyCell (HyCell report)Macro F1 (Macro-F1 of a logistic-regression classifier trained on the other batches, over the cell types of the held-out batch, averaged over held-out batches; per atlas or averaged over the seven atlases)0.857 fractionscIB human immune atlas, integer-count part (HyCell report)HyCell on annotation-transfer-lr, immune (HyCell report)
HyCell (HyCell report)Macro F1 (Macro-F1 of a logistic-regression classifier trained on the other batches, over the cell types of the held-out batch, averaged over held-out batches; per atlas or averaged over the seven atlases)0.768 fractionscIB human lung atlas, integer-count part (HyCell report)HyCell on annotation-transfer-lr, lung (HyCell report)
HyCell (HyCell report)Macro F1 (Macro-F1 of a logistic-regression classifier trained on the other batches, over the cell types of the held-out batch, averaged over held-out batches; per atlas or averaged over the seven atlases)0.832 fractionscIB human pancreas atlas, integer-count part (HyCell report)HyCell on annotation-transfer-lr, pancreas (HyCell report)
HyCell (HyCell report)Macro F1 (Macro-F1 of a logistic-regression classifier trained on the other batches, over the cell types of the held-out batch, averaged over held-out batches; per atlas or averaged over the seven atlases)0.961 fractionPBMC from two 10x Genomics runs (HyCell report)HyCell on annotation-transfer-lr, pbmc (HyCell report)
HyCell (HyCell report)Macro F1 (Macro-F1 of a logistic-regression classifier trained on the other batches, over the cell types of the held-out batch, averaged over held-out batches; per atlas or averaged over the seven atlases)0.847 fractionSeven held-out atlases of the HyCell reportHyCell on annotation-transfer-lr, seven-held-out-atlases (HyCell report)
HyCell (HyCell report)Scib batch correction score (Mean of five scIB batch-correction metrics (batch silhouette, iLISI, kBET, graph connectivity and principal-component regression), per atlas or averaged over the seven atlases)0.54 unitlessCross-tissue immune atlas, stratified sample (HyCell report)HyCell on batch-correction, cross-tissue-immune (HyCell report)
HyCell (HyCell report)Scib batch correction score (Mean of five scIB batch-correction metrics (batch silhouette, iLISI, kBET, graph connectivity and principal-component regression), per atlas or averaged over the seven atlases)0.527 unitlessSeven held-out atlases of the HyCell reportHyCell on batch-correction, seven-held-out-atlases (HyCell report)
HyCell (HyCell report)Scib bio conservation score (Mean of seven scIB biological-conservation metrics (isolated-label score, NMI and ARI of k-means and of Leiden clusters against the cell types, cell-type silhouette and cLISI), per atlas or averaged over the seven atlases)0.644 unitlessAsian Immune Diversity Atlas, sample (HyCell report)HyCell on bio-conservation, aida (HyCell report)
HyCell (HyCell report)Scib bio conservation score (Mean of seven scIB biological-conservation metrics (isolated-label score, NMI and ARI of k-means and of Leiden clusters against the cell types, cell-type silhouette and cLISI), per atlas or averaged over the seven atlases)0.757 unitlessMultiVINE brain vasculature, healthy donors (HyCell report)HyCell on bio-conservation, brain (HyCell report)
HyCell (HyCell report)Scib bio conservation score (Mean of seven scIB biological-conservation metrics (isolated-label score, NMI and ARI of k-means and of Leiden clusters against the cell types, cell-type silhouette and cLISI), per atlas or averaged over the seven atlases)0.633 unitlessCross-tissue immune atlas, stratified sample (HyCell report)HyCell on bio-conservation, cross-tissue-immune (HyCell report)
HyCell (HyCell report)Scib bio conservation score (Mean of seven scIB biological-conservation metrics (isolated-label score, NMI and ARI of k-means and of Leiden clusters against the cell types, cell-type silhouette and cLISI), per atlas or averaged over the seven atlases)0.681 unitlessscIB human immune atlas, integer-count part (HyCell report)HyCell on bio-conservation, immune (HyCell report)
HyCell (HyCell report)Scib bio conservation score (Mean of seven scIB biological-conservation metrics (isolated-label score, NMI and ARI of k-means and of Leiden clusters against the cell types, cell-type silhouette and cLISI), per atlas or averaged over the seven atlases)0.677 unitlessscIB human lung atlas, integer-count part (HyCell report)HyCell on bio-conservation, lung (HyCell report)
HyCell (HyCell report)Scib bio conservation score (Mean of seven scIB biological-conservation metrics (isolated-label score, NMI and ARI of k-means and of Leiden clusters against the cell types, cell-type silhouette and cLISI), per atlas or averaged over the seven atlases)0.725 unitlessscIB human pancreas atlas, integer-count part (HyCell report)HyCell on bio-conservation, pancreas (HyCell report)
HyCell (HyCell report)Scib bio conservation score (Mean of seven scIB biological-conservation metrics (isolated-label score, NMI and ARI of k-means and of Leiden clusters against the cell types, cell-type silhouette and cLISI), per atlas or averaged over the seven atlases)0.780 unitlessPBMC from two 10x Genomics runs (HyCell report)HyCell on bio-conservation, pbmc (HyCell report)
HyCell (HyCell report)Scib bio conservation score (Mean of seven scIB biological-conservation metrics (isolated-label score, NMI and ARI of k-means and of Leiden clusters against the cell types, cell-type silhouette and cLISI), per atlas or averaged over the seven atlases)0.700 unitlessSeven held-out atlases of the HyCell reportHyCell on bio-conservation, seven-held-out-atlases (HyCell report)
HyCell (HyCell report)Adjusted rand index (The same two-group split made within each donor, where chemistry no longer varies; mean ARI over the 12 donors)0.66 unitlessCross-tissue immune atlas, stratified sample (HyCell report)HyCell on cd4-split-ari-within-donor, cross-tissue-immune (HyCell report)
HyCell (HyCell report)Adjusted rand index (ARI between a two-group k-means split of the embeddings of the atlas's 2,778 CD4 T cells (1,343 naive or central-memory, 1,435 effector or effector-memory) and the naive-effector annotation, over all cells)0.64 unitlessCross-tissue immune atlas, stratified sample (HyCell report)HyCell on cd4-split-ari, cross-tissue-immune (HyCell report)
HyCell (HyCell report)Normalized mutual information (NMI between the two-group k-means split of the CD4 T cells and their sequencing chemistry; lower means the split ignores chemistry, which is what the analysis seeks)0.012 unitlessCross-tissue immune atlas, stratified sample (HyCell report)HyCell on cd4-split-nmi-chemistry, cross-tissue-immune (HyCell report)
HyCell (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.665 unitlessAsian Immune Diversity Atlas, sample (HyCell report)HyCell on composite-score, aida (HyCell report)
HyCell (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.775 unitlessMultiVINE brain vasculature, healthy donors (HyCell report)HyCell on composite-score, brain (HyCell report)
HyCell (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.697 unitlessCross-tissue immune atlas, stratified sample (HyCell report)HyCell on composite-score, cross-tissue-immune (HyCell report)
HyCell (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.735 unitlessscIB human immune atlas, integer-count part (HyCell report)HyCell on composite-score, immune (HyCell report)
HyCell (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.686 unitlessscIB human lung atlas, integer-count part (HyCell report)HyCell on composite-score, lung (HyCell report)
HyCell (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.717 unitlessscIB human pancreas atlas, integer-count part (HyCell report)HyCell on composite-score, pancreas (HyCell report)
HyCell (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.871 unitlessPBMC from two 10x Genomics runs (HyCell report)HyCell on composite-score, pbmc (HyCell report)
HyCell (HyCell report)Weighted composite score (S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, the authors' weighting, which gives the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equal weight; per atlas or averaged over the seven atlases)0.735 unitlessSeven held-out atlases of the HyCell reportHyCell on composite-score, seven-held-out-atlases (HyCell report)
HyCell (HyCell report)Spearman (Spearman correlation between the decoder's fold change and the counted pseudobulk fold change from naive B to plasma cells over the 12,879 genes detected in at least 1% of either state)0.81 unitlessCross-tissue immune atlas, stratified sample (HyCell report)HyCell on decoder-fold-change-spearman, cross-tissue-immune (HyCell report)
HyCell (HyCell report)Spearman (Mean over donors of the Spearman correlation between one donor's program for naive to effector CD4 T cells and the program of all other donors, from the HyCell decoder or from pseudobulk counts)0.957 unitlessCross-tissue immune atlas, stratified sample (HyCell report)HyCell on donor-consistency-cd4, cross-tissue-immune (HyCell report)
HyCell (HyCell report)Spearman (Mean over donors of the Spearman correlation between one donor's program for classical to nonclassical monocytes and the program of all other donors, from the HyCell decoder or from pseudobulk counts)0.940 unitlessCross-tissue immune atlas, stratified sample (HyCell report)HyCell on donor-consistency-monocyte, cross-tissue-immune (HyCell report)

Claims

Statements this source supports, each with where in the source it was checked.

  • HyCell report: bio first atlases: Ranking Among self-supervised models HyCell ranks first on biological conservation on five of the seven atlases and shares first place on the brain atlas. (Section 3.2 'Cell identity'; Figure 1b)
  • HyCell report: cd4 chemistry range: Downstream For seven of the other embeddings the CD4 split follows sequencing chemistry instead, with NMI between 0.31 and 0.55. (Section 4.1)
  • HyCell report: cd4 others bound: Downstream On the naive-effector CD4 split every other embedding, Stack and PCA included, reaches ARI at most 0.031. (Section 4.1)
  • HyCell report: cell level scope: Scope The benchmark measures cell-level tasks; sample-level tasks such as predicting a donor's disease state, which Stack's use of the other cells of a sample helps, are not part of it. (Section 3.2 'Overall score'; README 'Recorded result')
  • HyCell report: cell presentations: Training About one pass over the training cells: 78.2 million cell presentations (weights.json: 78,201,849). (Section 2.5; weights.json)
  • HyCell report: census release: Training data Human primary cells with raw counts from the CZ CELLxGENE Census release of 8 November 2025. (Section 2.5; NOTICE item 1)
  • HyCell report: composite weights: Scoring S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN, which weights the scIB total (0.6 Bio + 0.4 Batch) and annotation transfer equally; atlases are weighted equally in all averages. (Section 3.1 'Metrics')
  • HyCell report: corpus cells: Training data The corpus before removal is 95.0 million cells from 555 datasets. (Section 2.5)
  • HyCell report: dataset code entries: Architecture The decoder's dataset code has 555 entries, the Census count before removal, while training used 532 datasets; the report does not say what the 23 other entries hold. (Section 2.3; Table 1 'Conditioning'; weights.json n_datasets)
  • HyCell report: developer: Developer Released by Tencent Hunyuan; the report is by the 'Hyra Team'. (Report title page; repository README)
  • HyCell report: development set: Training evaluation overlap A separate development set of eight tissues, used for all design choices, was also removed from the corpus; its tissues are not named. (Section 2.5)
  • HyCell report: donor code entries: Architecture The donor code has 12,533 entries. (Section 2.3; Table 1 'Conditioning')
  • HyCell report: donor level properties: Scope HyCell embeds every cell on its own and is trained to discount donor and dataset effects; the report says properties shared by all cells of a donor, such as sex or disease state, are better read from embeddings aggregated per donor or sample. (Section 6; README 'Embed')
  • HyCell report: embedding dimension: Outputs The embedding is a 512-dimensional linear projection of the final [CLS] state. (Section 2.2; Table 1 'Embedding')
  • HyCell report: excluded atlases: Training evaluation overlap The seven atlases excluded from pretraining are the scIB human immune, pancreas and lung atlases (integer-count parts), PBMC from 10x Genomics pbmc4k and pbmc8k, the healthy donors of the MultiVINE brain vasculature atlas, a 40,322-cell sample of the cross-tissue immune atlas of Dominguez Conde et al. and a 41,140-cell sample of the Asian Immune Diversity Atlas. Every contributing dataset and collection was removed from the corpus. (Section 3.1 'Atlases'; Table 2; Section 2.5)
  • HyCell report: expressed genes: Inputs Up to 818 expressed genes enter a cell's set. (Section 2.1; Table 1 'Input')
  • HyCell report: finetune knn batches: Fine tuning The fine-tuned annotator is ahead of kNN on 17 of the 20 held-out batches. (Section 5)
  • HyCell report: finetune lr batches: Fine tuning The fine-tuned annotator is ahead of logistic regression on 19 of the 20 held-out batches. (Section 5)
  • HyCell report: finetune time: Training Fine-tuning takes one to three minutes on one H20 for 10,000 to 31,000 reference cells. (Section 5)
  • HyCell report: gd published bound: Downstream On the gamma-delta T against tissue-resident CD8 T split the published models reach ARI at most 0.06. (Section 4.1)
  • HyCell report: gene table parameters: Architecture The gene embedding table, shared by encoder and decoder, has 9.4 million parameters. (Table 1 'Parameters')
  • HyCell report: headline level: Headline The report calls HyCell's S of 0.735 'level with' Stack's 0.736 and does not claim a win; it prints no uncertainty for either. (Abstract; Section 3.2 'Overall score')
  • HyCell report: heads: Architecture Four attention heads. (Section 2.2; Table 1 'Encoder')
  • HyCell report: knn gap: Ranking In annotation transfer HyCell is second on kNN, 0.006 behind batch-grouped Stack. (Section 3.2 'Overall score')
  • HyCell report: label trained: Comparison condition SCimilarity and scPRINT are trained with cell-type labels. (Section 3.1 'Models')

Models, benchmarks and other records

Models (5)

Methods (1)

Configurations (14)

Benchmarks (1)

Protocols (16)

Datasets (8)

Evaluations (223)

Sources (7)

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.

Evidence table

Inspect claims, sources and review details

Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.

One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.

22 evidence rows matching the loaded filters

Claims, original sources and review scope · Release 2026-11-21-adc15377fab2
Property and statementOriginal source and locationReview and provenance
Archive note
The PDF is not committed: it is refetchable at the pinned commit and extract/parse.py refuses other bytes. Its gzipped text layer is artifacts/HyCell_Technical_Report.txt.gz.
Source metadata
HyCell: a Compact Embedding Model for Single-Cell Transcriptomes (Hyra Team, 2026-11-01)

Original source ↗

No field-specific location recorded

Version: Dated 2026-11-01 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (9 pages)
Retrieved: 2026-11-06T11:42:11Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.archive_note

Source artifact SHA-256: dc2cacf87ffae8525e81fede1b01f261d3a49cece9061465c36ad1fac6680a28

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and cell count asserted.

Inspected artifact

Archive sha256
465ce7e05942e9fee63a647d9351ff535b97add2da26bcd38df0590757105c23
Source metadata
HyCell: a Compact Embedding Model for Single-Cell Transcriptomes (Hyra Team, 2026-11-01)

Original source ↗

No field-specific location recorded

Version: Dated 2026-11-01 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (9 pages)
Retrieved: 2026-11-06T11:42:11Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.archive_sha256

Source artifact SHA-256: dc2cacf87ffae8525e81fede1b01f261d3a49cece9061465c36ad1fac6680a28

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and cell count asserted.

Inspected artifact

Artifact SHA-256
dc2cacf87ffae8525e81fede1b01f261d3a49cece9061465c36ad1fac6680a28
Source metadata
HyCell: a Compact Embedding Model for Single-Cell Transcriptomes (Hyra Team, 2026-11-01)

Original source ↗

No field-specific location recorded

Version: Dated 2026-11-01 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (9 pages)
Retrieved: 2026-11-06T11:42:11Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.artifact_sha256

Source artifact SHA-256: dc2cacf87ffae8525e81fede1b01f261d3a49cece9061465c36ad1fac6680a28

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and cell count asserted.

Inspected artifact

Artifact url
https://raw.githubusercontent.com/Tencent-Hunyuan/Hyra-results/0dcc533cc9f8f76c90ecbfe944df816b6672328a/AI4Science/hycell/paper/HyCell_Technical_Report.pdf
Source metadata
HyCell: a Compact Embedding Model for Single-Cell Transcriptomes (Hyra Team, 2026-11-01)

Original source ↗

No field-specific location recorded

Version: Dated 2026-11-01 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (9 pages)
Retrieved: 2026-11-06T11:42:11Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.artifact_url

Source artifact SHA-256: dc2cacf87ffae8525e81fede1b01f261d3a49cece9061465c36ad1fac6680a28

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and cell count asserted.

Inspected artifact

Extraction method
pdftotext -layout text layer, parsed by extract/parse.py with every row label and cell count asserted.
Source metadata
HyCell: a Compact Embedding Model for Single-Cell Transcriptomes (Hyra Team, 2026-11-01)

Original source ↗

No field-specific location recorded

Version: Dated 2026-11-01 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (9 pages)
Retrieved: 2026-11-06T11:42:11Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.extraction_method

Source artifact SHA-256: dc2cacf87ffae8525e81fede1b01f261d3a49cece9061465c36ad1fac6680a28

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and cell count asserted.

Inspected artifact

Licence
Apache-2.0 (repository LICENSE)
Source metadata
HyCell: a Compact Embedding Model for Single-Cell Transcriptomes (Hyra Team, 2026-11-01)

Original source ↗

No field-specific location recorded

Version: Dated 2026-11-01 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (9 pages)
Retrieved: 2026-11-06T11:42:11Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.licence

Source artifact SHA-256: dc2cacf87ffae8525e81fede1b01f261d3a49cece9061465c36ad1fac6680a28

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and cell count asserted.

Inspected artifact

Limitations
6 values
  • Author-released report, not peer reviewed. The developer of HyCell ran every model, so no result is independent of the developer.
  • No uncertainty is printed anywhere in the report: no interval, standard deviation or test for any score, including the 0.735 against 0.736 comparison with Stack.
  • The comparators do not run under equal conditions. Stack encodes sets of 256 cells with the cells of each batch together, so it uses batch labels at inference; SCimilarity and scPRINT are trained with cell-type labels; PCA is fitted to each evaluation atlas on 2,000 highly variable genes. HyCell embeds each cell from its own counts with no batch or donor information.
  • The benchmark measures cell-level tasks only. Stack draws on the other cells of a sample at inference, which the report says helps sample-level tasks such as predicting a donor's disease state; such tasks are not in the benchmark.
  • The composite S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN is the authors' chosen weighting. Table 4 prints only S per atlas, not its four components per atlas, so the per-atlas weighting cannot be checked.
  • The seven atlases are stated to be excluded from HyCell's pretraining corpus. The report does not say whether the ten published comparators were pretrained on cells of these atlases, and several are trained on CELLxGENE data, so the comparison is not shown to be held out for them.
Source metadata
HyCell: a Compact Embedding Model for Single-Cell Transcriptomes (Hyra Team, 2026-11-01)

Original source ↗

No field-specific location recorded

Version: Dated 2026-11-01 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (9 pages)
Retrieved: 2026-11-06T11:42:11Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.limitations

Source artifact SHA-256: dc2cacf87ffae8525e81fede1b01f261d3a49cece9061465c36ad1fac6680a28

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and cell count asserted.

Inspected artifact

Media type
application/pdf
Source metadata
HyCell: a Compact Embedding Model for Single-Cell Transcriptomes (Hyra Team, 2026-11-01)

Original source ↗

No field-specific location recorded

Version: Dated 2026-11-01 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (9 pages)
Retrieved: 2026-11-06T11:42:11Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.media_type

Source artifact SHA-256: dc2cacf87ffae8525e81fede1b01f261d3a49cece9061465c36ad1fac6680a28

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and cell count asserted.

Inspected artifact

Publication status
official_project_source
Source metadata
HyCell: a Compact Embedding Model for Single-Cell Transcriptomes (Hyra Team, 2026-11-01)

Original source ↗

No field-specific location recorded

Version: Dated 2026-11-01 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (9 pages)
Retrieved: 2026-11-06T11:42:11Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.publication_status

Source artifact SHA-256: dc2cacf87ffae8525e81fede1b01f261d3a49cece9061465c36ad1fac6680a28

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and cell count asserted.

Inspected artifact

Retrieved at
2026-11-06T11:42:11Z
Source metadata
HyCell: a Compact Embedding Model for Single-Cell Transcriptomes (Hyra Team, 2026-11-01)

Original source ↗

No field-specific location recorded

Version: Dated 2026-11-01 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (9 pages)
Retrieved: 2026-11-06T11:42:11Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.retrieved_at

Source artifact SHA-256: dc2cacf87ffae8525e81fede1b01f261d3a49cece9061465c36ad1fac6680a28

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and cell count asserted.

Inspected artifact

How to read the evidence

Sources and history

Release 2026-11-21-adc15377fab2 · Record review: source checked

Technical report for HyCell, with its architecture and training, a benchmark of frozen embeddings of eleven models on seven held-out atlases, downstream analyses and fine-tuning.

Release history

No supporting source is linked yet.

Read original source

Download this release (gzip)
Technical metadata and extraction receipts

Stable ID: hycell-20261106-source-report

areas
single-cell; cells-tissues
contexts
research
archive note
The PDF is not committed: it is refetchable at the pinned commit and extract/parse.py refuses other bytes. Its gzipped text layer is artifacts/HyCell_Technical_Report.txt.gz.
archive sha256
465ce7e05942e9fee63a647d9351ff535b97add2da26bcd38df0590757105c23
artifact sha256
dc2cacf87ffae8525e81fede1b01f261d3a49cece9061465c36ad1fac6680a28
artifact url
https://raw.githubusercontent.com/Tencent-Hunyuan/Hyra-results/0dcc533cc9f8f76c90ecbfe944df816b6672328a/AI4Science/hycell/paper/HyCell_Technical_Report.pdf
extraction method
pdftotext -layout text layer, parsed by extract/parse.py with every row label and cell count asserted.
licence
Apache-2.0 (repository LICENSE)
limitations
Author-released report, not peer reviewed. The developer of HyCell ran every model, so no result is independent of the developer.; No uncertainty is printed anywhere in the report: no interval, standard deviation or test for any score, including the 0.735 against 0.736 comparison with Stack.; The comparators do not run under equal conditions. Stack encodes sets of 256 cells with the cells of each batch together, so it uses batch labels at inference; SCimilarity and scPRINT are trained with cell-type labels; PCA is fitted to each evaluation atlas on 2,000 highly variable genes. HyCell embeds each cell from its own counts with no batch or donor information.; The benchmark measures cell-level tasks only. Stack draws on the other cells of a sample at inference, which the report says helps sample-level tasks such as predicting a donor's disease state; such tasks are not in the benchmark.; The composite S = 0.3 Bio + 0.2 Batch + 0.25 LR + 0.25 kNN is the authors' chosen weighting. Table 4 prints only S per atlas, not its four components per atlas, so the per-atlas weighting cannot be checked.; The seven atlases are stated to be excluded from HyCell's pretraining corpus. The report does not say whether the ten published comparators were pretrained on cells of these atlases, and several are trained on CELLxGENE data, so the comparison is not shown to be held out for them.
media type
application/pdf
publication status
official_project_source
retrieved at
2026-11-06T11:42:11Z
source type
vendor technical report
url
https://github.com/Tencent-Hunyuan/Hyra-results/blob/0dcc533cc9f8f76c90ecbfe944df816b6672328a/AI4Science/hycell/paper/HyCell_Technical_Report.pdf
version
Dated 2026-11-01 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (9 pages)
version note
The report is dated 2026-11-01 on page 1. The PDF's CreationDate and ModDate are both 2026-11-02T00:14:59+08 (pdfTeX-1.40.29), which is 2026-11-01T16:14:59Z, the same day in UTC. The printed date is what version records. No earlier PDF exists at the pinned commit, so whether a rebuild changed any content was not checked.
Related records

Suggest a correction