Changelog
Every model release, dated, newest first. A breaking entry means a client written against the previous release needs changing; the Models page describes the models in this release and the API reference the endpoints.
Bulk-conditioned generation (PBE+BCS)
| release_id | iclr27-pbe-bcs-main-20261016 |
| Models | breast-tumour-2k-v2 · t-all-2k-v2 · sepsis-pbmc-2k-v1 |
| Breaking change | Yes — clients written against the previous release need changing. |
Three models on the ICLR 2027 submission method — PBE, the patient-bulk encoder — a generator conditioned on each sample's own bulk profile — followed by BCS, a bulk-consistency correction of the generated expression: Breast tumour (breast-tumour-2k-v2, GSE176078 reference), T-ALL (t-all-2k-v2, SCPCP000003) and Sepsis PBMC (sepsis-pbmc-2k-v1, GSE279448/GSE279452 — new). Each uses a fixed 2,000-gene panel (disease-prior genes first, then the training data's highly variable genes (HVGs); never re-selected at inference).
Input contract has changed: submit the full raw-count matrix with every gene identifier spelled exactly as the model's list spells it. The lists are not symbols-only — T-ALL 17,475 Ensembl ids (ENSG…) and 38,244 HGNC symbols; Sepsis PBMC 11,253 Ensembl ids (ENSG…) and 28,947 HGNC symbols; Breast tumour HGNC symbols — and nothing is converted on the way in. The engine normalises over the cohort's complete common-gene set (16,864 / 55,719 / 40,200 genes for Breast tumour / T-ALL / Sepsis PBMC) before selecting the panel; a missing or duplicate common gene is an error, not a zero. GET /b2sc/models/{id}/required-genes serves the exact list for download, and the new-run page checks a file against it before upload.
Cell-type axes changed: breast 13 → 11 (ER+/HER2+/TNBC epithelial merged into Cancer Epithelial), T-ALL 12 → 10; Sepsis PBMC has 10. Results are not comparable with the 20260810/20260807 releases — keep the release_id with your results.
Outputs: H5AD X and NPZ expression are continuous BCS-corrected values, not counts and not log1pCP10K; 1,000 generated cells per sample (was 200). Every run uses the release's fixed default seed, 20260904; weight sets for seeds 20260905 and 20260906 are shipped but not selectable yet — there is no seed parameter. Results are written flat under the job — platform_manifest.json, run_manifest.json, proportions.csv (six decimal places; each row sums to 1 within rounding), cell_counts.csv, generated_cells.h5ad, generated_cells.npz, verification.json — and no analysis stage is offered for any model.
What this release claims: software reproducibility (the researcher's 9/9 cases bitwise on H100; on this platform a same-GPU repeat is bitwise identical for breast and T-ALL) and allocation integrity. Not claimed: an accuracy improvement over the models this release replaces. The researcher's handoff reports reproducibility, not accuracy; the one accuracy comparison it carries is against the researcher's own earlier paper model (A2+E01+S04), re-scored on stored per-patient aggregates, not against the retired platform models — and Sepsis PBMC has no predecessor to compare against at all. In that comparison not every metric improved: reference-gene recovery rose for all three cohorts, but mean-expression error rose for breast and T-ALL and disease-effect correlation fell for T-ALL and sepsis. Research Use Only.
Predecessors retired and refused: breast-tumour-2k and breast-tumour-5k (release 20260810) and t-all-2k (release 20260807) are no longer served. A submission naming one answers 400 INVALID_MODEL listing the three current ids. Jobs and results produced by the retired releases are removed with them.