rewirebio.iobenchmarks
Source

Ladon: An RNA Foundation Model for Predicting Patient Outcomes (Zenodo PDF)

2026 · Zenodo record 24157116 (version DOI 10.5281/zenodo.24157116; concept DOI 10.5281/zenodo.24157115), file Ladon-An-RNA-Foundation-Model-for-Predicting-Patient-Outcomes.pdf, 47 pages, dated 12 November 2026 · Project source

DOI: 10.5281/zenodo.24157116

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
ABISConcordance index (pairwise ordering accuracy: fraction of pairs of test bulks with different measured lineage shares put in the measured order, ties in inferred shares counted as half; mean over five lineages; 0.5 is chance)0.76 unitlessGiroux PBMC bulk RNA-seq test set for deconvolution (35 bulks, 17 donors)ABIS, PBMC deconvolution, pairwise ordering accuracy
ABISPearson (Pearson r between inferred and measured lineage shares across 35 PBMC bulks, mean over five lineages)0.79 unitlessGiroux PBMC bulk RNA-seq test set for deconvolution (35 bulks, 17 donors)ABIS, PBMC deconvolution, pearson r
BayesPrism (single-cell PBMC references)Concordance index (pairwise ordering accuracy: fraction of pairs of test bulks with different measured lineage shares put in the measured order, ties in inferred shares counted as half; mean over five lineages; 0.5 is chance)0.74 unitlessGiroux PBMC bulk RNA-seq test set for deconvolution (35 bulks, 17 donors)BayesPrism, PBMC deconvolution, pairwise ordering accuracy
BayesPrism (single-cell PBMC references)Pearson (Pearson r between inferred and measured lineage shares across 35 PBMC bulks, mean over five lineages)0.66 unitlessGiroux PBMC bulk RNA-seq test set for deconvolution (35 bulks, 17 donors)BayesPrism, PBMC deconvolution, pearson r
BulkFormer 147M (Ladon report comparator)AUROC (AUROC for RECIST complete or partial response against other evaluable responses, immunotherapy arms, equal mean over six held-out trials)0.629 unitlessImmunotherapy arms of six held-out trials with evaluable responseBulkFormer, held-out trials immunotherapy response AUROC
BulkFormer 147M (Ladon report comparator)Hazard ratio (hazard ratio, higher-risk half against lower-risk half at the median predicted risk, overall survival, Cox stratified by treatment arm, Efron ties, full follow-up; equal-trial geometric mean over six trials)1.63 unitlessSix held-out trials with overall survivalBulkFormer, held-out trials OS hazard ratio
BulkFormer 147M (Ladon report comparator)Hazard ratio (hazard ratio, higher-risk half against lower-risk half at the median predicted risk, progression-free survival, Cox stratified by treatment arm, Efron ties, full follow-up; equal-trial geometric mean over three trials)1.52 unitlessThree held-out trials with progression-free survivalBulkFormer, held-out trials PFS hazard ratio
BulkFormer 147M (Ladon report comparator)Concordance index (Uno's C-index, survival endpoint not specified beyond 'survival', equal mean over six TCGA target cancers, predictor fitted on CPTAC)0.60 unitlessTCGA six-cancer transfer target set (2,461 patients)BulkFormer, survival transfer CPTAC to TCGA, Uno C-index
BulkFormer 147M (Ladon report comparator)Concordance index (Uno's C-index, survival endpoint not specified beyond 'survival', pan-cancer (pairs across cancers counted), equal mean of the TCGA to CPTAC and CPTAC to TCGA transfers)0.72 unitlessTCGA and CPTAC transfer target sets combined (pan-cancer panel)BulkFormer, survival transfer pan-cancer, Uno C-index
BulkFormer 147M (Ladon report comparator)Concordance index (Uno's C-index, survival endpoint not specified beyond 'survival', equal mean over six CPTAC target cancers, predictor fitted on TCGA)0.65 unitlessCPTAC six-cancer transfer target set (1,128 patients)BulkFormer, survival transfer TCGA to CPTAC, Uno C-index
BulkRNABert, published Hugging Face checkpoint (Ladon report comparator)AUROC (AUROC for RECIST complete or partial response against other evaluable responses, immunotherapy arms, equal mean over six held-out trials)0.612 unitlessImmunotherapy arms of six held-out trials with evaluable responseBulkRNABert, held-out trials immunotherapy response AUROC
BulkRNABert, published Hugging Face checkpoint (Ladon report comparator)Hazard ratio (hazard ratio, higher-risk half against lower-risk half at the median predicted risk, overall survival, Cox stratified by treatment arm, Efron ties, full follow-up; equal-trial geometric mean over six trials)1.55 unitlessSix held-out trials with overall survivalBulkRNABert, held-out trials OS hazard ratio
BulkRNABert, published Hugging Face checkpoint (Ladon report comparator)Hazard ratio (hazard ratio, higher-risk half against lower-risk half at the median predicted risk, progression-free survival, Cox stratified by treatment arm, Efron ties, full follow-up; equal-trial geometric mean over three trials)1.60 unitlessThree held-out trials with progression-free survivalBulkRNABert, held-out trials PFS hazard ratio
BulkRNABert, published Hugging Face checkpoint (Ladon report comparator)Concordance index (Uno's C-index, survival endpoint not specified beyond 'survival', equal mean over six TCGA target cancers, predictor fitted on CPTAC)0.59 unitlessTCGA six-cancer transfer target set (2,461 patients)BulkRNABert, survival transfer CPTAC to TCGA, Uno C-index
BulkRNABert, published Hugging Face checkpoint (Ladon report comparator)Concordance index (Uno's C-index, survival endpoint not specified beyond 'survival', pan-cancer (pairs across cancers counted), equal mean of the TCGA to CPTAC and CPTAC to TCGA transfers)0.68 unitlessTCGA and CPTAC transfer target sets combined (pan-cancer panel)BulkRNABert, survival transfer pan-cancer, Uno C-index
BulkRNABert, published Hugging Face checkpoint (Ladon report comparator)Concordance index (Uno's C-index, survival endpoint not specified beyond 'survival', equal mean over six CPTAC target cancers, predictor fitted on TCGA)0.63 unitlessCPTAC six-cancer transfer target set (1,128 patients)BulkRNABert, survival transfer TCGA to CPTAC, Uno C-index
CIBERSORT reimplementation (Ladon report)Pearson (Pearson r between inferred and measured B/plasma share across 35 PBMC bulks)0.39 unitlessGiroux PBMC bulk RNA-seq test set for deconvolution (35 bulks, 17 donors)CIBERSORT, PBMC deconvolution per-lineage Pearson r
CIBERSORT reimplementation (Ladon report)Pearson (Pearson r between inferred and measured CD4 T share across 35 PBMC bulks)0.78 unitlessGiroux PBMC bulk RNA-seq test set for deconvolution (35 bulks, 17 donors)CIBERSORT, PBMC deconvolution per-lineage Pearson r
CIBERSORT reimplementation (Ladon report)Pearson (Pearson r between inferred and measured CD8 T share across 35 PBMC bulks)0.77 unitlessGiroux PBMC bulk RNA-seq test set for deconvolution (35 bulks, 17 donors)CIBERSORT, PBMC deconvolution per-lineage Pearson r
CIBERSORT reimplementation (Ladon report)Pearson (Pearson r between inferred and measured Monocyte share across 35 PBMC bulks)0.76 unitlessGiroux PBMC bulk RNA-seq test set for deconvolution (35 bulks, 17 donors)CIBERSORT, PBMC deconvolution per-lineage Pearson r
CIBERSORT reimplementation (Ladon report)Pearson (Pearson r between inferred and measured NK share across 35 PBMC bulks)0.27 unitlessGiroux PBMC bulk RNA-seq test set for deconvolution (35 bulks, 17 donors)CIBERSORT, PBMC deconvolution per-lineage Pearson r
CIBERSORT reimplementation (Ladon report)Concordance index (pairwise ordering accuracy: fraction of pairs of test bulks with different measured lineage shares put in the measured order, ties in inferred shares counted as half; mean over five lineages; 0.5 is chance)0.70 unitlessGiroux PBMC bulk RNA-seq test set for deconvolution (35 bulks, 17 donors)CIBERSORT, PBMC deconvolution, pairwise ordering accuracy
CIBERSORT reimplementation (Ladon report)Pearson (Pearson r between inferred and measured lineage shares across 35 PBMC bulks, mean over five lineages)0.60 unitlessGiroux PBMC bulk RNA-seq test set for deconvolution (35 bulks, 17 donors)CIBERSORT, PBMC deconvolution, pearson r
Clinical covariates encoded by BMRetriever-7B (Ladon report baseline)AUROC (AUROC for RECIST complete or partial response against other evaluable responses, immunotherapy arms, equal mean over six held-out trials)0.515 unitlessImmunotherapy arms of six held-out trials with evaluable responseClinical covariates (BMRetriever), held-out trials immunotherapy response AUROC
Clinical covariates encoded by BMRetriever-7B (Ladon report baseline)Hazard ratio (hazard ratio, higher-risk half against lower-risk half at the median predicted risk, overall survival, Cox stratified by treatment arm, Efron ties, full follow-up; equal-trial geometric mean over six trials)1.61 unitlessSix held-out trials with overall survivalClinical covariates (BMRetriever), held-out trials OS hazard ratio
Clinical covariates encoded by BMRetriever-7B (Ladon report baseline)Hazard ratio (hazard ratio, higher-risk half against lower-risk half at the median predicted risk, progression-free survival, Cox stratified by treatment arm, Efron ties, full follow-up; equal-trial geometric mean over three trials)1.66 unitlessThree held-out trials with progression-free survivalClinical covariates (BMRetriever), held-out trials PFS hazard ratio
Clinical covariates encoded by BMRetriever-7B (Ladon report baseline)Concordance index (Uno's C-index, survival endpoint not specified beyond 'survival', equal mean over six TCGA target cancers, predictor fitted on CPTAC)0.66 unitlessTCGA six-cancer transfer target set (2,461 patients)Clinical covariates (BMRetriever), survival transfer CPTAC to TCGA, Uno C-index
Clinical covariates encoded by BMRetriever-7B (Ladon report baseline)Concordance index (Uno's C-index, survival endpoint not specified beyond 'survival', pan-cancer (pairs across cancers counted), equal mean of the TCGA to CPTAC and CPTAC to TCGA transfers)0.71 unitlessTCGA and CPTAC transfer target sets combined (pan-cancer panel)Clinical covariates (BMRetriever), survival transfer pan-cancer, Uno C-index
Clinical covariates encoded by BMRetriever-7B (Ladon report baseline)Concordance index (Uno's C-index, survival endpoint not specified beyond 'survival', equal mean over six CPTAC target cancers, predictor fitted on TCGA)0.68 unitlessCPTAC six-cancer transfer target set (1,128 patients)Clinical covariates (BMRetriever), survival transfer TCGA to CPTAC, Uno C-index
Cohort mean isoform splitMAE (mean absolute error of INSR isoform percentage, 166 TCGA stomach tumours)12.2 percentTCGA stomach adenocarcinomas with INSR at least 10 TPM (166)cohort mean, INSR isoform split in 166 stomach tumours, MAE
Cohort-trained isoform ridge baseline (Ladon report)MAE (mean absolute error of INSR isoform percentage, 166 TCGA stomach tumours)6.6 percentTCGA stomach adenocarcinomas with INSR at least 10 TPM (166)cohort-trained baseline, INSR isoform split in 166 stomach tumours, MAE
Cohort-trained isoform ridge baseline (Ladon report)Pearson (Pearson r between predicted and measured INSR isoform percentages, 166 TCGA stomach tumours with INSR at least 10 TPM)0.84 unitlessTCGA stomach adenocarcinomas with INSR at least 10 TPM (166)cohort-trained baseline, INSR isoform split in 166 stomach tumours, Pearson r
Cohort-trained isoform ridge baseline (Ladon report)Pearson (median over 14,526 gene and cancer-type combinations (1,572 tumours) of Pearson r between predicted and measured isoform percentages, each isoform centred on its mean across tumours)0.69 unitlessTCGA tumours of eight held-out cancer types for isoform recovery (1,572)Cohort-trained baseline, isoform variation in eight held-out cancer types, median Pearson r
Differential gene expression signature (Ladon report)Concordance index (Uno's C-index, overall survival, mean over six trials weighted by evaluated patients; training pool TCGA plus non-trial cohorts plus other trials (largest pool, about 18k samples))0.595 unitlessSix held-out trials with overall survivalDGE, held-out trials OS Uno C-index, largest training pool
Differential gene expression signature (Ladon report)AUROC (AUROC for RECIST complete or partial response against other evaluable responses, immunotherapy arms, equal mean over six held-out trials)0.633 unitlessImmunotherapy arms of six held-out trials with evaluable responseDifferential gene expression, held-out trials immunotherapy response AUROC
Differential gene expression signature (Ladon report)Hazard ratio (hazard ratio, higher-risk half against lower-risk half at the median predicted risk, overall survival, Cox stratified by treatment arm, Efron ties, full follow-up; equal-trial geometric mean over six trials)1.52 unitlessSix held-out trials with overall survivalDifferential gene expression, held-out trials OS hazard ratio
Differential gene expression signature (Ladon report)Hazard ratio (hazard ratio, higher-risk half against lower-risk half at the median predicted risk, progression-free survival, Cox stratified by treatment arm, Efron ties, full follow-up; equal-trial geometric mean over three trials)1.74 unitlessThree held-out trials with progression-free survivalDifferential gene expression, held-out trials PFS hazard ratio
Differential gene expression signature (Ladon report)Concordance index (Uno's C-index, survival endpoint not specified beyond 'survival', equal mean over six TCGA target cancers, predictor fitted on CPTAC)0.60 unitlessTCGA six-cancer transfer target set (2,461 patients)Differential gene expression, survival transfer CPTAC to TCGA, Uno C-index
Differential gene expression signature (Ladon report)Concordance index (Uno's C-index, survival endpoint not specified beyond 'survival', pan-cancer (pairs across cancers counted), equal mean of the TCGA to CPTAC and CPTAC to TCGA transfers)0.59 unitlessTCGA and CPTAC transfer target sets combined (pan-cancer panel)Differential gene expression, survival transfer pan-cancer, Uno C-index
Differential gene expression signature (Ladon report)Concordance index (Uno's C-index, survival endpoint not specified beyond 'survival', equal mean over six CPTAC target cancers, predictor fitted on TCGA)0.63 unitlessCPTAC six-cancer transfer target set (1,128 patients)Differential gene expression, survival transfer TCGA to CPTAC, Uno C-index
DWLS reimplementation (Ladon report)Concordance index (pairwise ordering accuracy: fraction of pairs of test bulks with different measured lineage shares put in the measured order, ties in inferred shares counted as half; mean over five lineages; 0.5 is chance)0.65 unitlessGiroux PBMC bulk RNA-seq test set for deconvolution (35 bulks, 17 donors)DWLS, PBMC deconvolution, pairwise ordering accuracy
DWLS reimplementation (Ladon report)Pearson (Pearson r between inferred and measured lineage shares across 35 PBMC bulks, mean over five lineages)0.57 unitlessGiroux PBMC bulk RNA-seq test set for deconvolution (35 bulks, 17 donors)DWLS, PBMC deconvolution, pearson r
EPIC (BRef reference)Concordance index (pairwise ordering accuracy: fraction of pairs of test bulks with different measured lineage shares put in the measured order, ties in inferred shares counted as half; mean over five lineages; 0.5 is chance)0.76 unitlessGiroux PBMC bulk RNA-seq test set for deconvolution (35 bulks, 17 donors)EPIC, PBMC deconvolution, pairwise ordering accuracy
EPIC (BRef reference)Pearson (Pearson r between inferred and measured lineage shares across 35 PBMC bulks, mean over five lineages)0.80 unitlessGiroux PBMC bulk RNA-seq test set for deconvolution (35 bulks, 17 donors)EPIC, PBMC deconvolution, pearson r
Gene TPM features with academic-only context (Ladon report)AUROC (AUROC for RECIST complete or partial response against other evaluable responses, immunotherapy arms, equal mean over six held-out trials)0.613 unitlessImmunotherapy arms of six held-out trials with evaluable responseGene TPM, held-out trials immunotherapy response AUROC
Hallmark GSEA pathway signature (Ladon report)Concordance index (Uno's C-index, overall survival, mean over six trials weighted by evaluated patients; training pool TCGA plus non-trial cohorts plus other trials (largest pool, about 18k samples))0.578 unitlessSix held-out trials with overall survivalHallmark GSEA, held-out trials OS Uno C-index, largest training pool
Hallmark GSEA pathway signature (Ladon report)Concordance index (Uno's C-index, survival endpoint not specified beyond 'survival', equal mean over six TCGA target cancers, predictor fitted on CPTAC)0.57 unitlessTCGA six-cancer transfer target set (2,461 patients)Hallmark GSEA, survival transfer CPTAC to TCGA, Uno C-index
Hallmark GSEA pathway signature (Ladon report)Concordance index (Uno's C-index, survival endpoint not specified beyond 'survival', pan-cancer (pairs across cancers counted), equal mean of the TCGA to CPTAC and CPTAC to TCGA transfers)0.56 unitlessTCGA and CPTAC transfer target sets combined (pan-cancer panel)Hallmark GSEA, survival transfer pan-cancer, Uno C-index
Hallmark GSEA pathway signature (Ladon report)Concordance index (Uno's C-index, survival endpoint not specified beyond 'survival', equal mean over six CPTAC target cancers, predictor fitted on TCGA)0.59 unitlessCPTAC six-cancer transfer target set (1,128 patients)Hallmark GSEA, survival transfer TCGA to CPTAC, Uno C-index
HER2 expression (ERBB2 mRNA)AUROC (AUC for pathological complete response, HER2+ tumours, model fitted on TransNEO and applied unchanged)0.53 unitlessDFCI inflammatory breast cancer HER2+ patients (22, 9 pCR)HER2 expression, DFCI IBC HER2+, pCR AUC

Claims

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

  • Ladon report: access: Access No model weights, code repository or model licence is published with the report. As checked on 2026-11-15, the GitHub organisation blankbio lists 0 public repositories and the Hugging Face organisation BlankBio lists 0 models. The report's conclusion (page 43) and the announcement invite readers to contact Blank Bio to use the model. Access terms are unresolved; contact the provider. (Page 43, Conclusion, last sentence; Zenodo record Rights field)
  • Ladon report: adaptation description conflict: Adaptation description conflict The announcement says 'We fine-tune Ladon on academic immunotherapy cohorts' and that Ladon 'significantly outperforms all baselines'. The report describes supervised mid-training with academic response labels followed by a frozen encoder, PCA and TabICL, and its tests are one-sided and unadjusted (Introducing Ladon, section 'Ladon Predicts Response in Held-Out Clinical Trials'; report pages 11, 15 and 16)
  • Ladon report: all gene direction accuracy: All gene direction accuracy At the cutoff including all 9,207 expressed multi-isoform genes, Ladon's direction accuracy is 56% against 53% for direct quantification; the text does not say which of the two contrasts these values belong to (Page 27, sentence 'Ladon nevertheless stayed ahead at every cutoff down to all 9,207 expressed multi-isoform genes (56% compared with 53%)')
  • Ladon report: analysis design: Analysis design A benefit head on frozen Ladon embeddings estimates how the treatment-minus-control difference in 12-month PFS varies with molecular features (R-learning, Nie and Wager 2021), after a Ladon-derived score regresses out prognosis in both arms. Source trials: IMmotion150, IMmotion151, IMbrave150, OAK and IMpower133; five-fold cross-fitting with censoring-weighted logistic regression; ridge penalties 10^-4 to 10^3 chosen by leave-one-source-trial-out R-loss. Applied to POPLAR (atezolizumab against docetaxel) (Page 18 and page 22, Methods 'Benefit learning and target evaluation')
  • Ladon report: architecture: Architecture 3-billion-parameter transformer encoder-decoder (Page 2, 'Introducing Ladon', first sentence)
  • Ladon report: checkpoint identity: Checkpoint identity No checkpoint, model size per experiment or version is named. Figures 20 and 21 compare several model sizes and training steps, but are images with no values in the text layer (Pages 41 to 43, Figures 20 and 21 and Methods 'Reconstruction loss' and 'Clinical benchmark')
  • Ladon report: cohort description: Cohort description The announcement calls ARTEMIS and PBCP 'two additional trials'; the report treats the four external sets as cohorts from the Sammut et al. (2022) release and GEO GSE207248, with 35, 13, 18 and 22 patients (Introducing Ladon, section 'Ladon Recovers Cell Composition Signal'; report Figure 18 and page 40)
  • Ladon report: comparator gaps printed: Comparator gaps printed The comparator panel of Figure 8 prints 12-month high-low gaps of -2.0 percentage points (atezolizumab) and -2.6 percentage points (docetaxel) for its default selection. The panel header lists PD-L1 mRNA, AP/T, Plasma-cell and TLS3 as toggles and the PDF does not mark which is selected; values for the other three comparators are not printed (Figure 8 (page 19), comparator panel)
  • Ladon report: development setting: Development setting IMPRES was developed on melanoma immunotherapy cohorts; these trials are outside that development setting (Web Figure 6 page, DATA master.panel_f.rows, row 'IMPRES', field note)
  • Ladon report: development setting: Development setting TIDE was developed on melanoma anti-PD-1 / CTLA-4 cohorts; these trials are outside that development setting. The score is sign-reversed so higher values indicate expected response (Web Figure 6 page, DATA master.panel_f.rows, row 'TIDE', field note)
  • Ladon report: figure symbol: Figure symbol In Figure 2 the rows 'Differential gene expression' and 'Hallmark GSEA' carry a double dagger (‡) that the PDF does not explain; it applies to all three panels (Figure 2 (page 6), row labels)
  • Ladon report: five gene comparison: Five gene comparison In the five-gene analysis (CD44, INSR, KRAS, MKNK2, TPM1 in the five cancer types where at least four are eligible) Ladon exceeded the cohort-trained baseline in all 21 gene and cancer-type combinations; Figure 12b prints no values (Page 24 prose; Figure 12b (page 25))
  • Ladon report: held out protection: Held out protection Pretraining excluded the TCGA cohorts and included the ENCODE4 and blood bulk reference libraries; Ladon received gene-level expression only (Page 28, Methods 'Model input')
  • Ladon report: held out protection: Held out protection A random subset of single-cell studies contributed no synthetic bulks to post-training; these held-out studies included the single-cell study paired with the PBMC test bulks (Page 38, Methods 'Ladon post-training')
  • Ladon report: held out protection: Held out protection TCGA and CPTAC outcome labels are excluded from Ladon's encoder supervision and validation monitoring; the mid-trained encoder is trained on survival outcomes from other cohorts (Page 11, Methods 'Transfer between TCGA and CPTAC')
  • Ladon report: imbrave150 handling: Imbrave150 handling IMbrave150 has no patient-specific clinical covariates; its clinical-only hazard ratio is not estimable, the clinical-only summaries cover five OS and two PFS trials, and the combined predictor uses Ladon alone for IMbrave150 (Pages 13 and 14, Methods 'Survival in held-out trials')
  • Ladon report: in sample trial: In sample trial The T-effector signature was derived on POPLAR, so its POPLAR result is in-sample (Web Figure 6 page, DATA master.panel_f.rows, row 'T-effector / IFN-γ', field note)
  • Ladon report: ladon higher share: Ladon higher share Ladon's Pearson r is higher than the cohort-trained baseline's in 95.1% of the 14,526 gene and cancer-type combinations, and higher than the linear model on public data and the nearest public samples in 99.7% each (Figure 12a (page 25), 'Ladon higher in' labels)
  • Ladon report: marker association counts: Marker association counts Prose: of 304 marker-PFS associations significant before adjustment for Ladon prognosis, 24 remained significant (Benjamini-Hochberg within trial, arm and condition over 550 markers). Figure 7 prints '106 significant' and '0 retained' for its default trial selection; its toggle lists POPLAR, OAK and IMmotion151, while the Methods analyse four trials including IMmotion150; its caption says '550 total' although only markers significant before adjustment are displayed (Page 17 prose; Figure 7 (page 17) and caption; page 20, Methods 'Association models and endpoints')
  • Ladon report: median survival split: Median survival split Splitting OAK at the median predicted risk of Ladon + clinical covariates gives median overall survival of 7.9 months in the higher-risk half (n = 349) and 15.5 months in the lower-risk half (n = 350); dashed predicted curves are calibrated using OAK outcomes (Page 7 prose; Figure 4 (page 8) and caption)
  • Ladon report: poplar 12 month pfs gap: Poplar 12 month pfs gap Ladon score in POPLAR: 12-month PFS high-low gap 22.1 percentage points in the atezolizumab arm and 1.3 percentage points in the docetaxel arm; 28% against 6% progression-free at 12 months on atezolizumab, 14% against 13% on docetaxel. Patients at risk at month 0: atezolizumab high 43, low 52; docetaxel high 53, low 44 (Figure 8 (page 18), panel 'Ladon + TCGA' headers and at-risk tables; page 18 prose)
  • Ladon report: pretraining objective: Pretraining objective Pretrained with masked token reconstruction on isoform-level sequences of millions of bulk RNA-seq samples; self-supervised, so unlabelled samples are used (Page 2, 'Introducing Ladon')
  • Ladon report: printed results: Printed results No ClinBench score is printed: Figures 20 and 21 are images, the text layer carries no values, and the task list is not published (Pages 41 and 42, Figures 20 and 21)
  • Ladon report: reference data parity: Reference data parity Ladon's in-context examples are 1,419 patients from 33 academic cohorts plus patients from the other trials, and its encoder was mid-trained with response labels from academic cohorts; the gene-expression (gene TPM) baseline uses academic cohorts alone as context. BulkRNABert, BulkFormer and BMRetriever 'use the same PCA and TabICL procedure as Ladon'; whether their context includes the other trials is not stated explicitly, nor which context the DGE signature is fitted in (Pages 15 and 16, Methods 'Immunotherapy response')
  • Ladon report: response stratification: Response stratification Figure 9 shows objective response rates in high- and low-score groups of the POPLAR atezolizumab arm (87 of 95 patients with evaluable response) but prints no rate values (Figure 9 (page 20) and caption)

Models, benchmarks and other records

Models (4)

Methods (22)

Configurations (12)

Benchmarks (1)

Tasks (6)

Protocols (19)

Datasets (7)

Dataset subsets (16)

Evaluations (93)

Sources (3)

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.

31 evidence rows matching the loaded filters

Claims, original sources and review scope · Release 2026-11-21-c8bea25e5abb
Property and statementOriginal source and locationReview and provenance
Access note
Zenodo access_right is open, which covers downloading this PDF only. No model weights, code repository or model licence is published with the report. As checked on 2026-11-15, the GitHub organisation blankbio lists 0 public repositories and the Hugging Face organisation BlankBio lists 0 models. The report's conclusion (page 43) and the announcement invite readers to contact Blank Bio to use the model. Access terms are unresolved; contact the provider.
Source metadata
Ladon: An RNA Foundation Model for Predicting Patient Outcomes (Zenodo PDF)

Original source ↗

No field-specific location recorded

Version: Zenodo record 24157116 (version DOI 10.5281/zenodo.24157116; concept DOI 10.5281/zenodo.24157115), file Ladon-An-RNA-Foundation-Model-for-Predicting-Patient-Outcomes.pdf, 47 pages, dated 12 November 2026
Retrieved: 2026-11-15T20:29:10Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.access_note

Source artifact SHA-256: 64edf7c3c956c50a65aa01a0dcaa13f0cfef28ab4373ecfd50c431e67143893c

Hash scope: Hash scope not separately documented; inspect source record

Artifact SHA-256
64edf7c3c956c50a65aa01a0dcaa13f0cfef28ab4373ecfd50c431e67143893c
Source metadata
Ladon: An RNA Foundation Model for Predicting Patient Outcomes (Zenodo PDF)

Original source ↗

No field-specific location recorded

Version: Zenodo record 24157116 (version DOI 10.5281/zenodo.24157116; concept DOI 10.5281/zenodo.24157115), file Ladon-An-RNA-Foundation-Model-for-Predicting-Patient-Outcomes.pdf, 47 pages, dated 12 November 2026
Retrieved: 2026-11-15T20:29:10Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.artifact_sha256

Source artifact SHA-256: 64edf7c3c956c50a65aa01a0dcaa13f0cfef28ab4373ecfd50c431e67143893c

Hash scope: Hash scope not separately documented; inspect source record

Code licence
None stated. The report links no code, weights or model licence.
Source metadata
Ladon: An RNA Foundation Model for Predicting Patient Outcomes (Zenodo PDF)

Original source ↗

No field-specific location recorded

Version: Zenodo record 24157116 (version DOI 10.5281/zenodo.24157116; concept DOI 10.5281/zenodo.24157115), file Ladon-An-RNA-Foundation-Model-for-Predicting-Patient-Outcomes.pdf, 47 pages, dated 12 November 2026
Retrieved: 2026-11-15T20:29:10Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.code_licence

Source artifact SHA-256: 64edf7c3c956c50a65aa01a0dcaa13f0cfef28ab4373ecfd50c431e67143893c

Hash scope: Hash scope not separately documented; inspect source record

Doi
10.5281/zenodo.24157116
Source metadata
Ladon: An RNA Foundation Model for Predicting Patient Outcomes (Zenodo PDF)

Original source ↗

No field-specific location recorded

Version: Zenodo record 24157116 (version DOI 10.5281/zenodo.24157116; concept DOI 10.5281/zenodo.24157115), file Ladon-An-RNA-Foundation-Model-for-Predicting-Patient-Outcomes.pdf, 47 pages, dated 12 November 2026
Retrieved: 2026-11-15T20:29:10Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.doi

Source artifact SHA-256: 64edf7c3c956c50a65aa01a0dcaa13f0cfef28ab4373ecfd50c431e67143893c

Hash scope: Hash scope not separately documented; inspect source record

Licence
Copyright (C) 2026 Blank Bio. All rights reserved. (Zenodo record, Rights field)
Source metadata
Ladon: An RNA Foundation Model for Predicting Patient Outcomes (Zenodo PDF)

Original source ↗

No field-specific location recorded

Version: Zenodo record 24157116 (version DOI 10.5281/zenodo.24157116; concept DOI 10.5281/zenodo.24157115), file Ladon-An-RNA-Foundation-Model-for-Predicting-Patient-Outcomes.pdf, 47 pages, dated 12 November 2026
Retrieved: 2026-11-15T20:29:10Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.licence

Source artifact SHA-256: 64edf7c3c956c50a65aa01a0dcaa13f0cfef28ab4373ecfd50c431e67143893c

Hash scope: Hash scope not separately documented; inspect source record

Limitations
3 values
  • Company technical report deposited on Zenodo; not peer reviewed.
  • The PDF is a print of an interactive web page. Per-trial values, p-values and confidence intervals that the web version shows on hover or in collapsed data tables are not printed in the PDF.
  • All rights reserved: the PDF and its text layer are not archived in this repository.
Source metadata
Ladon: An RNA Foundation Model for Predicting Patient Outcomes (Zenodo PDF)

Original source ↗

No field-specific location recorded

Version: Zenodo record 24157116 (version DOI 10.5281/zenodo.24157116; concept DOI 10.5281/zenodo.24157115), file Ladon-An-RNA-Foundation-Model-for-Predicting-Patient-Outcomes.pdf, 47 pages, dated 12 November 2026
Retrieved: 2026-11-15T20:29:10Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.limitations

Source artifact SHA-256: 64edf7c3c956c50a65aa01a0dcaa13f0cfef28ab4373ecfd50c431e67143893c

Hash scope: Hash scope not separately documented; inspect source record

Media type
application/pdf
Source metadata
Ladon: An RNA Foundation Model for Predicting Patient Outcomes (Zenodo PDF)

Original source ↗

No field-specific location recorded

Version: Zenodo record 24157116 (version DOI 10.5281/zenodo.24157116; concept DOI 10.5281/zenodo.24157115), file Ladon-An-RNA-Foundation-Model-for-Predicting-Patient-Outcomes.pdf, 47 pages, dated 12 November 2026
Retrieved: 2026-11-15T20:29:10Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.media_type

Source artifact SHA-256: 64edf7c3c956c50a65aa01a0dcaa13f0cfef28ab4373ecfd50c431e67143893c

Hash scope: Hash scope not separately documented; inspect source record

Publication status
official_project_source
Source metadata
Ladon: An RNA Foundation Model for Predicting Patient Outcomes (Zenodo PDF)

Original source ↗

No field-specific location recorded

Version: Zenodo record 24157116 (version DOI 10.5281/zenodo.24157116; concept DOI 10.5281/zenodo.24157115), file Ladon-An-RNA-Foundation-Model-for-Predicting-Patient-Outcomes.pdf, 47 pages, dated 12 November 2026
Retrieved: 2026-11-15T20:29:10Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.publication_status

Source artifact SHA-256: 64edf7c3c956c50a65aa01a0dcaa13f0cfef28ab4373ecfd50c431e67143893c

Hash scope: Hash scope not separately documented; inspect source record

Retrieval url
https://zenodo.org/api/records/24157116/files/Ladon-An-RNA-Foundation-Model-for-Predicting-Patient-Outcomes.pdf/content
Source metadata
Ladon: An RNA Foundation Model for Predicting Patient Outcomes (Zenodo PDF)

Original source ↗

No field-specific location recorded

Version: Zenodo record 24157116 (version DOI 10.5281/zenodo.24157116; concept DOI 10.5281/zenodo.24157115), file Ladon-An-RNA-Foundation-Model-for-Predicting-Patient-Outcomes.pdf, 47 pages, dated 12 November 2026
Retrieved: 2026-11-15T20:29:10Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.retrieval_url

Source artifact SHA-256: 64edf7c3c956c50a65aa01a0dcaa13f0cfef28ab4373ecfd50c431e67143893c

Hash scope: Hash scope not separately documented; inspect source record

Retrieved at
2026-11-15T20:29:10Z
Source metadata
Ladon: An RNA Foundation Model for Predicting Patient Outcomes (Zenodo PDF)

Original source ↗

No field-specific location recorded

Version: Zenodo record 24157116 (version DOI 10.5281/zenodo.24157116; concept DOI 10.5281/zenodo.24157115), file Ladon-An-RNA-Foundation-Model-for-Predicting-Patient-Outcomes.pdf, 47 pages, dated 12 November 2026
Retrieved: 2026-11-15T20:29:10Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.retrieved_at

Source artifact SHA-256: 64edf7c3c956c50a65aa01a0dcaa13f0cfef28ab4373ecfd50c431e67143893c

Hash scope: Hash scope not separately documented; inspect source record

How to read the evidence

Sources and history

Release 2026-11-21-c8bea25e5abb · Record review: source checked

The Blank Bio technical report introducing Ladon, a bulk RNA-seq foundation model, and its evaluations on survival, immunotherapy response, treatment benefit, isoform recovery and cell composition.

1 source record and release history

Read original source

Download this release (gzip)
Technical metadata and extraction receipts

Stable ID: ladon-20261115-source-zenodo-pdf

areas
rna-transcriptomes
contexts
clinical_research
access note
Zenodo access_right is open, which covers downloading this PDF only. No model weights, code repository or model licence is published with the report. As checked on 2026-11-15, the GitHub organisation blankbio lists 0 public repositories and the Hugging Face organisation BlankBio lists 0 models. The report's conclusion (page 43) and the announcement invite readers to contact Blank Bio to use the model. Access terms are unresolved; contact the provider.
artifact sha256
64edf7c3c956c50a65aa01a0dcaa13f0cfef28ab4373ecfd50c431e67143893c
code licence
None stated. The report links no code, weights or model licence.
doi
10.5281/zenodo.24157116
licence
Copyright (C) 2026 Blank Bio. All rights reserved. (Zenodo record, Rights field)
limitations
Company technical report deposited on Zenodo; not peer reviewed.; The PDF is a print of an interactive web page. Per-trial values, p-values and confidence intervals that the web version shows on hover or in collapsed data tables are not printed in the PDF.; All rights reserved: the PDF and its text layer are not archived in this repository.
media type
application/pdf
publication status
official_project_source
retrieval url
https://zenodo.org/api/records/24157116/files/Ladon-An-RNA-Foundation-Model-for-Predicting-Patient-Outcomes.pdf/content
retrieved at
2026-11-15T20:29:10Z
review
artifact sha256: 64edf7c3c956c50a65aa01a0dcaa13f0cfef28ab4373ecfd50c431e67143893c; method: source-hash-verification; automated-source-review; method note: Independent review pass. Each pinned artifact was re-retrieved at review time and its SHA-256 recomputed: the Zenodo PDF matched (and its MD5 matched the Zenodo file checksum), and the three HTML artifacts matched byte for byte, although all three are mutable and the hash describes one retrieval. The PDF text layer regenerated with pdftotext 26.08.0 -layout reproduced the SHA-256 the source record's version_note states, so the text every value was read from is provably derived from the pinned bytes.; note: Artifact re-retrieved at review time; hash matched the source record.; retrieval url: https://zenodo.org/api/records/24157116/files/Ladon-An-RNA-Foundation-Model-for-Predicting-Patient-Outcomes.pdf/content; reviewed at: 2026-11-15T23:30:00Z; reviewer: claude; reviewer note: Claude review agent, independent of the extractor; no human review claimed. Transcription was checked against the pinned sources, which is not reproduction: nothing was rerun and no value was recomputed.
source locator
Whole document; title page (page 1) for authors, date and DOI
url
https://doi.org/10.5281/zenodo.24157116
version
Zenodo record 24157116 (version DOI 10.5281/zenodo.24157116; concept DOI 10.5281/zenodo.24157115), file Ladon-An-RNA-Foundation-Model-for-Predicting-Patient-Outcomes.pdf, 47 pages, dated 12 November 2026
version note
Downloaded twice on 2026-11-15 (about 20:22Z and 20:29Z) with identical bytes; the MD5 matches the Zenodo file checksum abc8168cf6059ee806b34ced791434fb. Values were read from a pdftotext 26.08.0 -layout text layer, SHA-256 2f5a65a1e6b9830dd2aa1e0155865462c34a58451a230657f697d110b637bd6d.
year
2026
Related records

Suggest a correction