Configuration: GEARS (Ahlmann-Eltze et al. 2025) Protocol: Adamson: expression prediction for unseen single perturbations (Ahlmann-Eltze et al. 2025 Fig. 2a and Extended Data Fig. 8a) Dataset: Adamson et al. 2016 Perturb-seq, K562 CRISPRi single perturbations (GEARS distribution, Ahlmann-Eltze et al. 2025) 0.726 Pearson delta (1,000 most highly expressed control genes; mean over held-out perturbations and splits)Computed by Rewire from the authors' source data unitless · higher
Uncertainty: Not reported by the source: No spread of the mean is stated; Rewire did not compute one
Coverage: 48/48 perturbation-split rows scored
Independent external evaluation · Source checked Evidence concern: excluded from comparisons Panel A repeats Source Data Fig. 2 Panel A row for row (48, 262 and 413 rows per method over two splits). Methods restricts the analysis to 73, 398 and 629 perturbations predicted by all models; the sheet does not have these counts.Source Data Extended Data Fig. 8, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Extended Data Fig. 8, sheet "Panel A"
The Fig. 2 legend gives "134, 210 and 24 unseen single perturbations across two test-training splits". In Panel A these are the split 1 counts only (Replogle K562 134, Replogle RPE1 210, Adamson 24 rows per method); split 2 has 128, 203 and 24, so the two-split totals are 262, 413 and 48. The Fig. 2a means recorded here use all Panel A rows; whether the plotted red lines use both splits is not stated beyond the sheet itself. Reviewer check 2026-10-11: using split 1 rows only would change the per-model means by up to 0.61 (L2 distance) and 0.06 (Pearson delta), and would change the model with the lowest mean L2 distance on Adamson and Replogle RPE1, so the choice is material.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Fig. 2 legend versus Source Data Fig. 2, sheet "Panel A"
Methods restricts the single-perturbation analysis to perturbations predicted by all models: 73 for Adamson, 398 for Replogle K562 and 629 for Replogle RPE1. No sheet of Source Data Fig. 2 has these counts (Panel A: 48, 262 and 413 per method over two splits; Panel C: 20, 214 and 224 test or val rows per method).Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Fig. 2, sheets "Panel A" and "Panel C"
The test and val rows in Panel C (20, 214 and 224 per method for Adamson, Replogle K562 and Replogle RPE1) match neither the Fig. 2 legend counts nor the Methods counts. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2, sheet "Panel C" versus the Fig. 2 legend and Methods
The main text says CPA was not included in the single-perturbation benchmark, but method "cpa" has rows in Panel C. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Main text paragraph 16 versus Source Data Fig. 2, sheet "Panel C"
Methods, coverage and source GEARS on Adamson (Ahlmann-Eltze et al. 2025)
perturbation-ahlmann-20261011-protocol-adamson-single
Aggregation: Not reported
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines ; Source Data Extended Data Fig. 8, Ahlmann-Eltze et al. 2025 · Source Data Extended Data Fig. 8 (MOESM10), sheet "Panel A", rows with dataset_name "adamson" and method "gears", column "r2_delta"; plotted as the red line in Extended Data Fig. 8a Configuration: Geneformer (Ahlmann-Eltze et al. 2025) Protocol: Adamson: expression prediction for unseen single perturbations (Ahlmann-Eltze et al. 2025 Fig. 2a and Extended Data Fig. 8a) Dataset: Adamson et al. 2016 Perturb-seq, K562 CRISPRi single perturbations (GEARS distribution, Ahlmann-Eltze et al. 2025) 0.726 Pearson delta (1,000 most highly expressed control genes; mean over held-out perturbations and splits)Computed by Rewire from the authors' source data unitless · higher
Uncertainty: Not reported by the source: No spread of the mean is stated; Rewire did not compute one
Coverage: 48/48 perturbation-split rows scored
Independent external evaluation · Source checked Evidence concern: excluded from comparisons Panel A repeats Source Data Fig. 2 Panel A row for row (48, 262 and 413 rows per method over two splits). Methods restricts the analysis to 73, 398 and 629 perturbations predicted by all models; the sheet does not have these counts.Source Data Extended Data Fig. 8, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Extended Data Fig. 8, sheet "Panel A"
The Fig. 2 legend gives "134, 210 and 24 unseen single perturbations across two test-training splits". In Panel A these are the split 1 counts only (Replogle K562 134, Replogle RPE1 210, Adamson 24 rows per method); split 2 has 128, 203 and 24, so the two-split totals are 262, 413 and 48. The Fig. 2a means recorded here use all Panel A rows; whether the plotted red lines use both splits is not stated beyond the sheet itself. Reviewer check 2026-10-11: using split 1 rows only would change the per-model means by up to 0.61 (L2 distance) and 0.06 (Pearson delta), and would change the model with the lowest mean L2 distance on Adamson and Replogle RPE1, so the choice is material.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Fig. 2 legend versus Source Data Fig. 2, sheet "Panel A"
Methods restricts the single-perturbation analysis to perturbations predicted by all models: 73 for Adamson, 398 for Replogle K562 and 629 for Replogle RPE1. No sheet of Source Data Fig. 2 has these counts (Panel A: 48, 262 and 413 per method over two splits; Panel C: 20, 214 and 224 test or val rows per method).Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Fig. 2, sheets "Panel A" and "Panel C"
The test and val rows in Panel C (20, 214 and 224 per method for Adamson, Replogle K562 and Replogle RPE1) match neither the Fig. 2 legend counts nor the Methods counts. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2, sheet "Panel C" versus the Fig. 2 legend and Methods
The main text says CPA was not included in the single-perturbation benchmark, but method "cpa" has rows in Panel C. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Main text paragraph 16 versus Source Data Fig. 2, sheet "Panel C"
Methods, coverage and source Geneformer on Adamson (Ahlmann-Eltze et al. 2025)
perturbation-ahlmann-20261011-protocol-adamson-single
Aggregation: Not reported
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines ; Source Data Extended Data Fig. 8, Ahlmann-Eltze et al. 2025 · Source Data Extended Data Fig. 8 (MOESM10), sheet "Panel A", rows with dataset_name "adamson" and method "geneformer", column "r2_delta"; plotted as the red line in Extended Data Fig. 8a Configuration: Linear model (G and P from training data) (Ahlmann-Eltze et al. 2025) Protocol: Adamson: expression prediction for unseen single perturbations (Ahlmann-Eltze et al. 2025 Fig. 2a and Extended Data Fig. 8a) Dataset: Adamson et al. 2016 Perturb-seq, K562 CRISPRi single perturbations (GEARS distribution, Ahlmann-Eltze et al. 2025) 0.717 Pearson delta (1,000 most highly expressed control genes; mean over held-out perturbations and splits)Computed by Rewire from the authors' source data unitless · higher
Uncertainty: Not reported by the source: No spread of the mean is stated; Rewire did not compute one
Coverage: 48/48 perturbation-split rows scored
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Panel A repeats Source Data Fig. 2 Panel A row for row (48, 262 and 413 rows per method over two splits). Methods restricts the analysis to 73, 398 and 629 perturbations predicted by all models; the sheet does not have these counts.Source Data Extended Data Fig. 8, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Extended Data Fig. 8, sheet "Panel A"
The Fig. 2 legend gives "134, 210 and 24 unseen single perturbations across two test-training splits". In Panel A these are the split 1 counts only (Replogle K562 134, Replogle RPE1 210, Adamson 24 rows per method); split 2 has 128, 203 and 24, so the two-split totals are 262, 413 and 48. The Fig. 2a means recorded here use all Panel A rows; whether the plotted red lines use both splits is not stated beyond the sheet itself. Reviewer check 2026-10-11: using split 1 rows only would change the per-model means by up to 0.61 (L2 distance) and 0.06 (Pearson delta), and would change the model with the lowest mean L2 distance on Adamson and Replogle RPE1, so the choice is material.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Fig. 2 legend versus Source Data Fig. 2, sheet "Panel A"
Methods restricts the single-perturbation analysis to perturbations predicted by all models: 73 for Adamson, 398 for Replogle K562 and 629 for Replogle RPE1. No sheet of Source Data Fig. 2 has these counts (Panel A: 48, 262 and 413 per method over two splits; Panel C: 20, 214 and 224 test or val rows per method).Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Fig. 2, sheets "Panel A" and "Panel C"
The test and val rows in Panel C (20, 214 and 224 per method for Adamson, Replogle K562 and Replogle RPE1) match neither the Fig. 2 legend counts nor the Methods counts. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2, sheet "Panel C" versus the Fig. 2 legend and Methods
The main text says CPA was not included in the single-perturbation benchmark, but method "cpa" has rows in Panel C. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Main text paragraph 16 versus Source Data Fig. 2, sheet "Panel C"
Methods, coverage and source Linear model (G and P from training data) on Adamson (Ahlmann-Eltze et al. 2025)
perturbation-ahlmann-20261011-protocol-adamson-single
Aggregation: Not reported
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines ; Source Data Extended Data Fig. 8, Ahlmann-Eltze et al. 2025 · Source Data Extended Data Fig. 8 (MOESM10), sheet "Panel A", rows with dataset_name "adamson" and method "lpm_selftrained", column "r2_delta"; plotted as the red line in Extended Data Fig. 8a Configuration: Mean (Ahlmann-Eltze et al. 2025) Protocol: Adamson: expression prediction for unseen single perturbations (Ahlmann-Eltze et al. 2025 Fig. 2a and Extended Data Fig. 8a) Dataset: Adamson et al. 2016 Perturb-seq, K562 CRISPRi single perturbations (GEARS distribution, Ahlmann-Eltze et al. 2025) 0.724 Pearson delta (1,000 most highly expressed control genes; mean over held-out perturbations and splits)Computed by Rewire from the authors' source data unitless · higher
Uncertainty: Not reported by the source: No spread of the mean is stated; Rewire did not compute one
Coverage: 48/48 perturbation-split rows scored
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Panel A repeats Source Data Fig. 2 Panel A row for row (48, 262 and 413 rows per method over two splits). Methods restricts the analysis to 73, 398 and 629 perturbations predicted by all models; the sheet does not have these counts.Source Data Extended Data Fig. 8, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Extended Data Fig. 8, sheet "Panel A"
The Fig. 2 legend gives "134, 210 and 24 unseen single perturbations across two test-training splits". In Panel A these are the split 1 counts only (Replogle K562 134, Replogle RPE1 210, Adamson 24 rows per method); split 2 has 128, 203 and 24, so the two-split totals are 262, 413 and 48. The Fig. 2a means recorded here use all Panel A rows; whether the plotted red lines use both splits is not stated beyond the sheet itself. Reviewer check 2026-10-11: using split 1 rows only would change the per-model means by up to 0.61 (L2 distance) and 0.06 (Pearson delta), and would change the model with the lowest mean L2 distance on Adamson and Replogle RPE1, so the choice is material.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Fig. 2 legend versus Source Data Fig. 2, sheet "Panel A"
Methods restricts the single-perturbation analysis to perturbations predicted by all models: 73 for Adamson, 398 for Replogle K562 and 629 for Replogle RPE1. No sheet of Source Data Fig. 2 has these counts (Panel A: 48, 262 and 413 per method over two splits; Panel C: 20, 214 and 224 test or val rows per method).Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Fig. 2, sheets "Panel A" and "Panel C"
The test and val rows in Panel C (20, 214 and 224 per method for Adamson, Replogle K562 and Replogle RPE1) match neither the Fig. 2 legend counts nor the Methods counts. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2, sheet "Panel C" versus the Fig. 2 legend and Methods
The main text says CPA was not included in the single-perturbation benchmark, but method "cpa" has rows in Panel C. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Main text paragraph 16 versus Source Data Fig. 2, sheet "Panel C"
Methods, coverage and source Mean on Adamson (Ahlmann-Eltze et al. 2025)
perturbation-ahlmann-20261011-protocol-adamson-single
Aggregation: Not reported
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines ; Source Data Extended Data Fig. 8, Ahlmann-Eltze et al. 2025 · Source Data Extended Data Fig. 8 (MOESM10), sheet "Panel A", rows with dataset_name "adamson" and method "mean", column "r2_delta"; plotted as the red line in Extended Data Fig. 8a Configuration: scBERT (Ahlmann-Eltze et al. 2025) Protocol: Adamson: expression prediction for unseen single perturbations (Ahlmann-Eltze et al. 2025 Fig. 2a and Extended Data Fig. 8a) Dataset: Adamson et al. 2016 Perturb-seq, K562 CRISPRi single perturbations (GEARS distribution, Ahlmann-Eltze et al. 2025) 0.728 Pearson delta (1,000 most highly expressed control genes; mean over held-out perturbations and splits)Computed by Rewire from the authors' source data unitless · higher
Uncertainty: Not reported by the source: No spread of the mean is stated; Rewire did not compute one
Coverage: 48/48 perturbation-split rows scored
Independent external evaluation · Source checked Evidence concern: excluded from comparisons Panel A repeats Source Data Fig. 2 Panel A row for row (48, 262 and 413 rows per method over two splits). Methods restricts the analysis to 73, 398 and 629 perturbations predicted by all models; the sheet does not have these counts.Source Data Extended Data Fig. 8, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Extended Data Fig. 8, sheet "Panel A"
The Fig. 2 legend gives "134, 210 and 24 unseen single perturbations across two test-training splits". In Panel A these are the split 1 counts only (Replogle K562 134, Replogle RPE1 210, Adamson 24 rows per method); split 2 has 128, 203 and 24, so the two-split totals are 262, 413 and 48. The Fig. 2a means recorded here use all Panel A rows; whether the plotted red lines use both splits is not stated beyond the sheet itself. Reviewer check 2026-10-11: using split 1 rows only would change the per-model means by up to 0.61 (L2 distance) and 0.06 (Pearson delta), and would change the model with the lowest mean L2 distance on Adamson and Replogle RPE1, so the choice is material.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Fig. 2 legend versus Source Data Fig. 2, sheet "Panel A"
Methods restricts the single-perturbation analysis to perturbations predicted by all models: 73 for Adamson, 398 for Replogle K562 and 629 for Replogle RPE1. No sheet of Source Data Fig. 2 has these counts (Panel A: 48, 262 and 413 per method over two splits; Panel C: 20, 214 and 224 test or val rows per method).Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Fig. 2, sheets "Panel A" and "Panel C"
The test and val rows in Panel C (20, 214 and 224 per method for Adamson, Replogle K562 and Replogle RPE1) match neither the Fig. 2 legend counts nor the Methods counts. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2, sheet "Panel C" versus the Fig. 2 legend and Methods
The main text says CPA was not included in the single-perturbation benchmark, but method "cpa" has rows in Panel C. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Main text paragraph 16 versus Source Data Fig. 2, sheet "Panel C"
Methods, coverage and source scBERT on Adamson (Ahlmann-Eltze et al. 2025)
perturbation-ahlmann-20261011-protocol-adamson-single
Aggregation: Not reported
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines ; Source Data Extended Data Fig. 8, Ahlmann-Eltze et al. 2025 · Source Data Extended Data Fig. 8 (MOESM10), sheet "Panel A", rows with dataset_name "adamson" and method "scbert", column "r2_delta"; plotted as the red line in Extended Data Fig. 8a Configuration: scGPT (Ahlmann-Eltze et al. 2025) Protocol: Adamson: expression prediction for unseen single perturbations (Ahlmann-Eltze et al. 2025 Fig. 2a and Extended Data Fig. 8a) Dataset: Adamson et al. 2016 Perturb-seq, K562 CRISPRi single perturbations (GEARS distribution, Ahlmann-Eltze et al. 2025) 0.595 Pearson delta (1,000 most highly expressed control genes; mean over held-out perturbations and splits)Computed by Rewire from the authors' source data unitless · higher
Uncertainty: Not reported by the source: No spread of the mean is stated; Rewire did not compute one
Coverage: 48/48 perturbation-split rows scored
Independent external evaluation · Source checked Evidence concern: excluded from comparisons Panel A repeats Source Data Fig. 2 Panel A row for row (48, 262 and 413 rows per method over two splits). Methods restricts the analysis to 73, 398 and 629 perturbations predicted by all models; the sheet does not have these counts.Source Data Extended Data Fig. 8, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Extended Data Fig. 8, sheet "Panel A"
The Fig. 2 legend gives "134, 210 and 24 unseen single perturbations across two test-training splits". In Panel A these are the split 1 counts only (Replogle K562 134, Replogle RPE1 210, Adamson 24 rows per method); split 2 has 128, 203 and 24, so the two-split totals are 262, 413 and 48. The Fig. 2a means recorded here use all Panel A rows; whether the plotted red lines use both splits is not stated beyond the sheet itself. Reviewer check 2026-10-11: using split 1 rows only would change the per-model means by up to 0.61 (L2 distance) and 0.06 (Pearson delta), and would change the model with the lowest mean L2 distance on Adamson and Replogle RPE1, so the choice is material.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Fig. 2 legend versus Source Data Fig. 2, sheet "Panel A"
Methods restricts the single-perturbation analysis to perturbations predicted by all models: 73 for Adamson, 398 for Replogle K562 and 629 for Replogle RPE1. No sheet of Source Data Fig. 2 has these counts (Panel A: 48, 262 and 413 per method over two splits; Panel C: 20, 214 and 224 test or val rows per method).Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Fig. 2, sheets "Panel A" and "Panel C"
The test and val rows in Panel C (20, 214 and 224 per method for Adamson, Replogle K562 and Replogle RPE1) match neither the Fig. 2 legend counts nor the Methods counts. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2, sheet "Panel C" versus the Fig. 2 legend and Methods
The main text says CPA was not included in the single-perturbation benchmark, but method "cpa" has rows in Panel C. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Main text paragraph 16 versus Source Data Fig. 2, sheet "Panel C"
Methods, coverage and source scGPT on Adamson (Ahlmann-Eltze et al. 2025)
perturbation-ahlmann-20261011-protocol-adamson-single
Aggregation: Not reported
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines ; Source Data Extended Data Fig. 8, Ahlmann-Eltze et al. 2025 · Source Data Extended Data Fig. 8 (MOESM10), sheet "Panel A", rows with dataset_name "adamson" and method "scgpt", column "r2_delta"; plotted as the red line in Extended Data Fig. 8a Configuration: UCE (Ahlmann-Eltze et al. 2025) Protocol: Adamson: expression prediction for unseen single perturbations (Ahlmann-Eltze et al. 2025 Fig. 2a and Extended Data Fig. 8a) Dataset: Adamson et al. 2016 Perturb-seq, K562 CRISPRi single perturbations (GEARS distribution, Ahlmann-Eltze et al. 2025) 0.724 Pearson delta (1,000 most highly expressed control genes; mean over held-out perturbations and splits)Computed by Rewire from the authors' source data unitless · higher
Uncertainty: Not reported by the source: No spread of the mean is stated; Rewire did not compute one
Coverage: 48/48 perturbation-split rows scored
Independent external evaluation · Source checked Evidence concern: excluded from comparisons Panel A repeats Source Data Fig. 2 Panel A row for row (48, 262 and 413 rows per method over two splits). Methods restricts the analysis to 73, 398 and 629 perturbations predicted by all models; the sheet does not have these counts.Source Data Extended Data Fig. 8, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Extended Data Fig. 8, sheet "Panel A"
The Fig. 2 legend gives "134, 210 and 24 unseen single perturbations across two test-training splits". In Panel A these are the split 1 counts only (Replogle K562 134, Replogle RPE1 210, Adamson 24 rows per method); split 2 has 128, 203 and 24, so the two-split totals are 262, 413 and 48. The Fig. 2a means recorded here use all Panel A rows; whether the plotted red lines use both splits is not stated beyond the sheet itself. Reviewer check 2026-10-11: using split 1 rows only would change the per-model means by up to 0.61 (L2 distance) and 0.06 (Pearson delta), and would change the model with the lowest mean L2 distance on Adamson and Replogle RPE1, so the choice is material.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Fig. 2 legend versus Source Data Fig. 2, sheet "Panel A"
Methods restricts the single-perturbation analysis to perturbations predicted by all models: 73 for Adamson, 398 for Replogle K562 and 629 for Replogle RPE1. No sheet of Source Data Fig. 2 has these counts (Panel A: 48, 262 and 413 per method over two splits; Panel C: 20, 214 and 224 test or val rows per method).Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Fig. 2, sheets "Panel A" and "Panel C"
The test and val rows in Panel C (20, 214 and 224 per method for Adamson, Replogle K562 and Replogle RPE1) match neither the Fig. 2 legend counts nor the Methods counts. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2, sheet "Panel C" versus the Fig. 2 legend and Methods
The main text says CPA was not included in the single-perturbation benchmark, but method "cpa" has rows in Panel C. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Main text paragraph 16 versus Source Data Fig. 2, sheet "Panel C"
Methods, coverage and source UCE on Adamson (Ahlmann-Eltze et al. 2025)
perturbation-ahlmann-20261011-protocol-adamson-single
Aggregation: Not reported
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines ; Source Data Extended Data Fig. 8, Ahlmann-Eltze et al. 2025 · Source Data Extended Data Fig. 8 (MOESM10), sheet "Panel A", rows with dataset_name "adamson" and method "uce", column "r2_delta"; plotted as the red line in Extended Data Fig. 8a Configuration: GEARS (Ahlmann-Eltze et al. 2025) Protocol: Adamson: expression prediction for unseen single perturbations (Ahlmann-Eltze et al. 2025 Fig. 2a and Extended Data Fig. 8a) Dataset: Adamson et al. 2016 Perturb-seq, K562 CRISPRi single perturbations (GEARS distribution, Ahlmann-Eltze et al. 2025) 3.87 L2 distance (1,000 most highly expressed control genes; mean over held-out perturbations and splits)Computed by Rewire from the authors' source data unit-unreported · lower
Uncertainty: Not reported by the source: No spread of the mean is stated; Rewire did not compute one
Coverage: 48/48 perturbation-split rows scored
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The Fig. 2 legend gives "134, 210 and 24 unseen single perturbations across two test-training splits". In Panel A these are the split 1 counts only (Replogle K562 134, Replogle RPE1 210, Adamson 24 rows per method); split 2 has 128, 203 and 24, so the two-split totals are 262, 413 and 48. The Fig. 2a means recorded here use all Panel A rows; whether the plotted red lines use both splits is not stated beyond the sheet itself. Reviewer check 2026-10-11: using split 1 rows only would change the per-model means by up to 0.61 (L2 distance) and 0.06 (Pearson delta), and would change the model with the lowest mean L2 distance on Adamson and Replogle RPE1, so the choice is material.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Fig. 2 legend versus Source Data Fig. 2, sheet "Panel A"
Methods restricts the single-perturbation analysis to perturbations predicted by all models: 73 for Adamson, 398 for Replogle K562 and 629 for Replogle RPE1. No sheet of Source Data Fig. 2 has these counts (Panel A: 48, 262 and 413 per method over two splits; Panel C: 20, 214 and 224 test or val rows per method).Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Fig. 2, sheets "Panel A" and "Panel C"
The test and val rows in Panel C (20, 214 and 224 per method for Adamson, Replogle K562 and Replogle RPE1) match neither the Fig. 2 legend counts nor the Methods counts. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2, sheet "Panel C" versus the Fig. 2 legend and Methods
The main text says CPA was not included in the single-perturbation benchmark, but method "cpa" has rows in Panel C. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Main text paragraph 16 versus Source Data Fig. 2, sheet "Panel C"
Methods, coverage and source GEARS on Adamson (Ahlmann-Eltze et al. 2025)
perturbation-ahlmann-20261011-protocol-adamson-single
Aggregation: Not reported
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines ; Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2 (MOESM4), sheet "Panel A", rows with dataset_name "adamson" and method "gears", column "l2"; plotted as the red line in Fig. 2a Configuration: Geneformer (Ahlmann-Eltze et al. 2025) Protocol: Adamson: expression prediction for unseen single perturbations (Ahlmann-Eltze et al. 2025 Fig. 2a and Extended Data Fig. 8a) Dataset: Adamson et al. 2016 Perturb-seq, K562 CRISPRi single perturbations (GEARS distribution, Ahlmann-Eltze et al. 2025) 3.91 L2 distance (1,000 most highly expressed control genes; mean over held-out perturbations and splits)Computed by Rewire from the authors' source data unit-unreported · lower
Uncertainty: Not reported by the source: No spread of the mean is stated; Rewire did not compute one
Coverage: 48/48 perturbation-split rows scored
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The Fig. 2 legend gives "134, 210 and 24 unseen single perturbations across two test-training splits". In Panel A these are the split 1 counts only (Replogle K562 134, Replogle RPE1 210, Adamson 24 rows per method); split 2 has 128, 203 and 24, so the two-split totals are 262, 413 and 48. The Fig. 2a means recorded here use all Panel A rows; whether the plotted red lines use both splits is not stated beyond the sheet itself. Reviewer check 2026-10-11: using split 1 rows only would change the per-model means by up to 0.61 (L2 distance) and 0.06 (Pearson delta), and would change the model with the lowest mean L2 distance on Adamson and Replogle RPE1, so the choice is material.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Fig. 2 legend versus Source Data Fig. 2, sheet "Panel A"
Methods restricts the single-perturbation analysis to perturbations predicted by all models: 73 for Adamson, 398 for Replogle K562 and 629 for Replogle RPE1. No sheet of Source Data Fig. 2 has these counts (Panel A: 48, 262 and 413 per method over two splits; Panel C: 20, 214 and 224 test or val rows per method).Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Fig. 2, sheets "Panel A" and "Panel C"
The test and val rows in Panel C (20, 214 and 224 per method for Adamson, Replogle K562 and Replogle RPE1) match neither the Fig. 2 legend counts nor the Methods counts. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2, sheet "Panel C" versus the Fig. 2 legend and Methods
The main text says CPA was not included in the single-perturbation benchmark, but method "cpa" has rows in Panel C. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Main text paragraph 16 versus Source Data Fig. 2, sheet "Panel C"
Methods, coverage and source Geneformer on Adamson (Ahlmann-Eltze et al. 2025)
perturbation-ahlmann-20261011-protocol-adamson-single
Aggregation: Not reported
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines ; Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2 (MOESM4), sheet "Panel A", rows with dataset_name "adamson" and method "geneformer", column "l2"; plotted as the red line in Fig. 2a Configuration: Linear model (G and P from training data) (Ahlmann-Eltze et al. 2025) Protocol: Adamson: expression prediction for unseen single perturbations (Ahlmann-Eltze et al. 2025 Fig. 2a and Extended Data Fig. 8a) Dataset: Adamson et al. 2016 Perturb-seq, K562 CRISPRi single perturbations (GEARS distribution, Ahlmann-Eltze et al. 2025) 4.08 L2 distance (1,000 most highly expressed control genes; mean over held-out perturbations and splits)Computed by Rewire from the authors' source data unit-unreported · lower
Uncertainty: Not reported by the source: No spread of the mean is stated; Rewire did not compute one
Coverage: 48/48 perturbation-split rows scored
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The Fig. 2 legend gives "134, 210 and 24 unseen single perturbations across two test-training splits". In Panel A these are the split 1 counts only (Replogle K562 134, Replogle RPE1 210, Adamson 24 rows per method); split 2 has 128, 203 and 24, so the two-split totals are 262, 413 and 48. The Fig. 2a means recorded here use all Panel A rows; whether the plotted red lines use both splits is not stated beyond the sheet itself. Reviewer check 2026-10-11: using split 1 rows only would change the per-model means by up to 0.61 (L2 distance) and 0.06 (Pearson delta), and would change the model with the lowest mean L2 distance on Adamson and Replogle RPE1, so the choice is material.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Fig. 2 legend versus Source Data Fig. 2, sheet "Panel A"
Methods restricts the single-perturbation analysis to perturbations predicted by all models: 73 for Adamson, 398 for Replogle K562 and 629 for Replogle RPE1. No sheet of Source Data Fig. 2 has these counts (Panel A: 48, 262 and 413 per method over two splits; Panel C: 20, 214 and 224 test or val rows per method).Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Fig. 2, sheets "Panel A" and "Panel C"
The test and val rows in Panel C (20, 214 and 224 per method for Adamson, Replogle K562 and Replogle RPE1) match neither the Fig. 2 legend counts nor the Methods counts. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2, sheet "Panel C" versus the Fig. 2 legend and Methods
The main text says CPA was not included in the single-perturbation benchmark, but method "cpa" has rows in Panel C. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Main text paragraph 16 versus Source Data Fig. 2, sheet "Panel C"
Methods, coverage and source Linear model (G and P from training data) on Adamson (Ahlmann-Eltze et al. 2025)
perturbation-ahlmann-20261011-protocol-adamson-single
Aggregation: Not reported
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines ; Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2 (MOESM4), sheet "Panel A", rows with dataset_name "adamson" and method "lpm_selftrained", column "l2"; plotted as the red line in Fig. 2a Configuration: Mean (Ahlmann-Eltze et al. 2025) Protocol: Adamson: expression prediction for unseen single perturbations (Ahlmann-Eltze et al. 2025 Fig. 2a and Extended Data Fig. 8a) Dataset: Adamson et al. 2016 Perturb-seq, K562 CRISPRi single perturbations (GEARS distribution, Ahlmann-Eltze et al. 2025) 3.94 L2 distance (1,000 most highly expressed control genes; mean over held-out perturbations and splits)Computed by Rewire from the authors' source data unit-unreported · lower
Uncertainty: Not reported by the source: No spread of the mean is stated; Rewire did not compute one
Coverage: 48/48 perturbation-split rows scored
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The Fig. 2 legend gives "134, 210 and 24 unseen single perturbations across two test-training splits". In Panel A these are the split 1 counts only (Replogle K562 134, Replogle RPE1 210, Adamson 24 rows per method); split 2 has 128, 203 and 24, so the two-split totals are 262, 413 and 48. The Fig. 2a means recorded here use all Panel A rows; whether the plotted red lines use both splits is not stated beyond the sheet itself. Reviewer check 2026-10-11: using split 1 rows only would change the per-model means by up to 0.61 (L2 distance) and 0.06 (Pearson delta), and would change the model with the lowest mean L2 distance on Adamson and Replogle RPE1, so the choice is material.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Fig. 2 legend versus Source Data Fig. 2, sheet "Panel A"
Methods restricts the single-perturbation analysis to perturbations predicted by all models: 73 for Adamson, 398 for Replogle K562 and 629 for Replogle RPE1. No sheet of Source Data Fig. 2 has these counts (Panel A: 48, 262 and 413 per method over two splits; Panel C: 20, 214 and 224 test or val rows per method).Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Fig. 2, sheets "Panel A" and "Panel C"
The test and val rows in Panel C (20, 214 and 224 per method for Adamson, Replogle K562 and Replogle RPE1) match neither the Fig. 2 legend counts nor the Methods counts. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2, sheet "Panel C" versus the Fig. 2 legend and Methods
The main text says CPA was not included in the single-perturbation benchmark, but method "cpa" has rows in Panel C. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Main text paragraph 16 versus Source Data Fig. 2, sheet "Panel C"
Methods, coverage and source Mean on Adamson (Ahlmann-Eltze et al. 2025)
perturbation-ahlmann-20261011-protocol-adamson-single
Aggregation: Not reported
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines ; Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2 (MOESM4), sheet "Panel A", rows with dataset_name "adamson" and method "mean", column "l2"; plotted as the red line in Fig. 2a Configuration: scBERT (Ahlmann-Eltze et al. 2025) Protocol: Adamson: expression prediction for unseen single perturbations (Ahlmann-Eltze et al. 2025 Fig. 2a and Extended Data Fig. 8a) Dataset: Adamson et al. 2016 Perturb-seq, K562 CRISPRi single perturbations (GEARS distribution, Ahlmann-Eltze et al. 2025) 3.88 L2 distance (1,000 most highly expressed control genes; mean over held-out perturbations and splits)Computed by Rewire from the authors' source data unit-unreported · lower
Uncertainty: Not reported by the source: No spread of the mean is stated; Rewire did not compute one
Coverage: 48/48 perturbation-split rows scored
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The Fig. 2 legend gives "134, 210 and 24 unseen single perturbations across two test-training splits". In Panel A these are the split 1 counts only (Replogle K562 134, Replogle RPE1 210, Adamson 24 rows per method); split 2 has 128, 203 and 24, so the two-split totals are 262, 413 and 48. The Fig. 2a means recorded here use all Panel A rows; whether the plotted red lines use both splits is not stated beyond the sheet itself. Reviewer check 2026-10-11: using split 1 rows only would change the per-model means by up to 0.61 (L2 distance) and 0.06 (Pearson delta), and would change the model with the lowest mean L2 distance on Adamson and Replogle RPE1, so the choice is material.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Fig. 2 legend versus Source Data Fig. 2, sheet "Panel A"
Methods restricts the single-perturbation analysis to perturbations predicted by all models: 73 for Adamson, 398 for Replogle K562 and 629 for Replogle RPE1. No sheet of Source Data Fig. 2 has these counts (Panel A: 48, 262 and 413 per method over two splits; Panel C: 20, 214 and 224 test or val rows per method).Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Fig. 2, sheets "Panel A" and "Panel C"
The test and val rows in Panel C (20, 214 and 224 per method for Adamson, Replogle K562 and Replogle RPE1) match neither the Fig. 2 legend counts nor the Methods counts. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2, sheet "Panel C" versus the Fig. 2 legend and Methods
The main text says CPA was not included in the single-perturbation benchmark, but method "cpa" has rows in Panel C. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Main text paragraph 16 versus Source Data Fig. 2, sheet "Panel C"
Methods, coverage and source scBERT on Adamson (Ahlmann-Eltze et al. 2025)
perturbation-ahlmann-20261011-protocol-adamson-single
Aggregation: Not reported
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines ; Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2 (MOESM4), sheet "Panel A", rows with dataset_name "adamson" and method "scbert", column "l2"; plotted as the red line in Fig. 2a Configuration: scGPT (Ahlmann-Eltze et al. 2025) Protocol: Adamson: expression prediction for unseen single perturbations (Ahlmann-Eltze et al. 2025 Fig. 2a and Extended Data Fig. 8a) Dataset: Adamson et al. 2016 Perturb-seq, K562 CRISPRi single perturbations (GEARS distribution, Ahlmann-Eltze et al. 2025) 4.95 L2 distance (1,000 most highly expressed control genes; mean over held-out perturbations and splits)Computed by Rewire from the authors' source data unit-unreported · lower
Uncertainty: Not reported by the source: No spread of the mean is stated; Rewire did not compute one
Coverage: 48/48 perturbation-split rows scored
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The Fig. 2 legend gives "134, 210 and 24 unseen single perturbations across two test-training splits". In Panel A these are the split 1 counts only (Replogle K562 134, Replogle RPE1 210, Adamson 24 rows per method); split 2 has 128, 203 and 24, so the two-split totals are 262, 413 and 48. The Fig. 2a means recorded here use all Panel A rows; whether the plotted red lines use both splits is not stated beyond the sheet itself. Reviewer check 2026-10-11: using split 1 rows only would change the per-model means by up to 0.61 (L2 distance) and 0.06 (Pearson delta), and would change the model with the lowest mean L2 distance on Adamson and Replogle RPE1, so the choice is material.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Fig. 2 legend versus Source Data Fig. 2, sheet "Panel A"
Methods restricts the single-perturbation analysis to perturbations predicted by all models: 73 for Adamson, 398 for Replogle K562 and 629 for Replogle RPE1. No sheet of Source Data Fig. 2 has these counts (Panel A: 48, 262 and 413 per method over two splits; Panel C: 20, 214 and 224 test or val rows per method).Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Fig. 2, sheets "Panel A" and "Panel C"
The test and val rows in Panel C (20, 214 and 224 per method for Adamson, Replogle K562 and Replogle RPE1) match neither the Fig. 2 legend counts nor the Methods counts. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2, sheet "Panel C" versus the Fig. 2 legend and Methods
The main text says CPA was not included in the single-perturbation benchmark, but method "cpa" has rows in Panel C. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Main text paragraph 16 versus Source Data Fig. 2, sheet "Panel C"
Methods, coverage and source scGPT on Adamson (Ahlmann-Eltze et al. 2025)
perturbation-ahlmann-20261011-protocol-adamson-single
Aggregation: Not reported
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines ; Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2 (MOESM4), sheet "Panel A", rows with dataset_name "adamson" and method "scgpt", column "l2"; plotted as the red line in Fig. 2a Configuration: UCE (Ahlmann-Eltze et al. 2025) Protocol: Adamson: expression prediction for unseen single perturbations (Ahlmann-Eltze et al. 2025 Fig. 2a and Extended Data Fig. 8a) Dataset: Adamson et al. 2016 Perturb-seq, K562 CRISPRi single perturbations (GEARS distribution, Ahlmann-Eltze et al. 2025) 3.94 L2 distance (1,000 most highly expressed control genes; mean over held-out perturbations and splits)Computed by Rewire from the authors' source data unit-unreported · lower
Uncertainty: Not reported by the source: No spread of the mean is stated; Rewire did not compute one
Coverage: 48/48 perturbation-split rows scored
Independent external evaluation · Source checked Evidence concern: excluded from comparisons The Fig. 2 legend gives "134, 210 and 24 unseen single perturbations across two test-training splits". In Panel A these are the split 1 counts only (Replogle K562 134, Replogle RPE1 210, Adamson 24 rows per method); split 2 has 128, 203 and 24, so the two-split totals are 262, 413 and 48. The Fig. 2a means recorded here use all Panel A rows; whether the plotted red lines use both splits is not stated beyond the sheet itself. Reviewer check 2026-10-11: using split 1 rows only would change the per-model means by up to 0.61 (L2 distance) and 0.06 (Pearson delta), and would change the model with the lowest mean L2 distance on Adamson and Replogle RPE1, so the choice is material.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Fig. 2 legend versus Source Data Fig. 2, sheet "Panel A"
Methods restricts the single-perturbation analysis to perturbations predicted by all models: 73 for Adamson, 398 for Replogle K562 and 629 for Replogle RPE1. No sheet of Source Data Fig. 2 has these counts (Panel A: 48, 262 and 413 per method over two splits; Panel C: 20, 214 and 224 test or val rows per method).Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Methods 'Single perturbation benchmark setup' paragraph 4 versus Source Data Fig. 2, sheets "Panel A" and "Panel C"
The test and val rows in Panel C (20, 214 and 224 per method for Adamson, Replogle K562 and Replogle RPE1) match neither the Fig. 2 legend counts nor the Methods counts. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2, sheet "Panel C" versus the Fig. 2 legend and Methods
The main text says CPA was not included in the single-perturbation benchmark, but method "cpa" has rows in Panel C. Fig. 2c was not extracted.Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Main text paragraph 16 versus Source Data Fig. 2, sheet "Panel C"
Methods, coverage and source UCE on Adamson (Ahlmann-Eltze et al. 2025)
perturbation-ahlmann-20261011-protocol-adamson-single
Aggregation: Not reported
Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines ; Source Data Fig. 2, Ahlmann-Eltze et al. 2025 · Source Data Fig. 2 (MOESM4), sheet "Panel A", rows with dataset_name "adamson" and method "uce", column "l2"; plotted as the red line in Fig. 2a