Configuration: 1-NN on GenePT embeddings (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 3.39% Proportion (share of perturbation-gene pairs predicted as differentially expressed (compare with the printed true DE prevalence))percent · unknown
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source 1-NN on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 1, row '1-NN', column 'Average' Configuration: Weighted 10-NN on GenePT embeddings (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 0.91% Proportion (share of perturbation-gene pairs predicted as differentially expressed (compare with the printed true DE prevalence))percent · unknown
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source 10-NN on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 1, row '10-NN', column 'Average' Configuration: Per-gene L2 logistic regression on GenePT embeddings (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 14.7% Proportion (share of perturbation-gene pairs predicted as differentially expressed (compare with the printed true DE prevalence))percent · unknown
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source Logistic regression on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 1, row 'Logistic regression', column 'Average' Configuration: Perturbation majority vote (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 0.03% Proportion (share of perturbation-gene pairs predicted as differentially expressed (compare with the printed true DE prevalence))percent · unknown
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source Majority vote on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 1, row 'Majority vote', column 'Average' Configuration: Scouter (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 64.1% Proportion (share of perturbation-gene pairs predicted as differentially expressed (compare with the printed true DE prevalence))percent · unknown
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source Scouter on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 1, row 'Scouter', column 'Average' Configuration: scVI perturbation model, mean output (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 73.4% Proportion (share of perturbation-gene pairs predicted as differentially expressed (compare with the printed true DE prevalence))percent · unknown
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source scVI mean on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 1, row 'scVI mean', column 'Average' Configuration: scVI perturbation model, sampled output (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 4.28% Proportion (share of perturbation-gene pairs predicted as differentially expressed (compare with the printed true DE prevalence))percent · unknown
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source scVI sampling on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 1, row 'scVI sampling', column 'Average' Configuration: VAE adapted from scGen (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 62.6% Proportion (share of perturbation-gene pairs predicted as differentially expressed (compare with the printed true DE prevalence))percent · unknown
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source VAE scGen on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 1, row 'VAE scGen', column 'Average' Configuration: 1-NN on GenePT embeddings (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 98.8% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q1 (0–0.016); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source 1-NN on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row '1-NN', column 'Q1 (0–0.016)' Configuration: Weighted 10-NN on GenePT embeddings (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 99.9% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q1 (0–0.016); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source 10-NN on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row '10-NN', column 'Q1 (0–0.016)' Configuration: Per-gene L2 logistic regression on GenePT embeddings (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 92.4% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q1 (0–0.016); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source Logistic regression on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row 'Logistic regression', column 'Q1 (0–0.016)' Configuration: Perturbation majority vote (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 100% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q1 (0–0.016); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source Majority vote on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row 'Majority vote', column 'Q1 (0–0.016)' Configuration: Scouter (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 22.6% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q1 (0–0.016); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source Scouter on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row 'Scouter', column 'Q1 (0–0.016)' Configuration: scVI perturbation model, mean output (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 14.3% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q1 (0–0.016); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source scVI mean on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row 'scVI mean', column 'Q1 (0–0.016)' Configuration: scVI perturbation model, sampled output (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 98.1% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q1 (0–0.016); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source scVI sampling on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row 'scVI sampling', column 'Q1 (0–0.016)' Configuration: VAE adapted from scGen (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 23.4% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q1 (0–0.016); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source VAE scGen on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row 'VAE scGen', column 'Q1 (0–0.016)' Configuration: 1-NN on GenePT embeddings (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 98.6% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q2 (0.016–0.034); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source 1-NN on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row '1-NN', column 'Q2 (0.016–0.034)' Configuration: Weighted 10-NN on GenePT embeddings (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 99.8% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q2 (0.016–0.034); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source 10-NN on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row '10-NN', column 'Q2 (0.016–0.034)' Configuration: Per-gene L2 logistic regression on GenePT embeddings (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 91.5% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q2 (0.016–0.034); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source Logistic regression on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row 'Logistic regression', column 'Q2 (0.016–0.034)' Configuration: Perturbation majority vote (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 100% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q2 (0.016–0.034); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source Majority vote on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row 'Majority vote', column 'Q2 (0.016–0.034)' Configuration: Scouter (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 24.9% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q2 (0.016–0.034); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source Scouter on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row 'Scouter', column 'Q2 (0.016–0.034)' Configuration: scVI perturbation model, mean output (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 16.1% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q2 (0.016–0.034); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source scVI mean on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row 'scVI mean', column 'Q2 (0.016–0.034)' Configuration: scVI perturbation model, sampled output (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 97.9% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q2 (0.016–0.034); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Independent external evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source scVI sampling on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row 'scVI sampling', column 'Q2 (0.016–0.034)' Configuration: VAE adapted from scGen (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 25.8% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q2 (0.016–0.034); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source VAE scGen on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row 'VAE scGen', column 'Q2 (0.016–0.034)' Configuration: 1-NN on GenePT embeddings (Sun et al. 2026) Protocol: Sun et al. 2026 unseen-perturbation DE classification (within cell line) Dataset subset: K562, RPE1, HepG2 and Jurkat test sets (Sun et al. 2026) 98.2% Accuracy (3-class (non-DE, up, down) accuracy within true |log2 FC| quantile Q3 (0.034–0.060); quantile bins per cell line, averaged over four cell lines)percent · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Methods describe results on both the full data and a PerturbQA-downsampled subset ('We annotate our results to distinguish between these two settings'), but the captions of Tables 1-4 do not say which setting they report. The printed true DE prevalence (1.75-4.41%) matches the full data's 95-98% non-DE share (Table 5), not the downsampled set; this is an inference, not a printed statement.Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Section 3.2 'Full vs. Curated Dataset'; Tables 1-5
Methods, coverage and source 1-NN on Sun et al. 2026 within-cell-line DE classification, four cell lines
lit-cells-networks-20261011-sun2026de-protocol-within-cell-line
Aggregation: Not reported
Beyond Expression Prediction: Benchmarking Differential Expression Classification in Single-Cell Perturbation Models · Table 2, row '1-NN', column 'Q3 (0.034–0.060)'