- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The main experiments cover 15 gene-level tasks, with an additional analysis of attention-based methods and external random forest (RF) configurations on nine binary classification tasks. The paper also examines sensitivity to probes, evaluation metrics, decision thresholds, and training budgets. Expression-weighted aggregation is used to assess how performance declines when gene representations are transferred to cell-level tasks.
Critical strengths
Equation (5) on PDF page 4 and the description on page 17 require all preprocessing and representation learning to use only the outer-training data, with refitting for every task–condition–operator–seed–fold combination. This helps prevent leakage from test-set feature statistics.
The authors acknowledge that all intervals overlap, attention shows no consistent advantage, method rankings depend on the metric, probe, and threshold, and gene-level gains do not necessarily transfer to cell-level tasks. The claim–evidence ledger in Appendix 15 and the limitations in Appendix 21 explicitly identify conclusions that the current results do not support.
Critical weaknesses
The caption of Table 7 on PDF page 18 states that “All operators share the view adapters,” whereas the text on the same page says that the three baselines “never touch the trained adapters.” These descriptions need to be reconciled.
Equation (7) on PDF page 18 uses an undefined function F. The text describes direct concatenation after budget allocation, but Table 8 lists “concatenation + projection.” Please clarify whether F is the identity mapping or an additional projection.
The text in Figures 4–14 on PDF pages 38–41 is generally too small to read comfortably. Figure 14 also presents pancreas and myeloid results, although its caption states that documentation for these datasets is outside the scope of the current report.
Questions for authors
Could the authors use the existing task-, fold-, and seed-level records to report paired difference intervals for the top two methods under each condition, together with the frequency with which each method ranks first? If differences smaller than the variability across seeds and folds are treated as ties, how many tasks still show a winner switch?
For tasks with only about 84 outer-training samples, how does PCA add produce 256 principal components? How does PCA then concat handle a requested dimension that exceeds the sample rank? What exactly does F in Eq. (7) denote?