- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies reference-based deconvolution under systematic reference shifts, with a focus on estimating cell-type fractions from bulk RNA-seq data using single-cell expression references. The authors propose PopLS, which aggregates marker-restricted post-fit residuals across related samples to estimate a shared direction for correcting fraction estimation bias. The paper provides theoretical analyses of residual geometry, bias identifiability, population residual concentration, and sufficient conditions for safe correction. Experiments on synthetic data, eight cancer pseudo-bulk datasets, and two real-bulk cohorts show that PopLS achieves lower estimation errors than the compared deconvolution methods.
Critical strengths
-
Using population-level residual information to correct systematic biases in reference-based deconvolution is well motivated. The theoretical analysis distinguishes between the components of reference shift that affect fraction estimates and those observable through reconstruction residuals. The identifiability analysis and the proposed structural bridge provide a useful framework for understanding when residual-based bias correction is possible.
-
The experimental evaluation is extensive, covering controlled synthetic experiments, eight cancer pseudo-bulk datasets, and two real-bulk cohorts.
Critical weaknesses
-
The theoretical analysis in Sections 3–4 is based on ordinary least squares, while the practical implementation in Appendix D uses reference-variance-based feature weights. It is unclear how the theoretical guarantees extend to the weighted estimator. Moreover, the safe correction condition in Theorem 4.5 involves unobservable quantities, whereas the implemented algorithm relies on an empirical gate and step-size selection. The connection between the theoretical guarantees and the actual implementation remains insufficiently established.
-
Table 1 shows that PopLS outperforms several existing methods, but the contribution of population-level residual aggregation is not fully isolated. The complete pipeline combines marker selection, residual aggregation, direction calibration, gating, and step-size selection. A controlled comparison with individual-sample residual correction under the same reference-shift conditions is missing, making it difficult to determine how much of the improvement comes specifically from sharing residual information across samples.
Questions for authors
-
The theoretical analysis uses ordinary least squares, while the practical implementation incorporates reference-variance-based feature weights (Appendix D). Could the authors clarify whether the residual projection identities and theoretical guarantees still hold for the weighted estimator?
-
Theorem 4.5 provides a sufficient condition for safe correction, but the implemented algorithm uses an empirical gate and step-size selection. Could the authors report how frequently the correction increases estimation errors for individual samples in the two real-bulk cohorts?
-
Could the authors provide a controlled comparison between population-aggregated and individual-sample residual correction under the same reference-shift conditions, while keeping the other components as consistent as possible? This would help clarify the contribution of population-level information sharing to the improvements in Table 1.
2 comments
Responses to Reviewer wvSR, Part 2/2
Q2. Frequency of increased sample-level error
We add a bulk-wise comparison to test how frequently our post-correction algorithm improves the initial estimates (which are given by a baseline, ReDeconv). Below we show the percentages of improved, worsened, and tied cases. We see that most of the time our correction improves. If the improved/worsened magnitudes were incorporated, the overall error rate of all bulks also decreases -- see Table 1(b) in the main text, compared with ReDeconv.
| Dataset | Bulks | Improved, n (%) | Worsened, n (%) | Tied, n (%) |
|---|---|---|---|---|
| Altman | 322 | 282 (87.6%) | 40 (12.4%) | 0 (0.0%) |
| Morandini | 156 | 104 (66.7%) | 52 (33.3%) | 0 (0.0%) |
| DREAM | 96 | 55 (57.3%) | 22 (22.9%) | 19 (19.8%) |
The above results have been added to Appendix E.3.3 and Table 11.
Q3/W2. Isolating population residuals vs. individual residuals on real-bulk data
We add an ablation study to compare the initial fits, single-residual correction, and population-residual correction on the three real-bulk datasets (results added to Appendix E.3.3, Table 10). Under the same experimental protocol, population-residual correction yields lower RMSE, higher Pearson correlation, and CCC than single-residual correction on all datasets.
| Dataset (N) | Metric | Initial | Individual | Population | Improvement |
|---|---|---|---|---|---|
| Altman (322) | RMSE ↓ | 0.13138 | 0.08734 | 0.08313 | +0.00421 |
| Pearson ↑ | 0.78994 | 0.91345 | 0.92302 | +0.00957 | |
| CCC ↑ | 0.78745 | 0.89606 | 0.90615 | +0.01009 | |
| Morandini (156) | RMSE ↓ | 0.06795 | 0.06286 | 0.06255 | +0.00032 |
| Pearson ↑ | 0.79163 | 0.82475 | 0.82654 | +0.00179 | |
| CCC ↑ | 0.77642 | 0.81321 | 0.81399 | +0.00077 | |
| DREAM (96) | RMSE ↓ | 0.06883 | 0.06804 | 0.06726 | +0.00078 |
| Pearson ↑ | 0.48056 | 0.47451 | 0.50425 | +0.02974 | |
| CCC ↑ | 0.46565 | 0.45090 | 0.48808 | +0.03718 |
Note that single-residual correction has achieved most of the improvement compared to the initial fits, which verifies the effect of our proposed structural bridge introduced in Section 4.3. Meanwhile, the further improvement by population-residual correction is relatively small, and this is due to the effective-sample-size law of residual error reduction, proven in Theorem 4.3. In our context, , the number of bulks, is much smaller than , the number of cells, so the increase of based on a large does not yield much decrease in residual estimation error. Further analysis comparing the decrease in bias reduction is similarly given in Appendix C.6, Proposition C.6.
Thank you again for your feedback, and we look forward to your further comments.
Responses to Reviewer wvSR, Part 1/2
We are very grateful for your constructive feedback. Below are point-by-point responses to the weaknesses and questions raised.
Q1/W1. Weighted least squares
Yes, the residual projection identities and theoretical guarantees still hold for the weighted least-squares (WLS). We extend ordinary least squares (OLS) theory to unconstrained WLS (Appendix C.7). In brief, let U be a full-column-rank reference expression matrix, G a symmetric positive-definite feature-weight matrix (i.e., gene weights). The WLS residual projector QG satisfies
\qquad Q_G^2=Q_G,\quad Q_GU=0,\quad Q_G^\top G=GQ_G.$$ $Q_G$ maps each bulk expression vector to its post-fit WLS residual. With $G=I$, it reduces to the OLS residual projector $Q=I-UU^\dagger$ (defined at the beginning of Section 4.1). These identities give the weighted bias-residual decomposition. Under the weighted shift-response and noise assumptions, the extension also gives conditional guarantees for direction recovery and squared fraction-error descent (Appendix C.7, Propositions C.7–C.8). **W1. Theoretical analysis vs. the actual implementation** The connection between our actual implementation (now summarized in Algorithm 1 at the end of Section 4.5) and the theoretical guarantees (Theorem 4.5) was indeed insufficiently explained. We thus add detailed descriptions on how the gate (Appendix C.8) and the empirical step size (Appendix C.5) are linked to the theoretical algorithm. Gate function. As the theoretical algorithm assumes reliable correcting conditions which may not be fully observable, we had introduced a gate function. The gate is to accept or reject a bias correction proposal for all estimates. Eq. (50) states accept iff. the following thresholds are satisfied: $$(p\geq0)\land(b\geq0)\land(r\geq0.443865)\land(a\geq0.0231)$$ where $p$ is the poplulation stability, $b$ the boostrap stability, $r$ the marker-residual enrichment, and $a$ the correction-signal magnitude. Once the thresholds are fixed, each decision uses only the reference and the observed bulk data. The non-negativity of $p$ and $b$ ensures the correct direction; the thresholds of $r$ and $a$ have been calibrated in simulation studies (Appendix E.2.3, Figure 8). The selected thresholds were searched optimal in terms of retaining the most correctly oriented candidates and rejecting all candidates labeled as harmful. Empirical step size selection. The step size selected for the theoretical algorithm was conservative and also requires unobservable quantities. Once the gate accepts a candidate direction $\mathbf v$, we select an empirical step using the criterion in Appendix C.5. For a candidate step $\lambda>0$, the proposed fraction update is $\mathbf f_i(\lambda)=\widehat{\mathbf f}_i+\lambda\mathbf v$. Its mean predicted response in marker-expression space is $$\bar{\mathbf{h}}(\lambda) = \frac{1}{m}\sum_{i=1}^{m} U_{\mathcal{M}} \bigl(\mathbf{f}_i(\lambda)-\widehat{\mathbf{f}}_i\bigr) = \lambda U_{\mathcal{M}}\mathbf{v}.$$ Given the observed population marker residual $\bar{\mathbf{r}}_{\mathcal{M}}$ and the reference-variance weight matrix $G_\lambda$, we compute $$\beta(\lambda) = \frac{\bar{\mathbf{h}}(\lambda)^\top G_\lambda \bar{\mathbf{r}}_{\mathcal{M}}} {\bar{\mathbf{h}}(\lambda)^\top G_\lambda \bar{\mathbf{h}}(\lambda)}, \qquad \widehat{\lambda} \in \arg\min_{\lambda\in\Lambda_{\mathrm{adm}}} \bigl(\beta(\lambda)-1\bigr)^2.$$ When $\beta(\lambda)=1$, the weighted projection of the observed residual onto the candidate response equals that response. Selecting the step with $\beta(\lambda)$ closest to one thus matches the predicted response to the observed residual along the proposed response direction. The selector uses observable quantities, while the fraction-error descent guarantee in Theorem 4.5 remains conditional on its sufficient criterion. We also held the recovered direction fixed and varied the step size to examine how MAE responds to the update magnitude (Figure 1(d), Appendices E.2.2 and E.3.1–E.3.2).