acceptodds
Under review as a conference paper at ICLR 2027

Identifying Bias Directions from Population Residuals in Reference-shifted Deconvolution

Abstract

We study reference-based linear deconvolution problems, which estimate component fractions from aggregate observations using a potentially shifted component profile. Our study is motivated by gene expression deconvolution. In biomedicine, bulk RNA-seq measures aggregate gene expression in a tissue; to infer cell-type composition, one deconvolutes it by referring to a cell-type gene expression profile, which is yet measured from separate tissues and by a different technology (scRNA-seq). Consequently, the reference profile can significantly differ from the true profile that generated the aggregates. Previous methods either neglect this shift or address it only in preprocessing; they then mostly use least squares to perform deconvolution for each tissue individually. Given that a cohort of similar bulk RNA-seq measurements is usually available, we propose to use post-fit residuals together to correct individual deconvolution bias. Our algorithm is developed from rigorous analysis. We first show that under a shared shift in reference, the population residuals alone cannot identify the fraction biases, motivating us to build a structural bridge that connects markered residuals to the population fraction error. Then, based on the latter, we study the conditions under which individual fraction bias can be safely corrected. The efficacy of our correction algorithm is proven under the PAC framework. Experimentally, we first conduct synthetic studies that demonstrate the residual geometry, bias identifiability, and PAC learning effectiveness. Second, we test an instantiation of our algorithm named PopLS on eight cancer pseudo-bulk datasets and two real-bulk cohorts, which further verify the benefits of population-based deconvolution.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Bulk RNA-seq deconvolution, Reference shifted deconvolution, Residual diagnostics, PAC learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_wvSROpenReview ↗
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
AuthorsOpenReview ↗

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 UU be a full-column-rank reference expression matrix, GG a symmetric positive-definite feature-weight matrix (i.e., gene weights). The WLS residual projector QGQ_G 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).
AuthorsOpenReview ↗

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.

DatasetBulksImproved, n (%)Worsened, n (%)Tied, n (%)
Altman322282 (87.6%)40 (12.4%)0 (0.0%)
Morandini156104 (66.7%)52 (33.3%)0 (0.0%)
DREAM9655 (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)MetricInitialIndividualPopulationImprovement
Altman (322)RMSE ↓0.131380.087340.08313+0.00421
Pearson ↑0.789940.913450.92302+0.00957
CCC ↑0.787450.896060.90615+0.01009
Morandini (156)RMSE ↓0.067950.062860.06255+0.00032
Pearson ↑0.791630.824750.82654+0.00179
CCC ↑0.776420.813210.81399+0.00077
DREAM (96)RMSE ↓0.068830.068040.06726+0.00078
Pearson ↑0.480560.474510.50425+0.02974
CCC ↑0.465650.450900.48808+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 TT introduced in Section 4.3. Meanwhile, the further improvement by population-residual correction is relatively small, and this is due to the O((mn)−1/2)O((mn)^{-1/2}) effective-sample-size law of residual error reduction, proven in Theorem 4.3. In our context, mm, the number of bulks, is much smaller than nn, the number of cells, so the increase of mm based on a large nn 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.

Reviewer_7RJYOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses reference-shifted linear deconvolution, where differences between the reference profile and the true profile lead to biased component fraction estimates. To address this problem, the authors propose PopLS, a population-based method that uses marker-restricted residuals across multiple samples to correct individual estimation biases.

Critical strengths
  1. The paper addresses an important limitation of existing reference-based deconvolution methods by leveraging population-level residual information to correct individual estimation biases.

  2. The paper provides extensive theoretical analyses of residual geometry, identifiability, and sufficient conditions for safe correction, supported by synthetic experiments and promising results on pseudo-bulk and real-bulk datasets.

Critical weaknesses
  1. PopLS estimates correction directions from marker-restricted residuals, but these directions may not necessarily match the true fraction biases. Although the paper introduces an observable gate to determine whether a correction should be applied, it remains unclear how reliably this gate can identify incorrect correction directions without knowing the true fraction biases. Further analysis of the gate's effectiveness would strengthen the proposed method.

  2. The experimental validation on real-world data is limited. Although the paper evaluates eight cancer pseudo-bulk datasets, the reference shifts are artificially introduced. Moreover, the real-bulk evaluation includes only two blood/PBMC cohorts with three and five cell types, respectively. Additional evaluations across different tissues and reference conditions would strengthen the generalizability of the proposed method.

Questions for authors

Please see the concerns detailed in the Weaknesses.

1 comment
AuthorsOpenReview ↗

Responses to Reviewer 7RJY

We are very grateful for your constructive feedback. Below are point-by-point responses to the weaknesses raised.

W1. Gate analysis

Appendix C.8 has been added to detail the gate algorithm and analysis. The objective of 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≥0)∧(b≥0)∧(r≥0.443865)∧(a≥0.0231)(p\geq0)\land(b\geq0)\land(r\geq0.443865)\land(a\geq0.0231) where pp is the poplulation stability, bb the boostrap stability, rr the marker-residual enrichment, and aa the correction-signal magnitude. Once the thresholds are fixed, each decision uses only the reference and the observed bulk data.

The non-negativity of pp and bb ensures the correct direction; the thresholds of rr and aa 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.

The rest of Appendix C.8 analyzes the gate limitations. Identical bulk observations and residuals can correspond to opposite true corrections, so passing the gate provides observable support for applying a cohort-level correction.

W2. New real‑bulk validation

Public bulk RNA‑seq datasets paired with known cell‑type proportions are rare, which limits direct validation of deconvolution methods across tissues and reference conditions (deconvData, Dietrich et al., 2026; Scaden, Menden et al., 2020). We tried our best and found a new dataset DREAM (GSE199324), a benchmark of 96 bulk RNA-seq measurements with known mixing proportions of nine cell types, as well as a separate single-cell RNA reference derived from Tabula Sapiens. Our method has also been shown to outperform others in deconvolution accuracy.

MethodReDeconvScadenMuSiCBayesPrismTAPEDECODECount BridgesPopLS (ours)
MAE0.06100.05950.11020.07760.05690.05650.07610.0440

The above results have also been added to Table 1(b) (Section 5.3 and Appendix E.3.2).

Thank you again for your feedback, and we look forward to your further comments.

Reviewer_pN4iOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the bias in cell-type proportion estimation caused by shifts in gene expression profiles. The authors propose a method called PopLS, which determines a common correction direction by aggregating the residuals of specific markers across samples from various cohorts. The main contributions of this work include:

  1. Analyzing the geometric structure of the residuals and the associated identifiability issues, while establishing a structural link between the residuals and the correction of cell-type proportions.
  2. Developing a framework for recovering the conditional direction, which ensures the safety of the update process.
  3. Demonstrating, through controlled simulation experiments, pseudo-bulk data covering eight cancer types, and two real-world bulk cohorts, that PopLS achieves superior performance compared to the benchmark methods evaluated.
Critical strengths
  1. The paper provides a clear analysis and description of the relationships among reference offsets, fitting residuals, and proportional errors, noting that the magnitude of the residuals and the direction of the correction must be evaluated separately. This helps readers understand the conditions under which the method holds true and represents a valuable contribution to the theoretical framework.
  2. The experiments include two real-world cohorts for which ground-truth proportions are available via flow cytometry; PopLS achieves low error rates in both. Furthermore, the method is subjected to systematic testing across multiple datasets to demonstrate its advantages.
Critical weaknesses
  1. While I can appreciate that aggregating residuals helps reduce random fluctuations, the stable residual signal—upon which this method relies to determine the correct direction for proportional correction—still requires a foundation of reliable structural connections. 2.The authors identify the limitations of existing methods as being restricted to sample-by-sample inference and failing to fully leverage cohort-level information. However, BayesPrism already updates its reference profiles by aggregating information across samples. While BayesPrism adjusts the expression model, PopLS incorporates a post-fitting correction to the proportions. In my view, both approaches address the same core issue and propose a similar solution; yet, the rationale for why the correction must be applied at a specific stage and why such a correction is necessary in the first place remains insufficiently explained. It remains a significant, unresolved question as to what unique advantages are gained by using population-level residuals to correct the proportions, and whether these advantages are supported by theoretical or empirical evidence.
Questions for authors

1.The authors need to address what unique advantages are gained by using population-level residuals to correct component proportions, and whether there is any theoretical or empirical evidence beyond performance evaluation metrics to support these advantages.

  1. The abstract employs repetitive sentence structures. Furthermore, Section 4 interweaves the algorithm construction process with theoretical results, making the overall logic difficult to follow. Including a framework diagram that outlines the key steps and the general structure of the paper would help improve clarity. Additionally, the heading "Our Results" should likely be changed to simply "Results."

3.In the context of the definitions in Section 4.3, Eqs. (13)–(14), and Corollary C.2, the paper allows AA to be rank-deficient. However, rank(A)<p\text{rank}(A) < p implies ϵrsp≥1\epsilon_{\text{rsp}} \geq 1, which in turn leads to bd≥∥d∗∥b_d \geq \lVert d^{\ast}\rVert. Since the corollary on direction recovery requires bd<∥d∗∥b_d < \lVert d^{\ast}\rVert, its conditions for applicability cannot be met in this scenario.

4.Sections 4.3–4.4 establish recovery guarantees under the condition that the bridge error is small; however, the question of how marker specificity—relative to the reference—ensures a properly calibrated response to unknown reference shifts requires further substantiation. The authors should explicitly state their structural assumptions regarding reference shifts and explain how the observed data support these assumptions.

2 comments
AuthorsOpenReview ↗

Responses to Reviewer pN4i, Part 1/2

We are very grateful for your constructive feedback. Below are point-by-point responses to the weaknesses and questions raised.

W1. Structural connection between residuals and bias directions

The foundation of reliable structural connections was indeed assumed for stable residual signals in our theoretical analysis. This assumption may not hold in practice because of unobservable quantities involved. To mitigate this issue, we revise our analysis as follows; afterwards, we would like to mention our empirical algorithm that was introduced in our first manuscript.

First, we strengthen the structural connection by adding a condition for direction recovery under a shared reference shift (Section 4.3, Eq. (13) and Appendix C.3, Eq. (24)):

\qquad \|N_{\mathcal M}\|_{\mathrm{op}}\le\eta.$$ Here $A$ maps compositions to systematic fraction corrections, and $V_0$ is a response template justified independently for the shift class. The remainder $N_{\mathcal M}=R_{\mathcal M}\Delta_Q-V_0A$ is the part of the shared reference shift visible on the marker genes that $V_0A$ does not capture. The bound $\eta$ limits how much this remainder can affect marker residuals per unit change in composition. Appendix C.3, Eq. (24), bounds the bridge error in terms of $\eta$ and the calibration error $$\kappa=\|DWV_0P_A-P_A\|_{\mathrm{op}}, \qquad P_A=AA^\dagger.$$ Under the stated assumptions, if $2b_d(\delta)+g_i<\|\mathbf d^\star\|_2$, the recovered direction has a positive inner product with the true correction for sample $i$ with high probability. Here $b_d(\delta)$ bounds the population recovery error, and $g_i$ accounts for composition differences and sample-specific coefficient noise (Section 4.4 and Appendix C.6, Corollary C.5). Second, in our actual implementation (Algorithm 1 at the end of Section 4.5), we had introduced a gate function that accept/reject a bias correction proposal for all estimates. Such a gate is computed based on observed quantities and then judges if the actual condition is reliable enough for correction. The gate's thresholds had also been calibrated (Appendix C.8). Please see our detailed response to Reviewer 7RJY, "W1. Gate analysis". **W2. How PopLS differs from BayesPrism** We agree that BayesPrism also uses population information to address reference shift. BayesPrism exploits population likelihood functions to update the shifted reference, which is used for a new round of deconvolution estimates (yet still done individually for each sample). Reference shift can persist after the update. Also, we argue that the updated reference loses interpretation of the deconvolution results, if one cares about the cell type proportions and expression. Our post-correction algorithm explicitly models the shift and uses population post-fit residuals to construct a shared fraction-correction signal. Thus, BayesPrism and our method exploit population information through different but complementary mechanisms: model-based reference adaptation and residual-guided fraction refinement. The above comparison has been added to Related Work. Moreover, our algorithm can post-correct BayesPrism's results. We add an experiment that uses BayesPrism’s final deconvolution estimates as the starting point for population residual correction. The table below compares the results before and after correction, demonstrating a successful combination of BayesPrism and our method (Appendix E.3.4, Table 12). DREAM is new real-bulk dataset added per Reviewer 7RJY's request. | Dataset (N) | Metric | BayesPrism | After population correction | Improvement | |---|---|---:|---:| ---: | | Altman (322) | RMSE ↓ | 0.23734 | 0.16417 | +0.07317 | | | Pearson ↑ | 0.15360 | 0.62494 | +0.47134 | | | CCC ↑ | 0.14497 | 0.58709 | +0.44211 | | Morandini (156) | RMSE ↓ | 0.14518 | 0.12832 | +0.01686 | | | Pearson ↑ | 0.43903 | 0.49296 | +0.05393 | | | CCC ↑ | 0.41644 | 0.48071 | +0.06426 | | DREAM (96) | RMSE ↓ | 0.14282 | 0.12078 | +0.02204 | | | Pearson ↑ | 0.31494 | 0.37907 | +0.06413 | | | CCC ↑ | 0.24259 | 0.31160 | +0.06900 |
AuthorsOpenReview ↗

Responses to Reviewer pN4i, Part 2/2

Q1. Advantages gained by using population residuals

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)MetricInitialIndividualPopulationImprovement
Altman (322)RMSE ↓0.131380.087340.08313+0.00421
Pearson ↑0.789940.913450.92302+0.00957
CCC ↑0.787450.896060.90615+0.01009
Morandini (156)RMSE ↓0.067950.062860.06255+0.00032
Pearson ↑0.791630.824750.82654+0.00179
CCC ↑0.776420.813210.81399+0.00077
DREAM (96)RMSE ↓0.068830.068040.06726+0.00078
Pearson ↑0.480560.474510.50425+0.02974
CCC ↑0.465650.450900.48808+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 TT introduced in Section 4.3. Meanwhile, the further improvement by population-residual correction is relatively small, and this is due to the O((mn)−1/2)O((mn)^{-1/2}) effective-sample-size law of residual error reduction, proven in Theorem 4.3. In our context, mm, the number of bulks, is much smaller than nn, the number of cells, so the increase of mm based on a large nn 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.

Q2. Writing and presentation

Abstract has been cleaned. Section 4 has been renamed “Results”. An overall algorithm flow, Algorithm 1 (an empirical version), has been added at the end of Section 4.

Q3. Rank deficiency issue

We agree that the original definition forces the response error to be at least one when AA is rank-deficient. We revise the definition to measure the error relative to the orthogonal projector onto the range of AA (Section 4.3, Eq. (14)):

=\bigl\|DWV_{\mathcal M}-AA^\dagger\bigr\|_{\mathrm{op}}.$$ The population correction $\mathbf d^\star$ lies in this subspace, so the revised response error can be below one even when $A$ is rank-deficient. Appendix C.3 derives the bridge bound for arbitrary rank. For a nonzero population correction, the sufficient condition for direction recovery remains $$b_d(\delta)<\|\mathbf d^\star\|_2.$$ The main rank-deficient simulation settings satisfy this condition and show strong alignment with the true population correction (Section 4.4, Corollary C.2 in Appendix C.4, and Appendix E.1.5, Figure 4 and Table 3). **Q4. Marker specificity under unknown reference shifts** We agree that markers selected from the reference may lose cell-type specificity in the target samples under an unknown shift. Several reference-based methods rely on reference-derived expression patterns remaining informative in the target mixtures. DWLS (Tsoucas et al., 2019) selects markers by differential expression in the single-cell reference, and ReDeconv (Lu et al., 2025) selects genes that distinguish cell types and show low within-type variability, with the expectation that their expression will be similar in the reference and mixtures. We add an explicit assumption on the shared shift in Appendix C.3 (Eq. (24)). Let $A$ map compositions to systematic fraction corrections and $P_A=AA^\dagger$ project onto its range. For a response template $V_0$ specified in advance or justified independently for the shift class, we assume $$R_{\mathcal{M}}\Delta_Q=V_0A+N_{\mathcal{M}}, \qquad \lVert N_{\mathcal{M}}\rVert_{\mathrm{op}}\le\eta, \qquad \kappa=\lVert DWV_0P_A-P_A\rVert_{\mathrm{op}},$$ with small $\eta$ and $\kappa$. These terms measure how closely the template describes the marker response and how accurately the decoder recovers the correction from that response, respectively. The controlled marker-specificity scan shows stronger alignment with the true correction when markers are more specific and the shared shift is sufficiently strong (Appendix E.2.1, Figure 1(c)). The known-shift simulations also show strong alignment in settings where the bridge error is small (Appendix E.1.5, Figure 4). The practical gate checks agreement across disjoint marker sets and residual enrichment on markers to assess whether the observed bulk data support the proposed correction direction (Appendices C.8 and D). Thank you again for your feedback, and we look forward to your further comments.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.