acceptodds
Under review as a conference paper at ICLR 2027

Weighted Hybrid PCA: Learning Shared Structure under Batch Effects

Abstract

Principal component analysis (PCA) is widely used to extract low-dimensional structure from high-dimensional data. However, batch effects, such as those due to experimental timing, equipment, or data sources, may dominate the leading principal components and obscure the shared variation of interest. When the corresponding labels are only partially observed, removing these effects without discarding useful observations becomes difficult. We model this setting using a factor model with a shared factor loading space and different intercepts across batches. We then propose weighted hybrid PCA, which combines covariance estimates from labeled and unlabeled observations. Using the observed labels, we select a weight from a safe set that excludes additional components caused by between-population variation. Under suitable conditions, the selected weight asymptotically minimizes both the factor detectability threshold and the factor loading space estimation error over the safe set. The lower threshold allows weaker factors to be detected. Simulations show substantially lower factor loading space estimation error for weights inside the safe set than for those outside it. The selected weight achieves the lowest estimation error, close to the oracle benchmark. On two A549 single-cell RNA perturbation datasets with p53-related expression responses, weighted hybrid PCA more efficiently captures the common within-population variation than the other methods and gives the highest average number and proportion of p53-associated principal components.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Principal component analysis, factor model, batch effects, random matrix theory

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_oYJmOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper studies PCA when data come from several batches that share the same factor loading space but have different means, and the batch label is known for only mm of the NN samples. It combines two covariance estimates, Sα=αSlab+(1−α)SmixS_\alpha = \alpha S_{\text{lab}} + (1-\alpha) S_{\text{mix}}, where SlabS_{\text{lab}} is computed from the labelled samples centred within each batch and SmixS_{\text{mix}} from the unlabelled samples centred globally.

Using random matrix theory, the authors define a "safe" set of weights for which the differences between batch means do not create extra spikes outside the noise bulk. They estimate the strength of these differences from the labelled data and choose the safe weight closest to m/Nm/N. Theorems 1–3 show that this choice recovers the correct number of factors, minimizes the detection threshold and the subspace error over the safe set, and matches oracle PCA when m/Nm/N is safe.

The method is validated by simulations (p=400p = 400) and on two sci-Plex A549 single-cell RNA subsets (Nutlin-3a and Pirarubicin).

Critical strengths

S1. The paper addresses a practical problem with careful theory. Batch effects can dominate the leading principal components, and throwing away the unlabelled samples wastes data. The method uses the partial labels to estimate how strong the batch differences are (Eq. 5) and then chooses one weight automatically, without cross-validation. Theorem 2 shows that this weight minimizes the detection threshold over the safe set and matches oracle PCA when m/Nm/N is safe, which directly answers the question the paper asks. I find this a principled and simple rule for a common situation.

S2. The results agree with the theory. In Fig. 1 the error drops sharply at the edge of the safe region, and the selected weight is close to the lowest error and to the oracle. On Nutlin-3a, the natural weight keeps about 37 components against about 10 for the proposed weight, which fits the explanation of spurious batch directions. The proposed method also has the highest held-out variance efficiency on both datasets (0.505 against 0.431 and 0.339 on Nutlin-3a; 0.450 against 0.372 and 0.277 on Pirarubicin). Code and data are provided.

Critical weaknesses

W1. The experiments are too narrow to support the practical claims.

The simulations use a single configuration: four equal batches, three factors with fixed strengths (5,4,3)(5, 4, 3) along the first coordinates (U=(e1,e2,e3)U = (e_1, e_2, e_3)), p=400p = 400, and m≥pm \ge p in every case. I would like to see unequal batches, small mm, more or weaker factors, heteroscedastic noise and some misspecification. In the simulations only different weights of the same method and the oracle are compared, with no outside method.

On the real data, only one competitor from the literature is used (mean-shift PCA, Li et al. 2026), and both datasets come from a single cell line and experiment. Simpler ways of using the labels are not compared either, for example projecting out the estimated batch-mean directions, or predicting the missing labels and centring within the predicted batches. The appendix contains proofs but no additional experiments.

Without broader evidence, I find it hard to judge whether the method is useful beyond the tested setting.

W2. Some key formulas are not explained in the main text, and the assumptions do not fit the application well.

At L142–143 a weight is called safe when (1−α)γμ≤tα(1-\alpha)\gamma_\mu \le t_\alpha, but the reason is not given there. tαt_\alpha only gets its meaning later (L222–225), as the level above which a spike from the unlabelled part separates from the noise bulk. Similarly, the estimator of γμ\gamma_\mu in Eq. (5), the edge equation (6), the bulk-edge estimate (9) and the component rule (10) are given without derivation; their justification appears only inside the appendix proofs (Lemma 1) and at L199–201. Algorithm 1 also appears (L147) before the quantities it uses are defined.

On the application side, the theory assumes noise with equal variance, while log-normalized scRNA-seq data are strongly heteroscedastic. The asymptotics assume many labelled samples, but the real data use about 20 labelled cells (3–5 per batch) with thousands of genes. The theory reveals labels at random, while the experiments reveal 10% within each batch (L362).

As a result, readers cannot easily follow or check the method, and the real-data results may not reflect the conditions the theory needs.

Questions for authors

Q1. Why is safety defined as (1−α)γμ≤tα(1-\alpha)\gamma_\mu \le t_\alpha? Could you give the intuition in Section 2?

Q2. How are Eqs. (5), (6), (9) and (10) obtained? Please explain them briefly or point to where each is justified.

Q3. This question is very important for my assessment. Do you claim to be the first to use a factor model with a shared loading space and batch-specific intercepts? If not, why are no other methods compared in the simulations?

Q4. Apart from mean-shift PCA (Li et al. 2026), are there other methods that could be compared on the real data?

Q5. This is a suggestion, and the answer could change my assessment. How does the method compare with projecting out the estimated batch-mean directions, or with predicting the missing labels and then centring within batches?

Q6. For the revision, could you add more simulation settings (unequal batches, small mm, more or weaker factors, heteroscedastic noise) and at least one more real dataset, even in the appendix?

Minor: in Table 1 the methods are compared at very different numbers of components (about 3 to 37). The p53 test uses nominal p<0.05p < 0.05 without multiple-testing correction. The error formula of Theorem 3 is not checked by any simulation.

6 comments
AuthorsOpenReview ↗

Author response (Part 1/2)

Thank you for your careful reading and constructive suggestions. Below, we address the comments and describe the corresponding revisions, including additional experiments.

W1: Experiments too narrow to support practical claims

Thank you for your suggestions on the experiments. We have expanded the simulations and added an independent real dataset. These additions are described in our response to Q6. Our response to Q4 and Q5 explains the choice of baselines and describes the additional baselines included in the revised experiments.

W2: Explanation of formulas and dependence on assumptions

Thank you for suggesting these clarifications. Our intention was to explain the population quantities and their roles in the algorithm before presenting the estimation formulas. Algorithm 1 provides an overview of the procedure. We have expanded these explanations in Section 2 (lines 153–158 of the revised manuscript). We have also added references to the relevant theory and proofs alongside the corresponding formulas (lines 175, 185, 199, and 205–206 of the revised manuscript).

We have added simulations with heteroscedastic noise to assess robustness beyond the equal-variance assumption, as described in Q6.

In the real-data experiments, we select labeled observations randomly within each batch to ensure that all batches are represented. Conditional on population memberships, labeling is independent of factor scores and noise. With the same labeling proportion across batches, the labeled and unlabeled samples retain the same limiting mixing proportions. These properties allow the conditional arguments in our proofs to extend to this labeling scheme.

We assess performance at labeling proportions of 5%–15% through the additional simulations described in our response to Q6. This range includes the 10% labeling proportion used in our single-cell RNA experiments.

Q1: Intuition for the safety condition

Selecting the number of PCs using the spectral edge, as in Eq. (10), is an established approach (Kritchman and Nadler, 2008). In the limit, the spectral edge gives the upper bound for eigenvalues generated by noise alone. In contrast, detectable factors produce eigenvalues above this edge.

However, between-population mean differences can also produce eigenvalues above the spectral edge. Here, γ^μ\widehat\gamma_\mu estimates the strength of these mean differences. After weighting, their estimated strength becomes (1−α)γ^μ(1-\alpha)\widehat\gamma_\mu. Our theory gives a threshold for this contribution, estimated by t^α\widehat t_\alpha. Eq. (7) defines a weight as safe when (1−α)γ^μ≤t^α(1-\alpha)\widehat\gamma_\mu\le\widehat t_\alpha. This prevents mean differences from generating additional eigenvalues above the spectral edge in the limit.

We have clarified the role of this threshold and the safety condition in Section 2 (lines 153–158 of the revised manuscript).

Q2: Derivation of Eq. (5), (6), (9), and (10)

We briefly explain the derivation and justification of each formula in turn.

Eq. (5) estimates the strength of the between-population mean differences. We estimate the population means and mixing proportions from the labeled observations. The correction term accounts for noise in the estimated means. Lemma 1 establishes the consistency of the resulting estimator.

We solve Eq. (6) for s^α\widehat s_\alpha. It estimates sαs_\alpha, the quantity used to calculate the spectral edge. The population equation for sαs_\alpha follows from Eq. (11). We set the derivative of its right-hand side to zero at the spectral edge. We then replace the parameters with their sample counterparts to obtain Eq. (6).

Eq. (9) estimates the spectral edge. We obtain it by substituting s^α\widehat s_\alpha into the sample counterpart of Eq. (11).

Eq. (10) applies the spectral edge criterion discussed in Q1. The added 1/log⁡p1/\log p term accounts for fluctuations around the estimated edge. Theorem 1 provides the theoretical justification for this selection rule.

Continued in Part 2.

AuthorsOpenReview ↗

Author response (Part 2/2)

Continuation of Part 1.

Q6: Additional simulations and an independent real dataset

Thank you for these suggestions. We have added simulations with unequal population sizes, labeling proportions of 5%–15%, more and weaker factors, and heteroscedastic noise. These labeling proportions correspond to 40–120 labeled observations across 10 populations. We also include mean-shift PCA (Li et al., 2026) and PCA after mean projection (Haghighatkhah et al., 2022) as additional baselines in these simulations. Weighted hybrid PCA achieves the lowest average factor loading space estimation error in all reported settings (Appendix B.2, Tables 5–6, lines 1514–1557 of the revised manuscript).

We have also added an experiment on the Quartet RNA dataset (Yu et al., 2024). This dataset contains measurements of four biological reference materials across 11 batches. We use partially observed batch labels for estimation and material identities only for evaluation. We use the SNR metric introduced by Yu et al. (2024) to assess material separation after projection onto the PCs selected by each method. Weighted hybrid PCA achieves the highest average SNR among the compared methods, close to that of fully labeled oracle PCA (Section 4.3 and Table 2, lines 459–495 of the revised manuscript).

Q3: Novelty and relation to existing factor models

Related factor models use fully observed population memberships to estimate shared factor loadings while modeling intercept differences (De Roover, 2021) or population-specific factors (De Vito et al., 2019). McNicholas (2010) also studies related factor models for semi-supervised classification, with the aim of predicting unknown population memberships. We study semi-supervised PCA for estimating the shared factor loading space, treating population-specific intercept differences as nuisance variation. We have clarified these distinctions in the Introduction (lines 112–118 of the revised manuscript).

Consider the Quartet RNA example (Yu et al., 2024) discussed in Q6. Differences between materials represent biological information of interest, whereas differences between batches are nuisance effects. In this example, classification methods such as McNicholas (2010) model biological differences between materials for classification. Our method models differences between batches to extract biological information despite these nuisance effects. We therefore do not include these classification methods as direct baselines for evaluating shared factor loading space estimation.

Q4 and Q5: Additional baselines

Thank you for suggesting these additional baseline methods. As explained in Q3, those factor models and semi-supervised classification methods are not suitable direct baselines for our setting. Mean-shift PCA (Li et al., 2026) treats between-population mean differences as nuisance variation. It aims to recover the underlying covariance structure. This makes it a suitable baseline for our problem. We have also included PCA after mean projection (Haghighatkhah et al., 2022) as a baseline in the revised simulations and real-data experiments.

We have not included prediction of missing batch labels followed by within-batch centering. In our experiments, simple classifiers produced highly unstable predictions of these labels, especially at low labeling proportions. Addressing this instability would require further methodological work beyond a straightforward baseline implementation.

Minor comments

Following Reviewer 1’s suggestion, we revised the noise-scale estimator as permitted by our theory. This leads to more conservative PC selection, especially under the natural weight. We have updated the experiments. Across methods, the mean PC counts are 1.42–5.95 for Nutlin-3a and 1.14–4.17 for Pirarubicin (Table 1, lines 433–446 of the revised manuscript). Weighted hybrid PCA retains its performance advantage on both datasets. As expected, excluding additional PCs caused by between-population mean differences leads to fewer selected PCs.

For the p53 association tests, we have adopted Benjamini–Hochberg-adjusted p-values to account for multiple testing (lines 1568–1570 of the revised manuscript).

We have added simulations comparing the empirical error with the theoretical value in Theorem 3 (Section 4.1 and Figure 2, lines 371–393 of the revised manuscript). Figure 2 shows that the empirical error approaches the theoretical value as the sample size increases.

Reviewer_oYJmOpenReview ↗

Thank you for the revision; three short follow-up questions

Thank you for the detailed response and the substantial revision. I checked the revised paper. The new simulations in Appendix B.2 (5–15% labels, unequal populations, more and weaker factors, heteroscedastic noise, and the mean-shift and mean-projection baselines), the Quartet RNA experiment with an oracle reference (Table 2), the check of Theorem 3 (Figure 2), the paired comparisons (Tables 3, 4 and 7), the Benjamini–Hochberg correction, and the new explanations in Section 2 and in the Introduction address my main concerns. I am inclined to raise my rating, and I would first like to see your answers to the three short questions below.

Q1. The noise estimator σ^2=tr(Smix)/p\hat\sigma^2 = \mathrm{tr}(S_{mix})/p also includes the variance of the factors and of the batch-mean differences, so it overestimates σ2\sigma^2 for finite pp. How large is this bias on the single-cell data, for example compared with the labeled estimator corrected for the lost degrees of freedom? It may explain why the selected numbers of components dropped so much (for example to 1.42 for labeled PCA on Nutlin-3a).

Q2. In Table 1 all methods find about one p53-associated component, and Table 7 shows a paired difference in count of only 0.09 and 0.05. The higher proportion therefore comes mainly from selecting fewer components. I suggest describing the p53 result in this way, rather than as more p53 signal being recovered. Similarly, on the Quartet data the SNR of weighted hybrid PCA (35.6 ± 2.7 dB) is close to that of labeled PCA (34.6 ± 2.1 dB), while the gain in distance to the oracle space is clear.

Q3. You mention that predicting the missing labels and then centring within the predicted batches gave very unstable results. Reporting these results briefly in the appendix would show why this simple approach is not enough.

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

This paper deals with the problem of extracting low-dimensional representations of examples modeled by a Gaussian mixture model (GMM) and proposes a semi-supervised method named weighted hybrid PCA. More specifically, the proposed method mixes covariance matrices extracted from supervised and unsupervised examples and carefully selects the mixture weight and the number of components to be extracted through theoretical analysis. Experimental evaluations with simple real datasets demonstrate that the proposed method performs better than simple baselines and the most closely related method, named mean-shift PCA.

Critical strengths

S1. The procedure of the proposed method is simple, as presented in Algorithm 1. Simplicity is invaluable, especially for engineers interested in this problem.

S2. Through theoretical analysis, the current paper demonstrated that, provided the data originate from a Gaussian mixture and certain assumptions are satisfied, the proposed method can extract low-dimensional representations modeled by the underlying distribution with sufficient accuracy.

Critical weaknesses

W1. The proposed method can be regarded as a special case of within-class (or within-group) PCA. We can find several papers on semi-supervised within-class PCA, like [Cai+ ICCV2007, Semi-Supervised Discriminant Analysis] or [Sugiyama+ Machine Learning 2010, Semi-Supervised Local Fisher Discriminant Analysis for Dimensionality Reduction]. I understand that the proposed method is not a complete copy-and-paste of the above methods and can automatically determine the number of components to be extracted and the weight for mixing supervised and unsupervised covariance matrices. However, the current paper does not appropriately justify the novelty of the proposed method among the vast array of existing methods.

W2. The simulation-based experiments presented in Section 4.1 should be revised. In particular, the current experimental setting prepares so many labeled examples, which is not so standard for semi-supervised learning. I understand that semi-supervised learning based on eigenproblems requires more examples than standard machine learning techniques, and the lack of supervised examples can have a devastating impact on the performance. However, semi-supervised learning with many supervised examples is not realistic.

Questions for authors

None

[Post-rebuttal comments]

I appreciate the author feedback. As a result, all concerns raised in the initial review have already been addressed in the revised paper. Although the current status might not be sufficient to increase the rating, I do not fight against the acceptance of this paper.

1 comment
AuthorsOpenReview ↗

Author response

Thank you for your careful reading and constructive suggestions. Below, we address the comments and describe the corresponding revisions, including additional experiments.

W1: Novelty and relation to existing methods

Thank you for pointing us to these related works. Cai et al. (2007) extend discriminant analysis by incorporating geometric information from unlabeled data. Sugiyama et al. (2010) combine local Fisher discriminant analysis with PCA. Both methods treat between-class differences as useful information and seek projections that preserve class separability. McNicholas (2010) also studies related factor models for semi-supervised classification, with the aim of predicting unknown population memberships. Our objective is to estimate the shared factor loading space, treating population-specific intercept differences as nuisance variation. These differences can generate additional principal components and interfere with estimation of the shared factor loading space. Related factor models use fully observed population memberships to estimate shared factor loadings while modeling intercept differences (De Roover, 2021) or population-specific factors (De Vito et al., 2019). We have clarified these distinctions in the Introduction (lines 112–118 of the revised manuscript).

Consider the Quartet RNA example (Yu et al., 2024), where four biological reference materials are measured across 11 batches. Differences between materials represent biological information of interest, whereas differences between batches are nuisance effects. In this example, the classification and discriminant methods discussed above model biological differences between materials for classification. Our method models differences between batches to extract biological information despite these nuisance effects. We have added an experiment on this dataset showing that weighted hybrid PCA better preserves separation between materials (Section 4.3 and Table 2, lines 459–495 of the revised manuscript).

W2: Simulations with fewer labeled observations

Thank you for this suggestion. We have added simulations with labeling proportions of 5%–15%, corresponding to 40–120 labels across 10 populations. These settings include unequal population sizes, more and weaker factors, and heteroscedastic noise. Weighted hybrid PCA achieves the lowest average factor loading space estimation error in all reported settings (Appendix B.2, Tables 5–6, lines 1514–1557 of the revised manuscript).

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

The paper studies shared factor-space estimation when batch means differ and only some batch labels are observed. Weighted hybrid PCA combines labeled and unlabeled covariance estimates, constructs an estimated safe set of weights, and selects the weight closest to the natural sample proportion. The paper provides asymptotic guarantees and evaluates the method on simulations and two single-cell datasets. The main contribution is the theoretical justification for deciding how much information to use from unlabeled observations. The safe-set analysis gives a clear rationale for the selection rule, and the estimator is straightforward to implement. I also appreciate the complete proofs and the use of held-out cells with training-only gene selection.

Critical strengths

The estimated safe set turns partial batch labels into a concrete weighting rule. Under the stated conditions, the selected weight minimizes the detection threshold and subspace error over that set, and can match oracle PCA’s asymptotic error. These results are the main reason I lean toward acceptance.

Critical weaknesses

My main concern is calibration with approximately 20 labels across six or four batches. Under the model, batch centering introduces a factor of (m−K)/m, approximately 0.70 or 0.80, into the noise estimator’s expectation. This estimate affects both weighting and component selection. The simulations use 400–800 labels and do not test this setting. I would like to see whether the HVE gains persist after accounting for the lost degrees of freedom.

Questions for authors

1.I am most concerned about calibration with few labels. The real-data experiments distribute approximately 20 labels across six or four batches, whereas simulations use 400–800 labels. With all batches represented, the noise estimator has conditional expectation (1−K/m)tr(Σ)/p. The centering factor is therefore approximately 0.70 or 0.80 here. Since this estimate affects both weight selection and the component cutoff, please provide feature dimensions and per-batch label counts. Do the HVE gains and selected component counts remain stable after accounting for the centering degrees of freedom?

2.I could not tell whether the stars in Figure 1 show errors obtained using each replication’s selected weight. The caption identifies median selected weights, while the curves average errors at fixed weights. Reading the curve at the median weight need not give the procedure’s average error. Please report errors at the individually selected weights and their paired differences from the baselines using the existing runs.

3.Remark 3 also needs clarification. Without asymptotic orthogonality, minimizing ϑα need not minimize the actual signal strength required for detection. Mean shifts along a true factor direction can help reveal that factor without creating another direction. Working through this case using K(z) would explain the guarantee more concretely. The general result remains useful when interpreted as optimizing a sufficient detection bound. 4.For Nutlin-3a, natural and weighted hybrid PCA retain approximately 37 and 10 components. Although HVE accounts for dimension in its denominator, comparisons at common component counts would help establish whether the estimated directions themselves improve. Fully labeled within-batch PCA would also be a useful reference. I am less persuaded by the p53 comparison, particularly the Pirarubicin averages of 1.49 and 1.45. Without paired uncertainty estimates and multiple-testing control, I would treat these counts as exploratory support.

1 comment
AuthorsOpenReview ↗

Author response

Thank you for your careful reading and constructive suggestions. Below, we address the comments and describe the corresponding revisions, including additional experiments.

W1 and Q1: Calibration with few labels

Thank you for your suggestion on noise-scale estimation with few labels. When few labels are available, it is preferable to estimate σ2\sigma^2 from the larger unlabeled sample using σ^2=tr⁡(Smix)/p\widehat{\sigma}^2=\operatorname{tr}(S_{\rm mix})/p. Our theory permits this choice. We have clarified this choice in lines 175–177 of the revised manuscript. We have also rerun the experiments using this estimator.

The feature dimensions are p=1920p=1920 for Nutlin-3a and p=712p=712 for Pirarubicin, with labeled batch counts (5,4,5,2,2,2)(5,4,5,2,2,2) and (6,5,6,3)(6,5,6,3), respectively. These details are now provided in lines 407–413 of the revised manuscript. With this estimator, the selected PC counts decrease, but weighted hybrid PCA retains its performance advantage on both datasets. Table 1 reports the updated results as mean ±\pm standard deviation over 100 repetitions.

We have also added simulations with labeling proportions of 5%–15%, corresponding to 40–120 labels across 10 populations. These settings include unequal population sizes, more and weaker factors, and heteroscedastic noise. Weighted hybrid PCA achieves the lowest mean factor loading space estimation error in all reported settings (Appendix B.2, Tables 5–6, lines 1514–1557 of the revised manuscript).

Q2: Performance at selected weights

The stars in Figure 1 mark the median selected weights and do not show the mean error at individually selected weights or the variation in weight selection across repetitions. Table 3 reports mean errors at individually selected weights and paired differences from oracle PCA, based on the same 100 repetitions. It also reports absolute deviations of the selected weights from their median within each setting (lines 1467–1484 of the revised manuscript). Table 4 (lines 1487–1505 of the revised manuscript) reports paired differences from natural hybrid PCA and labeled PCA, including means, standard deviations, and 95% confidence intervals for the mean differences.

Q3: Interpretation of the detection bound

Thank you for suggesting this clarification. Without asymptotic orthogonality, the selected weight asymptotically minimizes a sufficient detection bound over the estimated safe set. We have clarified this point in Remark 3 (lines 298–302 of the revised manuscript).

Q4: HVE gains and reliability of p53 comparisons

Higher HVE alone does not imply more accurate estimation of individual factor directions. Our goal is to improve estimation of the shared factor loading space by excluding additional PCs caused by between-population mean differences. Selecting the number of PCs using the spectral edge is an established approach (Kritchman and Nadler, 2008). This criterion alone cannot exclude these additional components. A core contribution of our work is to derive the spectral edge for the weighted model and combine it with safe weight selection. Comparisons at a common component count would therefore not fully evaluate our method.

Moreover, the component counts selected by natural and weighted hybrid PCA are closer with the estimator described in Q1. The mean counts differ by approximately three on Nutlin-3a and one on Pirarubicin (Table 1, lines 433–446 of the revised manuscript).

Similar numbers of p53-associated PCs are expected because removing additional PCs does not introduce new factors. The improvement is reflected in the increased proportion of p53-associated PCs. For the p53 association tests, we have adopted Benjamini–Hochberg-adjusted p-values to account for multiple testing (lines 1568–1570 of the revised manuscript). We have also added paired comparisons of p53-associated PC counts and proportions. Table 7 in Appendix B.3 reports the mean differences and their 95% confidence intervals (lines 1572–1579 of the revised manuscript).

To further evaluate our method on real data, we have added an experiment on the Quartet RNA dataset using measurements of four reference materials across 11 batches. Fully labeled within-batch PCA serves as an oracle reference. We report selected PC counts and squared projection distance to the oracle space. We also assess the separation of reference materials using the SNR metric introduced in the original Quartet study (Yu et al., 2024). Details are provided in Section 4.3 and Table 2 (lines 459–495 of the revised manuscript). Weighted hybrid PCA achieves the smallest mean distance to the oracle space and the highest mean SNR among the compared methods. Its SNR is close to that of oracle PCA.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.