acceptodds
Under review as a conference paper at ICLR 2027

Molecular Uncertainty Gradients: Exact Decomposition of Predictive Variance Onto Molecular Substructures

Abstract

Most explainability methods for molecular property prediction target predictions rather than uncertainty. Techniques for localizing predictive uncertainty to specific substructures remain comparatively underdeveloped, limiting our ability to identify which substructures drive uncertainty and which chemical patterns the model has not learned reliably. We present Molecular Uncertainty Gradients (MUG), a method that distributes a model's epistemic uncertainty, estimated as the variance across ensemble predictions, over the atoms of a molecule. By differentiating this variance with respect to intermediate atom representations, MUG computes signed atom-level contributions that sum exactly to the molecule’s predictive variance. These attributions can be aggregated over any group of atoms using a single backward pass. MUG requires no retraining and works with the dropout-based uncertainty estimates already standard in practice. We show that MUG localizes uncertainty to molecular substructures absent from the training data. Across six molecular property prediction benchmarks, MUG achieves the highest occlusion-controlled fidelity, and its decomposition holds. Crucially, masking the features of its highest-ranked atoms reduces model uncertainty substantially more than masking other fragments. We further demonstrate that MUG’s fragment-level uncertainty attributions can guide iterative library design toward compounds with favorable docking scores. Together, these results establish MUG as a complete and faithful approach to localizing predictive uncertainty, with practical applications in uncertainty-guided molecular design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
uncertainty quantification, explainable AI, ensemble methods, molecular property prediction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_d3xZOpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper presents MUG, a method to attribute uncertainty to specific molecular substructures. The paper provides a concrete motivation and cover six benchmarks as well as a molecular design task.

Critical strengths

The paper presents a specific and reasonable motivation. The problem of identifying which molecular substructures are related with predictive uncertainty is important in many domains.

Critical weaknesses

The novelty is limited. The method mostly takes existing gradient-based attribution methods and applies them to predictive variance.

Questions for authors

What is the main novelty of MUG compared with existing gradient-based methods?

How well does the exact decomposition work when the model assumptions are not satisfied?

Does exact decomposition necessarily mean accurate uncertainty attribution?

Can the authors provide more experiments and comparisons to support their claims?

2 comments
AuthorsOpenReview ↗

Response to Reviewer d3xZ (1/2)

We thank the reviewer for the questions on novelty and on the exact decomposition. We address each of them below, and all changes are highlighted in blue in the revised paper.

W1/Q1: What is the main novelty of MUG?

We thank the Reviewer for raising this question about MUG’s novelty. As Reviewer 4uic helpfully pointed out, Bley et al. (2025) provide an elegant and general framework for explaining predictive uncertainty through second-order effects. Their work clearly shows the value of investigating gradient methods for uncertainty attribution and offers an important reference point for discussing MUG. We have clarified this connection in the revised manuscript (Section 2; Appendix E). Please refer to our response to Reviewer 4uic for a detailed comparison between MUG and Bley et al. (2025).

In this context, MUG contributes an efficient formulation for molecular GNNs that obtains atom-level uncertainty attributions by directly differentiating predictive variance, without separately explaining each ensemble member’s prediction. We focus on final-layer atom representations, where the stated readout and head assumptions allow a signed, exact decomposition of predictive variance through a single backward pass, without retraining (Section 3.2, Theorem 1). Atom contributions can then be summed directly over any molecular fragment (Section 3.2). Our experiments demonstrate both the chemical relevance and practical usefulness of these attributions. Controlled fluorine/sulfone withholding experiments show that MUG localizes uncertainty to chemical substructures absent from training (Section 4.1, Figure 3b; Appendix H.4, Figure 8). In molecular design, these attributions guide fragment-level acquisition, increasing the docking hit rate from 47.2% to 56.2% compared with random fragment selection within the same acquisition framework (Section 5, Table 6). In the rebuttal version, we have also clarified the relationship to Integrated Gradients (Section 3.2) and added an input-space comparison targeting the same predictive variance (Appendix H.2, Table 16).

Additionally, we note that novelty of MUG lies also in applicability of this method in chemistry and drug design, where many major architectures (such as GNNs) can be easily adapted to use MUG without architectural changes. It also provides more interpretability for chemists operating on functional groups rather than single atoms (the completeness property, please also see our response to W1 of Reviewer wtzS).

Bley, F., Lapuschkin, S., Samek, W., & Montavon, G. (2025). Explaining Predictive Uncertainty by Exposing Second-Order Effects. Pattern Recognition, 160, 111171.

AuthorsOpenReview ↗

Response to Reviewer d3xZ (2/2)

Q2: How well does the exact decomposition work when the model assumptions are not satisfied?

The decomposition is no longer guaranteed to be exact when the assumptions are violated. As reported in our original submission (Appendix F.7; now Appendix H.8, Table 28 in the revised manuscript), completeness residuals are approximately 10−610^{-6}–10−510^{-5} for configurations satisfying the assumptions, compared with 0.049 for independently biased ensemble heads and 0.196 for input-level attribution. However, loss of completeness does not necessarily imply poor attribution rankings: a biased ReLU head yields a residual of 0.066 while retaining positive controlled fidelity (ΔFid=0.305\Delta\mathrm{Fid}=0.305). Attention pooling, by contrast, loses both: a residual of 0.296 and ΔFid=0.022\Delta\mathrm{Fid}=0.022.

Q3: Does exact decomposition necessarily mean accurate uncertainty attribution?

No, that is a separate but useful quality of attribution. Exact decomposition guarantees that contributions sum to predictive variance under the stated assumptions, but does not guarantee attribution accuracy (Section 3.2, Theorem 1). As shown in our paper, independent evaluations provide evidence of attribution quality and practical usefulness:

  • Controlled fidelity: MUG achieves the highest mean controlled fidelity among evaluated methods across six benchmarks, supporting attribution faithfulness (Section 4.2, Table 1).
  • Chemical withholding: The attribution share assigned directly to fluorine atoms increases from approximately 8% to 42%, with similar localization for withheld sulfone groups, supporting sensitivity to missing chemical training coverage (Section 4.1, Figure 3b; Appendix H.4, Figure 8).
  • Molecular design: MUG-guided fragment selection increases the docking hit rate from 47.2% to 56.2% compared with random fragment selection within the same framework (Section 5, Table 6).

Q4. Can the authors provide more experiments and comparisons to support their claims?

We would be grateful for more specific feedback on which claims or aspects of the experimental evaluation remain concerning, and which additional experiments or comparisons would help address these concerns. In response to Reviewer wtzS’s comments, we have added an input-space comparison with Integrated Gradients targeting the same predictive variance in the rebuttal version (Appendix H.2). Below, we summarize the experimental evidence already included in our original submission, with references to the revised manuscript:

  • Baseline comparisons: MUG achieves the highest mean controlled fidelity among evaluated methods across six benchmarks, supporting attribution faithfulness (Section 4.2, Table 1).
  • Chemical withholding: The attribution share assigned directly to fluorine atoms increases from approximately 8% to 42%, with similar localization for withheld sulfone groups, supporting sensitivity to missing chemical training coverage (Section 4.1, Figure 3b; Appendix H.4, Figure 8).
  • Prediction versus uncertainty attribution: MUG achieves positive uncertainty fidelity, whereas prediction-based Grad×Input yields negative uncertainty fidelity (Section 4.3, Table 3).
  • Architectures and ablations: configurations satisfying the assumptions achieve completeness residuals of approximately 10−610^{-6}–10−510^{-5}; ablations illustrate where the guarantee no longer holds (Section 4.5; Appendix H.8, Table 28).
  • Molecular design: MUG-guided fragment selection increases the docking hit rate from 47.2% to 56.2% compared with random fragment selection within the same framework (Section 5, Table 6).

We hope our responses and revisions address the Reviewer’s concerns. We would be happy to clarify any remaining questions during the discussion period.

Reviewer_wtzSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This works proposes MUG for attributing predictive uncertainty to atoms and molecular substructures. By applying the method of gradient-based attribution to ensemble predictive variance, the proposed MUG gives an exact variance decomposition under specific architectural assumptions. Some experiments evaluate its attribution fidelity and ability to identify unseen substructures, while a simulated molecular design study shows its potential for a uncertainty-guided fragment selection.

Critical strengths
  1. The work solves the explainability of molecular predictive uncertainty rather than predictions alone, showing a valuable perspective on understanding model uncertainty across different chemical structures.

  2. MUG based on existing gradient-based attribution methods and can be applied to MC Dropout uncertainty estimates without modifying the model architecture.

  3. The method is evaluated on 6 molecular property prediction tasks, with additional experiments on unseen substructures and occlusion fidelity.

Critical weaknesses
  1. The methodological novelty is relatively limited. MUG essentially applies the existing Gradient × Input approach to predictive variance, and its completeness property follows mainly from the homogeneity of predictive variance rather than a new attribution mechanism.

  2. The paper emphasizes Exact Completeness but does not compare MUG with methods such as Integrated Gradients applied to the same predictive variance target.

  3. The completeness guarantee relies on specific assumptions about the readout and prediction head. Table 5 shows that it no longer holds for R-MAT with attention pooling, which also yields negative occlusion fidelity.

Questions for authors
  1. Table 1 shows that MUG achieves almost identical controlled fidelity to Saliency applied directly to predictive variance (0.229 vs. 0.228 for regression and 0.172 vs. 0.168 for classification). Saliency even slightly outperforms MUG in top-ranked fidelity and selectivity rate on classification tasks. The advantage of MUG seems marginal. Could the authors provide further insights into what MUG offers beyond standard Saliency?

  2. I am not fully convinced why Completeness should be considered an important contribution to improving explanation quality. I could not find sufficient empirical evidence showing that the theoretical completeness guarantee actually improves uncertainty attribution.

  3. MUG primarily explains epistemic uncertainty, whereas the Yang & Li baseline in Table 1 uses aleatoric uncertainty attribution. I am curious about why methods targeting different types of uncertainty are directly compared?

  4. In Section 3.2, the attributions are computed from the final-layer GNN representations, which already aggregate information from neighboring nodes. How do the authors distinguish contributions from these learned representations from the actual chemical contributions of individual atoms?

  5. Some symbols are not clearly or consistently defined. For example, in Section 3.2, mm is defined as an atom index, but Eq. (5) refers to "fragment mm". I found this confusing when reading the manuscript.

6 comments
AuthorsOpenReview ↗

Response to Reviewer wtzS (1/3)

We thank the Reviewer for their thoughtful comments, which helped us improve the paper and clarify its contribution. We respond to each comment below. All changes are highlighted in blue in the revised paper.

W1: Novelty.

We agree with the Reviewer's description: MUG is Gradient × Input applied to the predictive variance, and its completeness follows from the homogeneity of that variance. We clarify that the novelty of MUG does not lie in new architectural elements but in the observation that predictive uncertainty gradients can provide valuable insight in chemistry and that this particular combination of methods gives us useful properties with practical implications for fields like drug discovery.

What is new. We apply gradient methods to predictive variance, not the prediction, which is not an obvious choice. Similar approaches have been studied to some extent in chemistry (Related Work, Section 2) and, as Reviewer 4uic notes, in general explainability literature (Bley et al. 2025). Because that work is the closer of the two, please see our response to Reviewer 4uic for the full comparison. Briefly, we focus on the final layer of GNNs, where each unit already encodes the atomic environment and not isolated atoms. We find that this setup gives MUG properties which typically break in deep architectures: the variance is positively homogeneous of degree two in those activations, so a single differentiation is exactly complete (Theorem 1). As a consequence, a single backward pass gives attributions equal, up to the constant factor T/2T/2, to zero-baseline IG at that layer (see W2 below).

Theoretical properties supporting applicability in chemistry. We observe that many major architectures used in chemistry, such as GNNs, can easily be adapted to use MUG by using dropout layers, and global pooling mechanisms and predictive heads are usually already homogeneous (A1, A2 and A5, Section 3.3). In Section 3.3 and Section 4.5 we provide examples of what models may break this condition. This gives us simple tools for inspecting uncertainty without introducing new architectures. We show that when these conditions are fulfilled, uncertainty can be analyzed on the fragment level (thanks to the completeness property, Theorem 1 and Eq. 6), which is useful for chemists analyzing compounds in terms of functional groups. We also show that zero attribution is a meaningful baseline: when the model is certain, every atom’s attribution is exactly zero (Corollary 1), while a negative value means the atom reduces ensemble disagreement (Section 3.2), which makes the outputs more intuitive to read. The signs are also necessary: rectifying the shares, as Grad-CAM does, forfeits the exact decomposition (Remark 1, Appendix D in the revised paper).

W2: Comparison with Integrated Gradients on Predictive Variance.

We agree that Integrated Gradients applied to the same predictive variance is an essential comparison. We implemented IG at the same final-layer representations used by MUG. Under our readout and prediction-head assumptions, predictive variance V(H)V(H) is homogeneous of degree two in the joint final-layer activations HH. Consequently,

= H \odot \int_0^1 \nabla V(\alpha H)\,\mathrm{d}\alpha = \frac{1}{2}\,H \odot \nabla V(H).$$ Aggregating over MC realizations and feature channels yields the variance-rescaled MUG atom-level scores, $\frac{T}{2}E_m^{\sigma^2}$, whose sum equals the predictive variance. We verified this equivalence on all test molecules across six datasets and observed a maximum absolute attribution difference below $6.4 \times 10^{-8}$ (Appendix H.2). This identity allows final-layer IG to be computed exactly with a single gradient evaluation. Thus, under the stated assumptions, variance-rescaled MUG provides a closed-form evaluation of zero-baseline IG for predictive variance and satisfies completeness at the final-layer representation level (Section 3.2). We additionally compared the methods in input space across six datasets. $MUG_{input}$ achieves higher classification $\Delta Fid$ than $IG_{input}$ (0.418 vs. 0.376), whereas $IG_{input}$ achieves higher regression $\Delta Fid$ (0.398 vs. 0.380). These results show complementary empirical strengths rather than consistent superiority of either method (Appendix H.2).
AuthorsOpenReview ↗

Response to Reviewer wtzS (2/3)

W3: Architectural Assumptions and Attribution Fidelity Beyond the Guarantee.

We agree that exact completeness is conditional on the readout and prediction head. We explicitly characterize these conditions in Section 4.5, “Scope and Limits of the Completeness Guarantee,” and include R-MAT to examine the method’s behavior outside this scope. Importantly, the homogeneity assumptions apply only to the mapping from the attributed final-layer atom representations to the output. The encoder itself need not be homogeneous and may contain nonlinear operations or attention. The guarantee covers mean, sum, and max pooling followed by a linear head with shared bias or a bias-free ReLU MLP. Our evaluated GCN, GIN, and D-MPNN configurations satisfy these conditions and achieve completeness residuals of approximately 10−610^{-6} - 10−510^{-5} (Table 5; Appendix H.8).

We acknowledge R-MAT’s negative controlled fidelity as an empirical limitation. To assess whether this behavior is specific to MUG, we additionally compared MUG, input Saliency, and final-layer Integrated Gradients on the complete Caco2 and BBBP test sets, using five seeds, the same predictive-variance target, and the original R-MAT occlusion protocol.

MethodCaco2 ΔFid\Delta\mathrm{Fid}BBBP ΔFid\Delta\mathrm{Fid}
MUG0.0029-0.0200
IG0.0035-0.0209
Saliency-0.0582-0.0562

On Caco2, MUG and IG achieve small positive ΔFid\Delta\mathrm{Fid}, whereas all three methods yield negative values on BBBP. Negative fidelity is therefore not unique to MUG.

We distinguish the architectural conditions for exact completeness from empirical fidelity under input occlusion: completeness provides an additive accounting of predictive variance under the stated assumptions, but does not by itself guarantee that masking highly attributed atoms will reduce uncertainty.

Q1: What MUG Offers Beyond Saliency.

We agree that Table 1 supports comparable occlusion fidelity rather than a substantial ranking advantage over Saliency. MUG’s main additional contribution is a signed, additive decomposition of predictive variance: under the stated architectural assumptions, its appropriately rescaled contributions sum exactly to the variance and can be aggregated into quantitative fragment contributions.

Appendix H.1, Table 14 verifies this additive consistency: the mean of dataset-median completeness residuals is 7.98×10−67.98 \times 10^{-6} for MUG, compared with 2.29×1032.29 \times 10^{3} for Saliency. Saliency scores can also be summed over atoms, but these sums are not guaranteed to represent fragment contributions to total predictive variance. The ranking-based fidelity metrics in Table 1 do not assess this property.

Q2. Why Completeness is a Contribution.

Two things are worth separating. Completeness does not improve the ranking, and we do not suggest it does: in Table 28 the biased ReLU head and input ×\times gradient both break the identity yet reach ΔFid\Delta\mathrm{Fid} of 0.3050.305 and 0.3160.316, above the 0.1580.158 of the complete configuration.

What it gives is that the shares add up. Any set of atoms has a share equal to the sum of its atoms', and that sum is still a fraction of the molecule's predictive variance. A ranking cannot be added, so it cannot say how much of the uncertainty a building block accounts for.

Section 5 is built on exactly that, and the revision now says so in its acquisition rules. It scores building blocks rather than atoms, takes the one with the larger share, and selects additional compounds containing it. Table 6 changes only that step: pick the block by a coin flip instead and the hit rate falls from 56.2%56.2\% to 47.2%47.2\%. Section 4.4 needs it too, where we divide each fragment's attribution by its atom count, which only means something if the fragment score is a sum over atoms.

The residual also tells us when the guarantee has lapsed, as it does for R-MAT's attention pooling in Table 5. A ranking gives no such signal.

Q3: Distinguishing Epistemic and Aleatoric Uncertainty in the Baseline Comparison.

Thank you for raising this point. Yang & Li’s framework provides both aleatoric and epistemic atom-level uncertainty estimates, and we evaluated both components. Their results, together with total uncertainty, are reported in Appendix H.5. We originally selected the aleatoric component for Table 1 because it achieved higher mean occlusion-controlled fidelity (ΔFid\Delta\mathrm{Fid}) than the epistemic component on all three regression datasets. Nevertheless, we agree that the uncertainty targets should be distinguished explicitly. We have therefore revised Table 1 to report both aleatoric and epistemic components. Would the Reviewer recommend retaining both components for completeness, or reporting only epistemic uncertainty to match MUG’s target?

AuthorsOpenReview ↗

Response to Reviewer wtzS (3/3)

Q4: Final-Layer Representations Already Aggregate Neighboring Atoms.

We agree with the Reviewer that atom-level representations in the final GNN layer already aggregate information from neighbouring atoms. We have made this explicit in the revision, where the shares are defined: each share is attributed to atom m together with its L-hop environment, rather than to the atom in chemical isolation (Section 3.2).

We consider that context appropriate rather than a limitation. An atom's contribution to a molecular property is not independent of its neighbors, since inductive and resonance effects are exactly how substituents change the electronics at a site, so a score carrying some neighborhood is closer to the chemistry than one pretending an atom acts alone. What we do not want is the context swamping the atom, which is what deep stacks and oversmoothing give you. Our models are shallow, with three message-passing layers on every dataset (Table 10), so the mixing stays local.

Q5: Inconsistent Symbol Definitions.

We thank the Reviewer for pointing this out and apologize for the confusion. We have corrected “fragment mm” to “atom mm” in Section 3.2, Eq. (5), and the accompanying text. We have also added Eq. (6) to define fragment-level attributions as sums over constituent atoms. The terminology has been corrected consistently in the Figure 2 caption, Appendix B, Proposition 1 in Appendix C, Remark 1 in Appendix D, and Appendix H.8. Throughout the manuscript, mm now denotes an atom and FF a fragment.


We hope our responses and revisions address the Reviewer’s concerns. We would be happy to clarify any remaining questions during the discussion period.

Reviewer_4uicOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Molecular Uncertainty Gradients (MUG), a method for identifying which parts of a molecule contribute to a GNN's predictive uncertainty. MUG estimates uncertainty from the variance across Monte Carlo dropout predictions and uses gradients of the final-layer node representations to assign uncertainty scores to individual atoms. The authors show that these scores sum to the predictive variance after rescaling (provided certain architectural assumptions are satisfied). They evaluate the method on molecular property prediction datasets through atom-masking experiments and by excluding specific chemical substructures from training.

Critical strengths

I think a major strength of the paper is its clear mathematical derivation of the uncertainty decomposition. The authors show how predictive variance can be decomposed into signed atom-level contributions and clearly explain the architectural assumptions under which the completeness guarantee holds. It is also useful that MUG can be applied to existing models without retraining or additional prediction heads.

Critical weaknesses

My main concern is the novelty of the proposed uncertainty decomposition compared to existing work. Bley et al. "Explaining Predictive Uncertainty by Exposing Second-Order Effects" (Pattern Recognition, Volume 160, 2025) introduce CovGI-marg, which decomposes predictive variance into signed feature-level contributions using the covariance of gradient-input attributions. MUG's attributions appear mathematically equivalent to CovGI-marg applied to the final-layer node representations (up to a constant scaling factor) yielding identical normalized attribution shares under MUG's stated assumptions. Also, Bley et al. describe a variance-conservation property closely related to MUG's completeness guarantee. This work is not cited or discussed in the manuscript. While the application to molecular GNNs and the experiments localizing uncertainty to unseen chemical substructures represent distinct contributions, it is unclear how much of the underlying mathematical formulation is novel relative to this prior work.

Questions for authors

Could the authors clarify the relationship between MUG and CovGI-marg introduced by Bley et al. (2025)? The two methods appear to produce equivalent normalized attributions when CovGI-marg is applied to final-layer node representations. Given that Bley et al. also establish a variance-conservation property, which theoretical contributions distinguish MUG from this prior work? A discussion of the mathematical relationship and any practical or computational differences would help clarify the paper's novelty. At minimum, the authors should cite Bley et al. and discuss its relationship to the proposed method.

3 comments

We thank the Reviewer for pointing us to Bley et al. (2025). We had not seen it, and its absence was a real gap in our related work. We fix it in the revised paper: Section 2 now discusses the work in its own paragraph, and new Appendix E derives the relationship in full.

We agree with the Reviewer's central mathematical observation. At our attribution point the marginal summary proposed in Bley et al. and our per-atom shares are equivalent up to a constant factor of 2/T2/T, so the normalized shares are identical. The appendix derives this directly. The covariance form, the signed contributions and the conservation of predictive variance are theirs, and the revised paper says so. We have revised our second contribution in the introduction accordingly, considering the following:

A single differentiation in MUG is enough. Only the pooling and the output head need to scale proportionally with their input; the message-passing stack below them can be arbitrarily deep, nonlinear and biased. One backward pass through σ2\sigma^2 then yields every atom's share, where the second-order construction needs one explanation per ensemble member. The compute is of the same order, since that backward pass runs through all TT passes; what is saved is the per-member explanations and the matrix. The revision also states the condition more carefully than our original text did. What conservation needs is that the shortfall each member's attribution leaves be the same across members, which holds at the final layer, where the shortfall is the shared head bias. At the input it does not hold in general, because each layer's bias passes through that pass's dropout mask. The single differentiation is exact at the last layer because the variance is degree-two homogeneous there, which a direct attribution at the input is not: a completeness residual of 1.26×10−51.26 \times 10^{-5} against 1.96×10−11.96 \times 10^{-1} on trained models (Table 28). That row compares two direct attributions of the variance, not their covariance construction, which, with Gradient × Input underneath, is likewise exact only where the shortfall is shared.

Ensembles with separate heads. Monte-Carlo dropout needs no correction because one output head serves every pass. A deep ensemble whose members each train their own head does, and our Proposition 1 gives the shortfall in closed form, −2TCovt(ft,bt)-\frac{2}{T}\mathrm{Cov}_t(f_t,b_t), measured in Table 28 as 4.85×10−24.85 \times 10^{-2} against 5.39×10−55.39 \times 10^{-5} with tied biases. Their construction covers both kinds of ensemble, given a conservative base attribution, so the correction is specific to the single differentiation.

Which summary conserves. Conservation belongs to the full matrix, so keeping only the diagonal discards the off-diagonal mass, and the diagonal is conservative only when the off-diagonal covariances sum to zero. Message passing does not give that, because adjacent atoms share a receptive field; the new appendix gives the gap exactly (Appendix E, Eq. 22). This does not conflict with their Table 1, where the diagonal ranks more reliably, since ranking robustness and exactness are different properties and we need the second.

Attributing at nodes rather than features. Atoms are nodes, so attributing at final-layer node representations gives one share per atom directly, and because the shares are exact and additive, any fragment's share is the sum over its atoms. Nothing is re-run, re-weighted or approximated at the fragment level. That is the granularity chemists often work at in practice, reasoning in atoms, functional groups and building blocks, and their paper runs no graph experiments.


We would be glad to know whether this framing seems right to you, and whether we have misread anything in your reference. We welcome further discussion of this or any other concern, so if any point needs more detail than the revision provides, please let us know.

Dear Reviewer 4uic,

We are glad that the Reviewer recognised our efforts, and we are even happier that our response satisfactorily addressed your concerns about the relationship between MUG and CovGI. Thank you also for clarifying which additional tests would be valuable for evaluating our method. Following your suggestions, we compared MUG with variance saliency for fragment selection in fragment library design and conducted direct measurements of runtime and memory usage for MUG and CovGI.

Fragment Selection in Fragment Library Design

We compared MUG with variance saliency using the same experimental design as in our original study. Results are reported as mean ± SD across five seeds.

TargetMethodAvg. score ↓Best score ↓Hit rate ↑Top-1% recall ↑
8PJKMUG−9.94 ± 0.02−12.36 ± 0.0052.32 ± 1.03%92.71 ± 0.88%
Variance saliency−9.93 ± 0.02−12.36 ± 0.0051.23 ± 1.21%90.08 ± 0.85%
5L2SMUG−9.28 ± 0.03−11.48 ± 0.0056.22 ± 1.95%89.61 ± 1.18%
Variance saliency−9.30 ± 0.02−11.48 ± 0.0056.88 ± 1.57%89.46 ± 3.58%

On 8PJK, MUG achieved a hit rate that was 1.09 percentage points (pp) higher than that for variance saliency, corresponding to approximately 14 additional hits per campaign. On 5L2S, its hit rate was 0.66 pp lower, corresponding to approximately 8 fewer hits per campaign. Both methods achieved comparable mean docking scores and identical best docking scores for each target. The largest observed difference was in top-1% recall (fraction of the top 1% of compounds ranked by docking score), where MUG achieved higher mean values than variance saliency on both targets: by 2.63 percentage points on 8PJK and 0.15 percentage points on 5L2S.

Runtime and Memory Measurements for MUG and CovGI-marg

We benchmarked MUG and CovGI-marg on 100 molecules from the Caco-2 test set across three GNN checkpoints, using 800 Monte Carlo dropout samples per molecule. All methods used final-layer atom representations. We measured CPU attribution time per molecule and peak memory usage, reporting both total peak memory and the increase relative to the baseline measured after loading the representations and prediction head.

MethodTime (ms/mol) ↓Peak memory increase (MiB) ↓Peak memory (MiB) ↓
CovGI-marg (GitHub)53.24 ± 4.1119.3 ± 1.32672.7 ± 1.4
CovGI-marg (batched autograd, full matrix)3.60 ± 1.31699.3 ± 12.53351.2 ± 12.1
CovGI-marg (Eq. (5), analytic gradients)1.26±0.52‾\underline{1.26 \pm 0.52}343.5 ± 13.02995.5 ± 12.5
MUG1.25 ± 0.50335.0±13.1‾\underline{335.0 \pm 13.1}2988.3±11.3‾\underline{2988.3 \pm 11.3}

The authors’ original CovGI-marg implementation computes gradients one ensemble member at a time, resulting in the longest runtime (53.23 ± 4.11 ms per molecule) but the smallest peak memory increase (19.3 ± 1.3 MiB). Computing all sample gradients in a single batched autograd call reduced the mean runtime to 3.60 ± 1.31 ms per molecule and 699.3 ± 12.5 MiB. Using the linearity of covariance in Eq. (5) of Bley et al. and an analytic gradient formulation enabled by the linear prediction head and mean pooling in our network, we reduced CovGI’s runtime and memory usage to levels comparable to those of MUG. Thus, MUG does not outperform the most efficient CovGI implementation tested in terms of runtime, but it is substantially faster than the authors’ original implementation.

Overall, MUG achieved slightly higher top-1% recall than variance saliency on both targets. MUG was also substantially faster than the authors’ original CovGI-marg implementation, while matching the runtime of the fastest optimized implementation we tested.

We hope that these additional experiments address the Reviewer’s remaining concerns and clarify the practical performance and scope of our contribution. We would be grateful if you could take these results and clarifications into account in your final assessment of our work.

Thank you for the additional experiments. According to your results, MUG and variance saliency achieve similar scores. It is unclear what advantage MUG offers in this respect. However, MUG does show 2.63 percentage points in top 1% recall on 8PJK. Therefore, the results indicate usefulness of uncertainty-guided fragment selection, but provide limited evidence that MUG's exact decomposition offers a consistent practical advantage over simpler uncertainty attribution methods. For runtime, MUG is substantially faster than the original CovGI-marg but is comparable to the optimized implementation. I acknowledge that MUG presents a framework that combines ideas and applies to the novel application of molecules. However, the experimental results do not suggest a meaningful improvement over conventional methods. I will therefore maintain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.