acceptodds
Withdrawn from ICLR 2027

SpeCliff: Spectral Decoupling of Molecular Graph Representations for Activity Cliff Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Graph neural networks often fail on molecular activity cliffs, where small structural edits cause large potency changes. We trace this failure to a spectral mismatch in the standard encoder-plus-mean-pooling design. Local substituent edits produce high-frequency graph signals, while linear Laplacian propagation attenuates higher frequencies and mean pooling dilutes localized changes. Motivated by this view, we propose SPECLIFF, a dual-branch architecture that combines a low-pass scaffold branch with an explicit high-pass branch based on Chebyshev-approximated spectral graph wavelets and bridge-edge decomposition. Across all 30 MoleculeACE tasks, SPECLIFF achieves the best average RMSE and wins on the largest number of datasets compared with prior baselines. Ablation studies, frequency-response analysis, and case studies show that the high-pass branch is critical and concentrates signal around substituent edits that drive activity cliffs.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular property prediction, graph neural networks, activity cliff

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_TACxOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that GNN failure on molecular activity cliffs stems not from oversquashing or oversmoothing during propagation, but from dilution at readout: localized substituent edits carry energy at nonzero graph frequencies, while mean pooling retains only the zero-frequency (constant) component, so a signal concentrated on a few atoms contributes weakly to the pooled graph vector. The proposed model, SpecLiff, keeps GraphCliff as a low-pass scaffold branch predicting μ, and adds a high-pass branch: Chebyshev-approximated spectral graph wavelets at two scales, bridge-decomposition features, and ℓ2-magnitude summation yielding z_high, from which an effect head predicts a correction τ, with ŷ = μ + τ. A matched-pair consistency loss supervises τ via τᵢ − τⱼ ≈ yᵢ − yⱼ on training cliff pairs. On all 30 MoleculeACE tasks, SpecLiff reports the best average RMSE (0.654 vs. 0.665 for GraphCliff) and the most dataset-level wins.

Critical strengths

The problem framing is the paper’s main contribution, and it is a good one. Relocating the failure from propagation to pooling is a clean and useful reframing. The empirical audit in Appendix A supports it constructively: attachment edges are shown not to be topological bottlenecks (lower betweenness, higher Ollivier-Ricci curvature than scaffold-internal edges, p = 1.0 in both one-sided tests), which argues directly against curvature-guided rewiring as the right remedy for this benchmark. The dilution account is intuitive, formalized at an appropriate level, and the paper is careful to state where each proposition does and does not apply.

Critical weaknesses

W1. The central mechanistic claim is asserted rather than demonstrated. The paper’s causal story has three links: (i) local structural differences between cliff pairs manifest as high-frequency graph signals; (ii) the chosen wavelet scales g_s(λ) = 1 − e^(−sλ) actually isolate that frequency band; (iii) what the high-pass branch isolates corresponds to the structural edit. None is directly verified.

Section 5.3 is the closest attempt, but it measures a different quantity: the spatial concentration of single-molecule representation energy on branch-labeled atoms, not the spectral content of the pairwise difference signal that Proposition 1 is about.

W2. RMSE_cliff is missing, and this is not optional for this benchmark. MoleculeACE was introduced specifically to argue for “activity-cliff-centered” metrics. A paper whose entire motivation is cliff performance, evaluated only on aggregate RMSE over all test molecules, leaves open whether the gain comes from cliff compounds at all or from the non-cliff majority. Given cliff ratios of 8.8–54%, an overall improvement of 0.011 could be produced either way.

Questions for authors

Q1 (W1). Can the authors provide direct evidence for the spectral premise? Concretely: (a) for aligned cliff pairs, compute the graph Fourier transform of the difference signal h(G) − h(G′) and report the energy distribution over λ, compared against non-cliff pairs of similar size; (b) report the selected s₁, s₂ per dataset together with the resulting filter gain curves, to show which band is actually passed; (c) quantify the correspondence between high-pass energy and the MCS-derived edit atoms (not the authors’ bridge labels), e.g. as an attribution AUC. Without (a) the motivating theory is untested on real data; without (c) Figure 5 remains anecdotal.

Q2 (W2). Please report RMSE_cliff for SpecLiff and all baselines, per dataset and averaged, using the official code.

Q3 . Please provide the full search space: ranges and priors for every tuned hyperparameter, the number of Optuna trials, and the selected values per dataset, in particular s₁, s₂, K, and β. Is any ordering (s₁ < s₂) or minimum separation enforced, and how often do the two selected scales end up close enough that the branch is effectively single-scale?

Q4. SpecLiff is GraphCliff plus an additional branch, yet reports lower training time per epoch (0.61 vs 0.94 s) and lower inference latency (11.03 vs 11.72 ms). Could the authors clarify: are the two models compared under matched batch size, hidden dimension, and number of message-passing layers?

Q5. Two ablations would disentangle architecture from supervision: (a) Low-pass + pair loss, i.e. no high-pass branch but with pairwise supervision applied to μ (μᵢ − μⱼ ≈ yᵢ − yⱼ). If this recovers much of the gain, the contribution is the supervision, not the spectral branch. (b) High-pass branch with β = 0 but retuned, rather than inheriting full-model hyperparameters. Given w/o pair loss (0.698) and w/o high-pass (0.702) differ by only 0.004, does the gain in fact require both components jointly? If so, the paper’s narrative, in which the high-pass branch is the primary contribution and the pair loss a training detail, may need rebalancing.

Q6. How many qualifying cliff pairs exist per dataset in train split? Given single-seed results and a 0.011 margin, can the authors report mean ± std over at least 3–5 seeds for the full model and the ablations? Specifically, is the pair-loss ablation gap statistically distinguishable from seed noise?

Q7. Why is the kernel fixed rather than learned? Appendix C calls the branch “a fixed-kernel Chebyshev approximation rather than a freely learned ChebNet filter”, but it is the same object as ChebNet, at the same O(K|E|d) cost, with coefficients frozen. ChebNet, GPR-GNN, BernNet and ChebNetII learn these end-to-end; none is cited or compared. Two controls would settle the branch’s contribution: (a) the same branch with learned coefficients, and (b) the same branch with a low-pass kernel. If (b) also beats the no-high-pass variant, the gain is added capacity, not high-pass filtering. Corollary 1 only requires |h(λ)| ≥ η above the cutoff — a constraint on the response, not a reason to freeze it.

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

The authors highlight a recurring failure of molecular graph neural networks when dealing with "activity cliffs"—pairs of molecules whose potency changes significantly despite only minor structural modifications. They argue that this failure stems from a "spectral mismatch" inherent in the standard architecture, which typically employs an encoder followed by mean pooling. The paper explains that local substituent modifications generate high-frequency signals, whereas Laplacian-based propagation tends to attenuate these high-frequency components, and mean pooling further dilutes these localized changes. To address this, the authors propose a dual-branch decomposition approach that utilizes both low-frequency components (representing the molecular scaffold) and high-frequency components (capturing substituent effects). This method, termed SPECLIFF, decomposes the representation into a scaffold-based low-frequency term and a high-frequency correction term sensitive to substituent modifications. The authors validate their approach by comparing test RMSE across 30 regression tasks from the MoleculeACE benchmark. Furthermore, they conduct additional experiments to confirm the effective preservation of high-frequency signals, perform an ablation study to assess the contribution of individual components, and compare inference latency and training time against GraphCliff (reproduced under identical GPU conditions).

Critical strengths
  • It is logical to analyze and address the causes of failure related to the "activity cliff" by employing a dual-branch design that incorporates both low-frequency and high-frequency components.

  • The model demonstrates consistent performance across benchmarks.

  • Through additional experiments and ablation studies, the authors clarify the specific contributions of individual components.

  • The study provides a comparison of inference and training times, measured on the same GPU.

Critical weaknesses
  • While "dilution" may be one of the causes of performance degradation, other factors—such as data noise, measurement errors, label scaling, or biases like the "cliff ratio"—could also play a role; therefore, further verification or additional experiments are required to investigate this.

  • To demonstrate the generalizability of the proposed method, experiments should be conducted using backbones other than GraphCliff.

  • A limitation regarding the treatment of mean pooling is that variations in the number of atoms introduce terms such as normalization adjustments, making it difficult to encapsulate these effects within a simple theorem; consequently, certain theorems can only be stated for graphs of the same size that have been aligned.

Questions for authors
  • This paper is based on the observation that "edits causing cliffs are primarily concentrated in small branches (pendant structures) separated by bridges." However, it is unclear whether the same alternative approach (high-frequency preservation) is applicable to other types of cliffs—such as those involving positional isomers, stereochemistry, or large-scale scaffold rearrangements. Does this suggest a potential gap between the types of edits covered by the theory and the diverse range of edits actually present in the benchmarks?

  • Could you demonstrate the generalization performance of the proposed method by conducting experiments using backbones other than GraphCliff?

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

The paper proves a few things about how MPNNs may underperform on molecules with activity cliffs as a result of the structure of the pendant graph and the architecture of message-passing and mean-pooling readout. The theoretical analysis includes a bound on the spectral energy of the signal from the pendant branch, how linear message-passing attenuates high-frequency signals, and a bound on difference in the mean-pooled representations of graphs with localized changes.

Architecturally, it uses a shared encoder followed by two branches: one from GraphCliff as the low-pass branch, and the added high-pass branch, where high-pass filtering with the Chebyshev basis is used, along with magnitude sum pooling and additional branch features. There is the additional training objective penalizing deviation of the correction difference of the matched pair from the true potency difference.

The results show the architectural and training additions on GraphCliff improve on it in MoleculeACE, with ablations showing removing each component worsens average performance under the same hyperparameters.

Critical strengths

The improvements over GraphCliff on MoleculeACE appear solid.

Critical weaknesses
  1. The contribution of high-pass filtering is not isolated.

The added branch combines high-pass filtering with magnitude sum-pooling. The current ablations do not establish whether high-pass filtering is necessary or whether the improvement can be obtained through magnitude sum-pooling without high-pass filtering.

The no-high-pass ablation removes the entire branch and pair loss; bypassing only the high-pass filters while retaining magnitude sum-pooling, the correction head, and pair supervision would better isolate the contribution of spectral filtering.

Also, the matched-pair consistency loss introduces the additional assumption that potency differences should be assigned to the correction head, which is not established by the theory.

  1. The published ACANet numbers appear better than the in-house replication and the paper's model. Also, MAPCliff-WMGR (https://doi.org/10.1021/acs.jcim.5c01915) reports overall RMSE of 0.624, better than the paper's 0.654. The paper should address the ACANet discrepancy, and the paper should either include MAPCliff-WMGR for comparison or explain its exclusion.

  2. Previous work on MoleculeACE also include RMSE_{cliff}, or RMSE on the cliff subset, along with RMSE_{all}. The paper only reports RMSE_{all}. Since the paper is addressing the phenomenon arising from cliff, the results of the cliff subset should also be reported.

Questions for authors

More results:

  1. Have you experimented on magnitude sum-pooling with the pair consistency loss without high-pass filtering? This would help separate the contributions of filtering and readout.

  2. What are the evaluation results on the cliff subset, RMSE_{cliff}?

Clarification:

  1. Under Proposition 2's assumptions, mean pooling extracts only the constant mode, so propagation's behavior does not affect the mean-pooled representations. What additional role does propagation attenuation play in the explanation?

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

Related papers

Discussion (0)

Sign in to comment.