acceptodds
Under review as a conference paper at ICLR 2027

The Shift–Utility Gap: Representation Shift Does Not Determine Predictive Utility in 3D Molecular Prediction

Abstract

Three-dimensional (3D) molecular models can exploit geometric information beyond coordinate-independent two-dimensional (2D) topology, but they also inherit a deployment vulnerability: predictions depend on the coordinate realization supplied at inference time. We study a problem distinct from representation-shift detection itself. A strongly shifted 3D representation can still outperform a 2D expert for a particular molecule, while an apparently in-distribution representation need not provide additional predictive value. We call this mismatch the shift–utility gap. On OpenADMET-ExpansionRx, end-to-end Uni-Mol falls from macro on original coordinates to after coordinate randomization; a matched Frad analysis drops from to . Across six external regression datasets, clean-versus-random shift is detected almost perfectly (AUROC 0.9957–1.0000), whereas clean sample-level utility discrimination remains near chance (AUROC 0.4430–0.5256) and randomized-condition utility varies substantially (0.4206–0.6539). We introduce CORE (COordinate-shift-aware Routing of Experts), which combines shift evidence with 2D/3D predictions, disagreement, endpoint context, and a bounded 2D-anchored robust path. Under randomized coordinates, CORE reaches macro versus for endpoint-fixed static mixing (BH-adjusted ); under original coordinates static mixing is slightly higher ( versus ), but not significantly so (). Full-architecture transfer on Lipo and ESOL shows the same qualitative failure-containment pattern beyond OpenADMET. The main contribution is therefore the empirical distinction between representation shift and relative predictive utility; CORE is one selective-routing response to that distinction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular property prediction, 3D molecular representations, representation shift, relative predictive utility, mixture of experts, robust machine learning, ADMET

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_i7ypOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates the robustness of 3D molecular prediction models to coordinate perturbations. The authors identify a shift–utility gap: although coordinate perturbations can substantially degrade the overall predictive performance of 3D models, the magnitude of representation shift does not reliably indicate whether a 3D model will outperform a 2D model on an individual sample.

Based on this observation, the authors propose CORE, a dynamic expert-routing framework that combines 2D and 3D molecular predictions with representation-shift descriptors, prediction disagreement, and a bounded 2D-anchored robust pathway.

The authors validate their findings on OpenADMET and six external molecular regression datasets, demonstrating the distinction between representation-shift detection and sample-level predictive utility. Experiments with Uni-Mol and Frad illustrate the vulnerability of 3D models to coordinate perturbations. Additional experiments, including ablation studies and evaluations on alternative conformers, show that CORE improves robustness over static expert fusion under severe coordinate randomization, while offering limited advantages under clean coordinates and more realistic conformer variations.

Critical strengths
  • The paper identifies an important distinction between representation shift and sample-level predictive utility. Experiments across multiple molecular datasets show that representation shift can be detected reliably, yet its magnitude does not consistently indicate whether a 3D model outperforms a 2D model for individual molecules.
  • The paper provides extensive experiments across multiple datasets, including evaluations with different 3D encoders, ablation studies, and statistical significance tests. The proposed CORE framework also demonstrates statistically significant improvements over static expert fusion under severe coordinate randomization. These experiments lend credibility to the empirical findings, although the practical advantages of CORE under more realistic conformational variations remain limited.
Critical weaknesses
  • CORE primarily combines existing techniques, including representation-shift detection and dynamic expert routing, without introducing a substantially new methodological idea. More importantly, the ablation studies show no statistically significant improvement over a simpler two-expert dynamic gating baseline. This raises concerns about whether the additional architectural complexity is necessary or provides meaningful benefits, weakening the paper's methodological contribution.
  • Although CORE significantly outperforms static expert fusion under severe coordinate randomization, this perturbation type is included during training and may not reflect realistic molecular conformational variations. As shown in Figure 4, CORE does not significantly outperform static fusion on held-out ETKDG and MMFF conformers. Consequently, it remains unclear under which realistic conditions CORE should be preferred over simpler alternatives and what substantial practical advantages it offers. This limits the significance and applicability of the proposed approach.
Questions for authors
  • Could the authors clarify the key advantages of CORE and why practitioners should prefer it over simpler alternatives, such as static expert fusion or a two-expert dynamic gate? Given that CORE shows no statistically significant improvement over the simpler dynamic gate and offers limited advantages on held-out ETKDG/MMFF conformers, what specific deployment scenarios would justify its additional complexity? Which experimental results most strongly support this practical advantage?
2 comments
AuthorsOpenReview ↗

Response to Reviewer i7yp

For a practitioner, we would start with the two-expert gate if adaptive routing is needed, and with static fusion if the main objective is accuracy on ETKDG/MMFF conformers. Even under severe randomization, where Full showed its clearest gain over Static, Full achieved a macro-R² of 0.570981 versus 0.569236 for the two-expert gate. The incremental difference was +0.001745 (BH-adjusted q = 0.716582).

We examined this choice further in a paired comparison on identical outer-test molecule–endpoint records. All three ETKDG conformers and the MMFF-lowest conformer were held out from training and model selection.

Held-out conditionFull COREStatic RouterTwo-expert GateFull − Gate (corrected 95% CI)
ETKDG, mean across three conformers0.5839100.5933980.581411+0.002499 [−0.002807, 0.007804]
MMFF-lowest0.5942450.6026640.592856+0.001389 [−0.004842, 0.007620]

We also examined the error a practitioner might encounter from an unfavorable conformer. For each molecule–endpoint pair, we took the largest absolute error across its four conformers, then averaged within each endpoint. Against each baseline, Full had a lower mean worst-conformer error on five of nine endpoints. All 18 endpoint-wise intervals included zero. These exploratory intervals use an approximate correction for repeated-CV dependence and do not provide simultaneous coverage across comparisons.

Under the severe randomization condition included in training, Full retained more accuracy than endpoint-specific static fusion (random-coordinate ΔR² = +0.024463, BH-adjusted q = 0.013301). The two-expert gate captured most of that numerical gain. The evidence supports adaptive routing relative to static fusion in this stress test. The direct comparison with the two-expert gate provides insufficient evidence that the third expert adds enough predictive benefit to justify its complexity. We will make that distinction explicit in the discussion and limit our recommendation for Full CORE accordingly.

Thank you for the detailed response and the additional analyses. I appreciate the authors' transparency in distinguishing the benefits of adaptive routing from the incremental contribution of Full CORE.

The new results largely confirm my original concerns. While adaptive routing provides meaningful improvements over static fusion under severe coordinate randomization, Full CORE does not demonstrate a statistically significant advantage over the simpler two-expert gate. Furthermore, under the more realistic held-out ETKDG and MMFF conformer settings, static fusion remains competitive and achieves higher average performance.

I appreciate the authors' willingness to clarify these limitations and narrow their practical recommendations. However, the additional evidence does not substantially address my concerns regarding the methodological novelty and practical necessity of the full framework. The empirical observation about the distinction between representation shift and predictive utility remains valuable, but the added complexity of Full CORE is not sufficiently justified by the reported results.

Therefore, I will maintain my original rating.

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

This paper investigates the shift–utility gap in 3D molecular prediction, showing that representation shift does not reliably indicate whether a 3D model outperforms a 2D counterpart on individual samples. The authors propose CORE, an adaptive routing framework that combines 2D/3D predictions, representation-shift descriptors, and a bounded robust expert. Experiments on OpenADMET and six external datasets demonstrate severe sensitivity of 3D models to coordinate perturbations. CORE improves robustness over static fusion under randomized coordinates, while remaining competitive under clean conditions.

Critical strengths

This research question is well-motivated:

Conformer dependence at inference is a real obstacle to deploying 3D encoders, and the second-encoder (Frad) replication plus six external datasets is more validation effort than average submissions on this topic make.

Experimental protocols are good.

I have seen the following details: 1.Fold-specific PCA bases and scalers fitted on inner-training data only. 2. ETKDGv3/MMFF-lowest conformers are excluded from every selection step including threshold tuning. 3. A fixed coordinate archive built before downstream extraction. Many papers in this area are looser.

Critical weaknesses

1.The paper seems to be premature and the overall writing is poor. And Figure 1(b), which illustrates the paper’s central contribution, is labeled a conceptual illustration. Table 1 and Table 6 are identical.

2.I can find a clear limited novelty and missing related work in this paper:

The central observation that representation shift does not necessarily imply poor predictive utility appears to be a natural extension of established findings rather than a fundamentally new insight. Prior work has already distinguished OOD detection from prediction reliability (Guerin et al., AAAI 2023; Averly & Chao, ICCV 2023; Liang et al., TMLR 2024).

More importantly, this distinction has also been studied specifically in molecular property prediction: Uncertainty Quantification in Molecular Machine Learning for Property Predictions under Data Shifts (JCIM 2026) discusses the mismatch between molecular distributional distance and prediction error. Hamakawa & Miyao (JCIM 2025), which is cited in passing but never engaged with, shows that 3D models can remain competitive with 2D fingerprints even with inaccurate conformers. Additionally, Mole-PAIR (He et al., 2025) has explored learned OOD detection for molecular foundation models, including Uni-Mol, but is not discussed or compared against.

The proposed CORE architecture also largely builds on established mixture-of-experts routing and multimodal fusion techniques. Given that the shift–utility gap is primarily demonstrated using three handcrafted PCA-based descriptors, and that the bounded robust expert and consistency loss provide no significant independent improvements, the paper's conceptual and methodological novelty remains limited. The authors should better distinguish their contributions from these closely related works and clarify what fundamentally new understanding or capability CORE provides.

3.The experiment is not convincing enough for me.

The abstract leads with CORE at 0.5710 versus 0.5465 for static mixing under randomization. But Table 7 shows the 2D expert at 0.567139 under the same condition, and Table 8 gives q = 0.322977 for CORE versus the 2D expert there. Across different conditions, static mixing is numerically ahead on original (0.6022 vs 0.5943), on ETKDG mean (0.5934 vs 0.5839), and on MMFF-lowest (0.6027 vs 0.5942)

4.The ablations do not support the architecture. The baseline comparisons are insufficient, as most of the evaluated methods are either simplified variants of CORE or basic 2D/3D predictors, making them closer to ablation studies than comparisons against competitive alternative approaches.

5.Appendix K states that the 15 paired values are repeated model fits over five distinct test partitions, and that the correction carries no seed-within-fold term. The effective sample size is therefore closer to 5. Given that q = 0.0133 is the paper’s only significant positive result, it should not rest on a test the authors themselves describe as lacking the relevant hierarchical term.

Questions for authors

1.What is the fold-level (n = 5) result for CORE versus static mixing under randomized coordinates? Alternatively, a variance decomposition separating seed variability from fold variability would address the same concern.

2.Under randomized coordinates, Table 8 reports q = 0.322977 for CORE versus the 2D expert. What does CORE provide over simply disabling the 3D pathway in this condition?

While AI-assisted research is becoming increasingly common, authors remain responsible for ensuring the scientific rigor and quality of their submissions. I encourage the authors to conduct a more thorough evaluation and carefully validate their claims before submission to a venue such as ICLR.

2 comments
AuthorsOpenReview ↗

Response to Reviewer thoM

Thank you for questioning the unit of replication. Our statement that the three seeds shared five fixed test folds was incorrect. We checked the archived molecule IDs. Within each seed, the five scaffold test folds were disjoint and together covered all 7,608 molecules. Test membership changed across seeds. Averaging the three results called “fold 0,” for example, would combine three different test sets. Because the seed changed both the partition and the model fit, this design cannot separately estimate initialization variance.

Random-coordinate result. Within each seed–fold block, we matched Full and Static on the same molecule–endpoint records and labels, then computed the difference in nine-endpoint macro-R². Full achieved 0.570981 versus 0.546518 for Static, a mean paired difference of +0.024463. Full led in 14 of 15 blocks. It won all five folds in two repetitions and four of five in the third. We used the Nadeau–Bengio-style standard error, SD of the paired differences × √(1/15 + 1/4), with 14 degrees of freedom as an approximate correction for repeated-CV dependence. The two-sided p-value was 0.003990, with BH-adjusted q = 0.013301 across the ten clean/random comparisons. We will correct the split description throughout the manuscript and include all 15 paired differences in Appendix K.

Static had the higher clean macro-R² (0.602184 versus 0.594331, q = 0.325591). Using the same prediction runs, we also compared the clean-to-random drop directly. Full lost 0.023350 and Static lost 0.055666. Full had a smaller drop in every block, with a mean paired difference of −0.032316 (BH-adjusted q = 0.000355 across the five degradation comparisons).

Why not disable 3D? If severe corruption is known in advance, disabling 3D is a reasonable choice. Under randomization, Full exceeded the True No3D ablation by only 0.003842 (q = 0.322977). On original coordinates, the difference was +0.027192 (q = 0.000930). The latter result motivates studying selective use of 3D when coordinate quality is unknown. The random-coordinate result alone gives little reason to prefer Full to a correctly triggered 2D fallback.

We will consolidate Tables 1 and 6 and replace the schematic in Figure 1b with empirical results. We will also discuss Hamakawa and Miyao’s finding that 3D models can remain competitive with imperfect conformers, alongside Mole-PAIR’s learned molecular OOD detection. Our experiments ask whether a score identifies which of two experts has lower error on the same molecule and endpoint. We evaluated that question for three shift descriptors. A head-to-head routing comparison with a learned detector such as Mole-PAIR remains open.

Thank you for the clarification and additional results.

I stay by my point that the overall contribution is very limited and more that of a workshop, as illustrated in the W1 and W2. The rebuttal only discuss W3 to W5, and the improvement is limited. Moreover, the author also confirms the statement mistakes in the paper. Overall, this paper is clearly below other papers in my batch. Therefore, I will maintain my original score.

Reviewer_Vd4SOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates the distinction between representation shift and sample-level predictive utility in 3D molecular property prediction, termed the shift–utility gap. The authors observe that, although coordinate perturbations can substantially degrade the overall predictive performance of 3D molecular models, a shifted 3D representation may still yield more accurate predictions than a 2D expert for individual molecules. Conversely, an apparently in-distribution 3D representation does not necessarily provide additional predictive value.

Motivated by this observation, the authors propose CORE (COordinate-shift-aware Routing of Experts), a framework that adaptively combines 2D and 3D predictions using representation-shift descriptors, expert disagreement, endpoint information, and a bounded 2D-anchored robust pathway.

Critical strengths

The paper highlights an empirical observation: even when coordinate perturbations cause a substantial degradation in the overall predictive performance of 3D molecular models, this does not necessarily imply that a 2D model is more accurate for every individual molecule. The distinction between performance degradation and sample-level relative predictive utility is worth noting when designing molecular prediction systems.

Critical weaknesses

While the paper presents an empirical observation, the novelty and significance of the overall contribution appear somewhat limited. The distinction between aggregate performance degradation and sample-level predictive utility is relatively intuitive. Furthermore, the empirical advantages of CORE over simpler fusion strategies are limited to severe coordinate perturbations.

Questions for authors
  1. Could the authors further clarify the novelty and practical significance of the shift–utility gap?
  2. Given that realistic conformational variations are unlikely to induce shifts as severe as coordinate randomization, how relevant are CORE's robustness improvements to practical molecular property prediction?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Vd4S

Thank you for asking what the shift–utility gap changes in practice. A descriptor can detect coordinate randomization almost perfectly and still offer little guidance about which expert will make the smaller error for an individual molecule with its original coordinates. Across six external regression datasets, the three descriptors we tested distinguished original from randomized representations with AUROCs of 0.9957–1.0000. On original coordinates, their AUROCs for discriminating between cases where 2D or 3D had the smaller squared error ranged from 0.4430 to 0.5256.

The OpenADMET threshold experiment shows the consequence for an actual routing decision. We selected a threshold on validation data for the PCA reconstruction error. On original test coordinates, 2D was more accurate in 56.24% of the cases routed to 3D. Among the cases routed back to 2D, 3D would have been more accurate in 42.39%. The rule achieved a macro-R² of 0.5382, compared with 0.5671 for the 2D-only baseline. Measuring both types of routing error makes the practical cost of using this shift score explicit. The relationship also varies by dataset and condition: on randomized ESOL inputs, the descriptors retained some information about relative utility (AUROC 0.6212–0.6539).

We agree that randomization is a severe stress test. The router encountered this type of corruption during training. Under randomization, CORE improved macro-R² over static fusion by 0.0245 (BH-adjusted q = 0.0133). On the held-out ETKDG conformers, static fusion had the higher mean macro-R² (0.5934 versus 0.5839), with the same ordering on MMFF-lowest. For a pipeline using conformers like these, static fusion is a reasonable default. We will make this distinction explicit in the manuscript and replace the schematic in Figure 1b with the measured shift and utility AUROCs.

Thank you for the additional clarification and experiments. I appreciate the authors' efforts to address my concerns.

The additional resutls show that the improvements of CORE over static fusion are primarily observed under severe coordinate perturbations, which may limit the practical significance of the method.

Therefore, I will maintain my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.