acceptodds
Under review as a conference paper at ICLR 2027

Tabular In-Context Learning for Structural Generalization in Molecular Property Prediction

Abstract

Molecular property prediction helps identify compounds with desired characteristics by estimating molecular properties from chemical structures. In applications such as drug discovery and materials design, this often requires generalizing to molecules that are structurally different from those in the labeled training data. Recent studies have applied tabular foundation models (TFMs) to molecular property prediction through in-context learning, but their effectiveness under such structural train–test shifts remains underexplored. To address this gap, we evaluate TFM-based approaches in this setting and introduce MolRIFT (MOLecular Relational Inference with Frozen TFMs), a framework that incorporates explicit molecular comparisons into tabular in-context learning to improve structural generalization in molecular property prediction. In MolRIFT, one TFM makes an initial prediction from molecule-level context, while a second TFM uses molecular pairs and their prediction-error differences to refine it. Across 58 MoleculeACE and Polaris tasks, combining CheMeleon representations with TabPFN-3 wins a majority of tasks against each evaluated baseline, while the MolRIFT framework improves this strong predictor on 46 of the 58 tasks. The gains also persist across four molecular representations and three TFM backbones, supporting MolRIFT as a broadly applicable approach. The code and datasets are available at https://anonymous.4open.science/r/MolRIFT-8773/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Tabular Foundation Model, Molecular Property Prediction, Structural Generalization

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.20 over 5 reviews, above 96% of the others.
Reviewer_jdcAOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MoLRIFT adapts the pairwise differential regression approach from PADRE for use with a frozen TabPFN-3 + CheMeleon pipeline. It employs a second, frozen TFM to predict the out-of-sample error differential of the global model, applying corrections via a combination of bidirectional averaging and similarity-gated aggregation. However, the global TFM baseline is already quite strong; while the pairwise correction achieves more wins than losses across tasks, the improvement in raw RMSE is marginal, and the confidence interval for the comparison against a simple kNN-residual transfer method includes zero.

Critical strengths
  1. Provides complete procedures and hyperparameters for OOF cross-fitting, anchor selection, bidirectional inference and gating, enabling replication or ablation. This is valuable for exploring relative prediction within the tabular ICL framework.
  2. Relative to tree models/GNNs requiring per-task hyperparameter search MoLRIFT's frozen pipeline achieves higher Elo with less end-to-end wall-clock time, placing it on the observed Pareto frontier.
Critical weaknesses
  1. Against kNN residual transfer (using stored OOF errors from similar labeled molecules without learning pairwise differences), MoLRIFT's RMSE reduction is 0.00402 with 95% CI including zero ([−0.00110, 0.00930], Table 3). Against single-molecule residual, the improvement is 0.00768—statistically distinguishable from zero but numerically tiny. The incremental advantage of pairwise correction over "directly reusing neighbor residuals" is not established.
  2. Causilo (Cho et al., 2026) achieves a higher Elo score on TabArena with an inference time approximately 40% that of TabPFN-3.5; however, it is employed solely as a robustness backbone rather than being included in the primary comparisons or serving as a relation predictor. Furthermore, global TFM fine-tuning—conducted under an equivalent computational budget—is entirely absent from the evaluation. If a single round of lightweight fine-tuning could yield greater gains than pairwise correction, then the added engineering complexity would lack justification.
Questions for authors
  1. The relation TFM's supervision signal depends on OOF residuals systematically underestimates errors on structurally distant molecules—as indicated by CheMeleon's failure on activity cliffs and stereochemistry—is the relation TFM merely fitting an already optimistic noise pattern? Are there diagnostic experiments that can distinguish learning a transferable correction relation from fitting spurious correlations in OOF residuals?
  2. Under a matched computational budget, what is MoLRIFT's incremental advantage over directly fine-tuning the global TFM for a few steps? If fine-tuning yields larger gains, is the engineering complexity of OOF construction + relation context + bidirectional inference + gating justified?
3 comments

We thank the reviewer for the feedback. We clarify the evidence for learning useful corrections from OOF errors and the value of our approach relative to residual transfer and global fine-tuning.


W1. Magnitude and consistency of the gains over residual baselines.

MolRIFT’s learned pairwise correction improves on both kNN residual transfer and the single-molecule residual predictor across most tasks, although the mean RMSE gains are modest. Moreover, mean RMSE alone does not show how broadly these gains occur, because larger losses on a few tasks can offset smaller gains across many others.

In structural generalization, more accurate predictions can help prioritize unfamiliar compounds for experimental testing. We therefore believe that improving predictions consistently across diverse tasks is as important as the absolute magnitude of improvement. MolRIFT shows this consistency against both residual baselines, achieving lower RMSE and MAE on most of the 47 regression tasks in each comparison.

ComparatorRMSE W/LMAE W/L
kNN residual transfer34/1335/12
Single-molecule residual35/1238/9

We have added the full per-task RMSE and MAE comparisons against both residual baselines in Appendix C.2, Tables 13–16 (pp. 26–29 of the revised manuscript), so that readers can assess the magnitude of these differences alongside their consistency across tasks.


Q1. OOF errors and transfer to difficult molecules.

Although OOF errors are smaller than structural-test errors, MolRIFT’s corrections track unseen errors better than matched kNN and shuffled-label controls. We first confirmed the difference in error magnitude on MoleculeACE.

Error evaluationMean MAE
Training OOF0.496
Outer structural test0.706

We then tested whether relationships learned from these smaller OOF errors remain useful on the held-out molecules in our primary structural splits. Across 47 regression tasks, we measured the Spearman correlation between each method’s predicted correction and the adjustment actually needed by the global prediction, yq−μqy_q-\mu_q.

ComparatorMolRIFT’s increase in correlation [95% CI]
kNN residual transfer+0.0407 [0.0167, 0.0657]
Shuffled relation labels+0.0415 [0.0167, 0.0678]

MolRIFT’s higher correlation with actual test errors provides evidence that the learned pair relationships provide useful information about unseen molecules beyond directly reusing neighbor residuals. The comparison with shuffled labels further supports the role of the true pair labels in capturing these transferable relationships.


W2/Q2. Backbone choice and comparison with global fine-tuning.

Causilo was released in mid-September, leaving insufficient time to repeat the full experimental suite with it before submission. We nevertheless included it in our backbone experiments, where Causilo serves as both the global and relation TFM.

Because MolRIFT constructs molecular-relation contexts while keeping the TFM parameters unchanged, the same framework can be applied to different TFM backbones. Our Causilo results illustrate how advances in foundation models can translate into better molecular predictions, and newer models such as Kumo Tabular could provide further gains within the same framework.

MolRIFT achieves lower error than the tested lightweight fine-tuning of the global TFM at comparable measured runtimes. In this additional evaluation, both methods use the same TabPFN-3 checkpoints, molecular features, and outer splits across all 58 tasks. Fine-tuning uses AdamW with a learning rate of 10−510^{-5} for up to eight epochs. For each split, we use the measured MolRIFT runtime as the budget for fine-tuning and prediction, reserving time for prediction. The table reports the MolRIFT reference from this matched evaluation, with runtime measured from precomputed features through prediction.

MethodMoleculeACE RMSEMedian total runtime (s)
MolRIFT0.8850920.3
Global fine-tuning0.9411518.0
ComparisonResult
Tasks won by MolRIFT48/58
MoleculeACE RMSE reduction [95% CI]0.05606 [0.03551, 0.07831]

These results support the additional computation used by MolRIFT’s molecular-relation contexts over the tested lightweight fine-tuning alternative.

The author's answer addresses my concerns, and I will increase the score to 3, but as I am not familiar with the relevant field, I will keep the confidence score unchanged.

Thank you for reviewing the additional results and reconsidering the paper. We’re glad the OOF diagnostics and fine-tuning comparison helped address your concerns.

Reviewer_x49WOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

This work considers the performance of tabular foundation models (TFMs) to molecular property prediction through in-context learning under structural train-test shifts. The authors introduce MolRift to incorporate explicit molecular comparisons into tabular in-context learning to improve structural generalization in molecular property prediction. Empirical results on MoleculeACE demonstrate the effectiveness of proposed method.

Critical strengths

The evaluation is generally comprehensive, covering various benchmarks (MoleculeACE, Lo-Hi, DrugOOD) and 12 encoder-backbone combinations (Figure 3), and the proposed methods has consistent improvements across these settings.

Critical weaknesses
  1. While results on rank-based Elo seems good, the native-metric effect sizes are very small. Mean MoleculeACE RMSE improves only from 0.8932 to 0.8849 (Table 1; roughly 0.9% relative), the RMSE-reduction intervals for the two-global-TFM ensemble (0.00547 [−0.00032, 0.01093]) and kNN-residual (0.00402 [−0.00110, 0.00930]) controls include zero (Table 3), and the error-difference pair target is not established over property differences (31/58 wins, 0.88496 vs. 0.88383, interval includes zero; App. C.4). Elo converts small per-task advantages into large ratings, so the framing that molecular comparisons improve structural generalization currently outruns the measured magnitudes.

  2. The scope of the gains is narrower than the structural-generalization framing. Under DrugOOD assay/size/scaffold shifts, MOLRIFT (0.6687 mean ROC-AUC) trails its own global baseline (0.6704) and MoleOOD (0.6946) (Table 20); on publication-year splits the effect vanishes or reverses (App. D.3); and at the query level only 54.8–57.2% of regression predictions improve, with the improvement fraction non-monotone in anchor similarity (Table 21). What survives is a benefit specific to the scaffold-plus-similarity-filtered split family, which the authors partly acknowledge, but the title-level claim is not bounded accordingly. In addition, the evaluation sits squarely on activity-cliff benchmark data without engaging that literature: MoleculeACE ships with task-trained reference implementations of SVM, kNN, RF, XGBoost, GCN, GAT, MPNN, and AttentiveFP, but the benchmark of Table 1 evaluates only tree ensembles, Chemprop, MiniMol/CheMeleon heads, and iMoLD — and re-splits the tasks with a custom scaffold-plus-similarity filter rather than the canonical protocol, so the results are not comparable with any number on the MoleculeACE leaderboard. More broadly, there are some works (e.g., [1]) explicitly targeting activity cliffs, as is the foundational activity-landscape literature [2]. For a paper whose central claim is about precisely these hard pairs, evaluating on their data while neither citing nor comparing to this line of work weakens both the positioning and the empirical scope.

References

[1] Curriculum-aware Training for Discriminating Molecular Property Prediction Models. ICLR 2025

[2] On outliers and activity cliffs why qsar often disappoints. Journal of Chemical Information and Modeling

Questions for authors
  1. (Clarification) Given the mean MoleculeACE RMSE delta of about 0.008 (0.8932 to 0.8849, Table 1), can you characterize the practical magnitude — e.g., per-task relative RMSE reductions or an effect on candidate ranking — rather than Elo and win counts?

  2. (Clarification) App. C.4 shows the property-difference target is not worse than the error-difference target (0.88383 vs. 0.88496, interval includes zero). What evidence justifies residual-difference targets as the paper's default, given this null result?

  3. (Clarification) Why were MoleculeACE's canonical reference algorithms (SVM, kNN, XGBoost, GCN, GAT, MPNN, AttentiveFP) not evaluated? Even one task-trained reference point — e.g., XGBoost or the tuned SVM — under your primary protocol would let readers place the TFM-based numbers relative to the leaderboard the benchmark community uses. Relatedly, was the custom scaffold-plus-similarity re-splitting motivated by a specific flaw in the canonical protocol, and would the conclusions hold under the official split?

  4. (Clarification) The paper runs on activity-cliff benchmark data but cites none of the recent activity-cliff learning literature (e.g., [1] in the Weakness) How does MOLRIFT relate to these works, and does the descriptive cliff-subset result (App. E.2: mean cliff RMSE 0.9565 vs. 0.9706, no paired interval) survive as a claimed advantage against methods trained specifically to discriminate cliffs?

  5. (Clarification) Under DrugOOD (Table 20) the pair stage lowers mean ROC-AUC below the global baseline. Do you have a diagnosis, e.g., does the 0.20 anchor-acceptance rule admit structurally similar but shift-irrelevant anchors?

  6. (Clarification) Table 16 shows 16 query anchors beating the 32-anchor reference on 38/58 tasks, and the adaptive gate loses to the mean-matched fixed coefficient on 26/32 (App. C.7). Were the anchor cap of 32 and τ=0.20\tau = 0.20 prespecified, or selected on these tasks? If prespecified, stating this explicitly would help, since the reference configuration then appears slightly suboptimal rather than tuned.

  7. (Suggestion) A query-level reliability signal beyond the similarity gate would strengthen the story, since roughly 42–47% of regression predictions worsen even at the higher similarity bins (Table 21).

4 comments

We thank the reviewer for the feedback. We clarify the magnitude and scope of MolRIFT’s gains, its relation to existing molecular prediction methods, and the rationale for our evaluation and design choices.


W1/Q1. Magnitude and consistency of predictive gains.

To show the magnitude of these improvements at the task level, we have added the full per-task RMSE and MAE results, together with relative RMSE and MAE reductions against the global predictor and each control, to Appendix C.2, Tables 13–16 (pp. 26–29 of the revised manuscript).

Although MolRIFT’s mean RMSE gains over the global predictor and other configuration variants are modest, we consider its consistent improvements across most tasks meaningful. Moreover, mean RMSE alone does not show how broadly these gains occur, because larger losses on a few tasks can offset smaller gains across many others.

In structural generalization, more accurate predictions can help prioritize unfamiliar compounds for experimental testing. We therefore believe that improving predictions consistently across diverse tasks is as important as the absolute magnitude of improvement. MolRIFT shows this consistency against both the global predictor and the simpler controls, achieving lower RMSE and MAE on most of the 47 regression tasks in each comparison.


W1/Q2. Choice of correction target.

We chose error differences as MolRIFT’s prediction target because they directly express our goal of correcting the global predictor’s remaining error. The target is the observed property difference minus the difference predicted by the global model. This lets the relation predictor start from a similar anchor’s known error and estimate how much that error changes for the query.

Plus, mean RMSE in Appendix C.4 can be influenced by a large performance gap on a few tasks. For example, excluding just CHEMBL2835_Ki reverses the original comparison, with error differences yielding a lower mean RMSE (0.890483 vs. 0.891337). To complement this aggregate comparison, the table below reports the number of tasks on which error differences yield lower RMSE or MAE than property differences.

BenchmarkLower RMSELower MAE
MoleculeACE17/3018/30
Polaris regression12/1715/17
All regression tasks29/4733/47

Although the error-difference target was prespecified independently of these test-set comparisons, the results also support its competitive predictive performance. We have added the per-task RMSE and MAE results for both targets to Appendix C.4, Tables 19–20.


W2/Q3. Baseline selection and the evaluation protocol.

We did not initially include the full set of MoleculeACE reference algorithms because we prioritized recent methods such as MiniMol and CheMeleon for their strong predictive performance. To provide the direct comparison you requested, we have now evaluated the missing reference algorithms under our primary structural split. MolRIFT achieves lower mean RMSE and MAE than all eight baselines in the table below. This comparison covers all 30 MoleculeACE tasks, with MolRIFT using CheMeleon + TabPFN-3.

MethodMean RMSEMean MAERMSE W/LMAE W/L
MolRIFT0.884960.69665——
SVM (archived)0.950260.7577928/227/3
XGBoost (archived)0.924060.7348126/427/3
kNN1.017560.7926529/128/2
GBM0.973220.7765528/228/2
GCN1.127660.9141730/030/0
GAT1.149230.9391830/030/0
MPNN1.084720.8764029/129/1
AttentiveFP1.097470.8837030/030/0

The SVM and XGBoost rows reuse existing runs under the same primary structural split, while the other six baselines were evaluated for this comparison.

Regarding the evaluation split, we chose our protocol to study structural generalization, not because we consider MoleculeACE’s official split flawed. The official protocol splits molecules within structural clusters, preserving structural similarity between training and test sets. Our question is whether a model can predict properties for unseen chemical series when close labeled analogues are unavailable. We therefore combine scaffold holdout with a fingerprint-similarity cutoff to enforce both scaffold separation and limited similarity between training and test molecules.

The gains also extend beyond this split construction. With Morgan-based anchor selection unchanged, MolRIFT outperforms the global predictor on 38/58 tasks under each of the ECFP4- and MACCS-based DataSAIL partitions (Section 5.2, p. 7 of the revised manuscript).

W2/Q4. Relation to activity-cliff research.

Our study addresses prediction for structurally unfamiliar chemical series, while activity-cliff research focuses on distinguishing structurally similar molecules with sharply different properties. MolRIFT uses molecular pairs to estimate differences in the global predictor’s errors and correct predictions for unfamiliar molecules. Because these molecules can also belong to activity-cliff pairs, the analysis in Appendix E.2 (p. 39 of the revised manuscript) examines whether correction also helps on this subset within our structural split. This analysis compares MolRIFT only with the global predictor, so the reported cliff-subset RMSE reduction does not establish an advantage over methods trained specifically to discriminate activity cliffs.

To clarify how this analysis relates to activity-cliff research, we have revised the Purpose and setup paragraph in Appendix E.2 to cite Maggiora [1] and Yang et al. [2] and explain the scope of the comparison.

References

[1] Maggiora, G. M. On Outliers and Activity Cliffs—Why QSAR Often Disappoints. Journal of Chemical Information and Modeling, 46(4), 1535, 2006.

[2] Yang, H., Yao, Q., and Kwok, J. Curriculum-aware Training for Discriminating Molecular Property Prediction Models. ICLR, 2025.


W2/Q5. Why correction can hurt under DrugOOD.

A possible reason is that error relationships learned from training pairs become less reliable when assay conditions or structure–property relationships change. Our primary splits test new scaffolds within each property task. DrugOOD additionally includes assay and size shifts. These changes can affect the global predictor’s errors in ways that the training pairs do not capture.

Because the 0.20 threshold checks only structural similarity, it can still admit anchors whose errors are not useful for correcting the query under these shifts. The IC50 results are consistent with this explanation. Mean anchor similarity is nearly identical across shifts, but correction improves AUC under the size shift and lowers it under the assay and scaffold shifts.

IC50 shiftMean anchor similarityAUC change [95% CI]
Assay0.410-0.00287 [-0.00415, -0.00155]
Scaffold0.409-0.00613 [-0.00728, -0.00498]
Size0.406+0.00344 [0.00225, 0.00463]

Q6. How were the default settings chosen?

The anchor cap of 32 and the gating settings were prespecified and shared across tasks, independently of these test-set comparisons. As suggested, we have now explicitly stated this in Section 4.3 (p. 6 of the revised manuscript).


Q7. Reliability signals beyond similarity.

We agree that a query-level reliability signal beyond similarity could strengthen MolRIFT. Validation data could help identify signals that predict useful corrections and determine how they should control correction strength. At the same time, our design aims to preserve adaptation without task-specific tuning, a key advantage of tabular in-context learning. We therefore used common correction settings across tasks and minimized reliance on validation-based search. Developing such a signal without task-specific tuning is therefore a promising direction for future work.

Thanks for the detailed reply. I find the additional discussion on other baselines and AC-related research insightful. I increase my score above the acceptance bar.

Thank you for reconsidering the paper. Your questions helped us clarify the evaluation and its relation to activity-cliff research, and we’re glad the additional comparisons and discussion were helpful.

Reviewer_QDG2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MolRIFT, a two-stage tabular in-context learning framework for molecular property prediction under structrual shifts. MolRIFT uses a frozen TFM to predict molecular properties at first, and uses a second frozen TFM to predict pairwise differences in prediction errors to refine the initial predictions through similarity-weighted corrections. Across 58 tasks, MolRIFT improves its global TFM baseline on 46 tasks, with consistent gains across different molecular representations and TFM backbones. The paper also provides extensive ablations and evaluations under alternative distribution shifts.

Critical strengths
  1. Well-motivated practical problem: Generalization to OOF molecules is important in molecular discovery.
  2. Broad experimental setup: The paper evaluates 58 tasks, multiple structural split protocols, four molecular experimental evaluations, and three TFM backbones. And the ablations include meaningful alternatives.
  3. Clean writing.
Critical weaknesses
  1. Limited novelty: Pairwise molecular property prediction is already established in PADRE and DeepDelta, combining embeddings with frozen TFMs is a simple and general usage. MolRIFT's main contribution appears to be integrating these ideas through OOF residual correction and heuristic aggregation, which is a largely incremental engineering design.
  2. Modest improvements in native metrics: MolRIFT improves the global predictor on 46/58 tasks, but the mean MoleculeACE RMSE improvement is approximately 0.9%. Against the kNN-residual, shuffled-relation-label, and two-TFM-ensemble controls, the RMSE-reduction intervals all include zero (Table 3), so the learned pairwise relation adds little in magnitude beyond simpler residual smoothing.
  3. The advantage under genuinely difficult structural shifts is less convincing: The approach relies on molecular similarity and attenuates corrections for dissimilar anchors. At the strictest similarity cutoff (0.40), the RMSE improvement is only 0.0013.
Questions for authors
  1. see weaknesses.
  2. Given that property-difference prediction performs comparably to residual-difference prediction, what is the essential advantage of the proposed residual formulation beyond integrating pairwise examples into a second TFM?
  3. Does the method provide meaningful benefits specifically for the most structurally unfamiliar molecules? Can the authors better characterize when relational correction helps versus when it merely introduces additional error?
  4. For binary classification, Eq. 10 adds an unbounded correction to the predicted probability, and the paper describes no clipping or link function. How are corrected outputs constrained to valid probabilities?
4 comments

We thank the reviewer for the feedback. We clarify MolRIFT’s contribution beyond existing pairwise prediction methods and the magnitude and scope of its benefits under structural generalization.


W1/Q1. Novelty and the comparison with simple PADRE+TFM approach.

MolRIFT’s novelty lies in constructing molecular-relation contexts for error-difference prediction and using them to correct a strong molecule-level TFM under structural generalization. As you note, and as we discuss in our Related Work section, pairwise molecular prediction and the use of molecular embeddings with TFMs are established research directions. However, these ideas alone do not specify how to construct relation contexts, what the relation model should predict, or how its predictions should improve the global model. We address these questions through the proposed context and correction design, together with systematic experiments and analyses of its effectiveness and limitations.

To test whether directly combining an existing pairwise method with a TFM is sufficient, we implemented PADRE-TFM with both RDKit2D and CheMeleon representations.

MethodMoleculeACE RMSEMolRIFT RMSE W/L (47 tasks)
PADRE-TFM RDKit2D0.9510442/5
PADRE-TFM CheMeleon0.9094336/11
MolRIFT, similarity-only0.88505Reference

MolRIFT outperforms both variants, showing the value of how we construct and use molecular relations beyond directly combining PADRE with a TFM.


W2/Q1. Magnitude and consistency of predictive gains.

Although MolRIFT’s mean RMSE gains over the global predictor and other configuration variants are modest, we consider its consistent improvements across most tasks meaningful. Moreover, mean RMSE alone does not show how broadly these gains occur, because larger losses on a few tasks can offset smaller gains across many others.

In structural generalization, more accurate predictions can help prioritize unfamiliar compounds for experimental testing. We therefore believe that improving predictions consistently across diverse tasks is as important as the absolute magnitude of improvement. MolRIFT shows this consistency against both the global predictor and the simpler controls, achieving lower RMSE and MAE on most of the 47 regression tasks in each comparison.

ComparatorRMSE W/LMAE W/L
Global predictor40/740/7
Two-global-TFM ensemble38/937/10
kNN residual34/1335/12
Shuffled relation labels34/1335/12

We have added the full per-task RMSE and MAE results in Appendix C.2, Tables 13–16 (pp. 26–29 of the revised manuscript).


W3/Q1. Benefit under the strictest structural split.

We agree that the mean RMSE improvement becomes smaller as the similarity cutoff is reduced from 0.60 to 0.40. At the 0.40 cutoff, the mean MoleculeACE RMSE reduction is 0.0013, although MolRIFT still outperforms the global predictor on 38 of the 52 retained tasks. As discussed above, we consider consistent improvements across diverse tasks valuable for predicting unfamiliar chemical series.

We initially chose the 0.60 cutoff based on prior work using Tanimoto-based splits in reaction prediction [1, 2]. We have now added this explanation to Section 4.1 (p. 6 of the revised manuscript). However, as you point out, the gains become smaller below our default cutoff, and making relational correction more effective under more severe structural shifts remains an important direction for future work.

References

[1] Kovács, D. P., McCorkindale, W., and Lee, A. A. Quantitative interpretation explains machine learning models for chemical reaction prediction and uncovers bias. Nature Communications, 12, 1695, 2021.

[2] Wang, Y. et al. Retrosynthesis prediction with an interpretable deep-learning framework based on molecular assembly tasks. Nature Communications, 14, 6155, 2023.

Q2. Choice of residual differences.

We chose residual (error) differences as MolRIFT’s prediction target because they directly express our goal of correcting the global predictor’s remaining error. The target is the observed property difference minus the difference predicted by the global model. This lets the relation predictor start from a similar anchor’s known error and estimate how much that error changes for the query.

Plus, mean RMSE in Appendix C.4 can be influenced by a large performance gap on a few tasks. For example, excluding just CHEMBL2835_Ki reverses the original comparison, with error differences yielding a lower mean RMSE (0.890483 vs. 0.891337). To complement this aggregate comparison, the table below reports the number of tasks on which error differences yield lower RMSE or MAE than property differences.

BenchmarkLower RMSELower MAE
MoleculeACE17/3018/30
Polaris regression12/1715/17
All regression tasks29/4733/47

Although the error-difference target was prespecified independently of these test-set comparisons, the results also support its competitive predictive performance. We have added the per-task RMSE and MAE results for both targets to Appendix C.4, Tables 19–20.


Q3. When does correction help structurally unfamiliar molecules?

Our analysis shows a net MAE reduction in the groups with similarity of 0.30 or higher, but no clear average benefit below 0.30. Using MoleculeACE test predictions from our main evaluation split, we grouped molecules by their similarity to the closest training molecule. For each group, we measured the net MAE reduction relative to the global predictor, including both decreases and increases in error and averaging across tasks.

Similarity to the closest training moleculeNet MAE reduction [95% CI]
Below 0.30−0.00107 [−0.00407, 0.00189]
0.30 to below 0.45+0.00947 [0.00051, 0.01936]
0.45 or above+0.01188 [0.00559, 0.01813]

The positive net reduction in the 0.30–0.45 group shows that correction can help even when the closest training molecule has limited similarity, with error reductions outweighing error increases on average. Below 0.30, however, the average change is small and the confidence interval includes zero, so a benefit is not established for the most structurally distant group.


Q4. Valid probabilities.

The clipping step was implemented but was omitted from Eq. 10. We have corrected Eq. 10 in Section 3.4 of the revised manuscript (p. 5) to match the implementation.

Thank you for the response and revisions. The PADRE-TFM comparisons, per-task RMSE/MAE results, and similarity-stratified analysis strengthen the empirical case for MolRIFT and clarify its limitations. The clipping clarification resolves my concern about Eq. (10).

I retain reservations about the incremental novelty, modest effect sizes, and uncertain benefits under severe structural shifts. The rebuttal has addressed several of my concerns, and I am now leaning toward acceptance. I am therefore increasing my score from 2 to 3.

Thank you for taking the time to read our responses and reconsider the paper. We’re glad the additional experiments were helpful. We appreciate your feedback, as your questions pushed us to explain more clearly what MolRIFT contributes and where its benefits remain limited.

Reviewer_5z2GOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes MolRIFT, a framework that combines a global tabular foundation model with a second model predicting pairwise error differences to refine molecular property predictions. The improvements are consistent under the primary structural splits.

Critical strengths

The gains extend across representations and TFM backbones.

Critical weaknesses
  1. The need for the specific correction design is not clearly established. For example, error differences vs property differences; similarity-only weighting, removing MMP priority, and 16 anchors vs default setting.
  2. The predictive gains are modest, while the inference cost is large. Mean MACE RMSE decreases from 0.893 to 0.885, while predicting 10,000 molecules takes about 229s vs 7s for the global baseline. The global baseline also performs better on DrugOOD.
Questions for authors
  1. Given the comparable performance of the property differences setting, what advantage motivates predicting error differences?
  2. Can a simplified configuration, such as a model with fewer anchors and similarity-only weighting, retain the main gains while reducing inference cost?
  3. What explains the lack of consistent improvement under DrugOOD and chronological shifts? When does the correction help?
3 comments

We thank the reviewer for the feedback. We clarify our correction design, the value of its predictive gains relative to the additional computation, and the conditions under which correction helps.


W1/Q1. Why error differences, and why these component settings?

We chose error differences as MolRIFT’s prediction target because they directly express our goal of correcting the global predictor’s remaining error. The target is the observed property difference minus the difference predicted by the global model. This lets the relation predictor start from a similar anchor’s known error and estimate how much that error changes for the query.

Plus, mean RMSE in Appendix C.4 can be influenced by a large performance gap on a few tasks. For example, excluding just CHEMBL2835_Ki reverses the original comparison, with error differences yielding a lower mean RMSE (0.890483 vs. 0.891337). To complement this aggregate comparison, the table below reports the number of tasks on which error differences yield lower RMSE or MAE than property differences.

BenchmarkLower RMSELower MAE
MoleculeACE17/3018/30
Polaris regression12/1715/17
All regression tasks29/4733/47

Although the error-difference target was prespecified independently of these test-set comparisons, the results also support its competitive predictive performance. We have added the per-task RMSE and MAE results for both targets to Appendix C.4, Tables 19–20.

The default weighting, MMP priority, and anchor count were also prespecified as practical choices for constructing and combining molecular-relation contexts. We view the comparable performance with similarity-only weighting, without MMP priority, and with fewer anchors as further support for our core idea of using molecular-relation contexts to correct the global predictor’s remaining errors.


W2/Q2. Predictive gains, inference cost, and simpler configurations.

Although MolRIFT’s mean RMSE gains are modest, they are consistent across tasks, and we consider the additional computation worthwhile when evaluated over the full workflow. In molecular property prediction under structural generalization, the method must handle unfamiliar chemical series across endpoints with different structure–property relationships. Consistent gains across endpoints therefore provide important evidence of usefulness alongside the average improvement. Default MolRIFT achieves lower RMSE on 40/47 regression tasks and lower MAE on 40/47 tasks than the global baseline, showing that its benefits extend across a broad range of endpoints.

Although MolRIFT adds inference cost, it achieves higher Elo with less total computation than the evaluated supervised methods with hyperparameter optimization and the fine-tuned baseline (Figure 4, p. 9). These results motivate the default configuration of MolRIFT as a practical way to obtain strong predictive performance without task-specific hyperparameter or configuration search.

A configuration with eight anchors and similarity-only weighting retains the mean MoleculeACE RMSE gain while reducing measured relation-stage time. We jointly tested these changes on the 286 applicable splits across all 58 tasks, keeping the other settings fixed. Both configurations were evaluated using the same CheMeleon features, global predictions, and TabPFN-3 checkpoint. The default MolRIFT value below comes from this additional matched evaluation.

ConfigurationMoleculeACE RMSEMedian relation-stage query time (s)
Global0.89315N/A
Default MolRIFT0.885091.99
8 anchors, similarity-only0.884621.38

These results suggest that our core idea of correcting global predictions through molecular-relation contexts remains effective across different component settings. This flexibility allows MolRIFT to be adapted to scenarios where inference efficiency is a priority.


W2/Q3. DrugOOD, chronological shifts, and when correction helps.

A possible reason is that error relationships learned from training pairs become less reliable when assay conditions or structure–property relationships change. DrugOOD includes assay and size shifts, while chronological splits separate earlier from later measurements. These changes can affect the global predictor’s errors in ways that the training pairs do not capture.

In our experiments, correction helps when predicting molecules with unseen scaffolds for the same property task and when ranking similar compounds in lead optimization. In the first setting, MolRIFT outperforms the global predictor on 46/58 tasks, even with limited train–test fingerprint similarity. In the second setting, the official Lo-Hi results show an increase in mean cluster Spearman from 0.3004 to 0.3233 (Appendix D.2).

Thank you for the response. It alleviates my concerns, and I maintain my positive rating.

Thank you for taking the time to read our response. We’re glad the explanations and the additional experiment with a simpler configuration helped address your concerns.

Reviewer_QeEFOpenReview ↗
Rating
4(+1)
Confidence
3(+1)
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper focuses on the issue of structural generalization in molecular property prediction. The author proposes MolRIFT, a relationship inference framework based on a frozen table-based model. This method first uses the molecular encoder and Global TFM for the initial property prediction, and then calculates the residuals of the training molecules through five-fold Out-of-Fold prediction. Subsequently, the author constructs a relationship context consisting of molecular pairs and their prediction error differences, and uses the second frozen TFM to infer the error changes of the test molecules relative to the reference molecules. Finally, the initial predictions are corrected through molecular similarity, uncertainty weighting, and reliability gating. The method is validated on 58 tasks in MoleculeACE and Polaris. MolRIFT outperforms Global TFM in 46 of these tasks, and shows consistent performance improvement across multiple combinations of molecular encoders and table-based models.

Critical strengths

I really appreciate the research approach of the author. Currently, many studies are focusing on designing increasingly complex architectures. I really like the author's "back to basics" approach, which predicts different new molecules based on the known relative relationships. The author also verified their design through a wealth of experiments.

Critical weaknesses

However, the results in C.4 also demonstrated that the differences in prediction properties could achieve the same or even better performance. So why did the author design the current architecture? Then the author designed too many key components, which is a very engineering approach. It is hoped that future research can design new ideas to continue this line of thinking.

Questions for authors

I would like the author to explain why it is necessary to predict the error differences. Additionally, as demonstrated in E.1, the improvement ratio per sample for the regression task does not monotonically increase with similarity. Moreover, Appendix C.7 does not prove that adaptive gating is significantly better than moderate fixed coefficients. This indicates whether the reliability assumption based on similarity holds. Although I have these concerns, I still believe that the quality of this paper is trustworthy.

3 comments

We thank the reviewer for the feedback. We clarify the rationale for predicting error differences and the role of similarity in controlling correction strength.


W/Q. Choice of error differences.

We chose error differences as MolRIFT’s prediction target because they directly express our goal of correcting the global predictor’s remaining error. The target is the observed property difference minus the difference predicted by the global model. This lets the relation predictor start from a similar anchor’s known error and estimate how much that error changes for the query.

Plus, mean RMSE in Appendix C.4 can be influenced by a large performance gap on a few tasks. For example, excluding just CHEMBL2835_Ki reverses the original comparison, with error differences yielding a lower mean RMSE (0.890483 vs. 0.891337). To complement this aggregate comparison, the table below reports the number of tasks on which error differences yield lower RMSE or MAE than property differences.

BenchmarkLower RMSELower MAE
MoleculeACE17/3018/30
Polaris regression12/1715/17
All regression tasks29/4733/47

Although the error-difference target was prespecified independently of these test-set comparisons, the results also support its competitive predictive performance. We have added the per-task RMSE and MAE results for both targets to Appendix C.4, Tables 19–20.


Q. Similarity and reliability.

We use similarity as a reliability signal to limit correction strength when close reference molecules are unavailable. This does not necessarily imply that the probability of improvement increases monotonically with similarity. Anchor weights determine relative contributions, while the gate controls the overall correction strength. Although moderate fixed coefficients are also competitive, the gate provides a simple way to adjust correction strength to the references available for each query. To evaluate the benefit of limiting correction strength itself, we compare MolRIFT with and without the gate in Table 3 (p. 9). Removing the gate increases mean MoleculeACE RMSE by 0.03443, and default MolRIFT performs better on 45/58 tasks.

Thank you for your reply. You have solved all my problems. I will maintain a positive evaluation and increase my confidence.

Thank you for reading our response and following up. We’re glad we could resolve your concerns, and we appreciate your support.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary for the Area Chair

Dear Area Chair,

Thank you and the reviewers for the careful evaluation and discussion. We summarize the main concerns and how we addressed them through additional experiments, analyses, and manuscript revisions.

  • Limited novelty beyond combining pairwise prediction with TFMs (QDG2). To test whether directly combining existing ideas is sufficient, we implemented PADRE-TFM with RDKit2D and CheMeleon representations. MolRIFT outperforms both variants, supporting the value of its molecular-relation context and correction design for improving a strong global TFM.

  • Small native-metric gains relative to Elo and task wins (QDG2, x49W, jdcA). To address whether rankings overstate the improvements, we added per-task RMSE, MAE, and relative error reductions against the global predictor and controls to Appendix C.2, Tables 13–16. We acknowledge the modest mean gains while making their magnitude directly assessable alongside their consistency. MolRIFT lowers both RMSE and MAE on 40/47 regression tasks against the global predictor and improves over each simpler control on most tasks.

  • Justification for the correction target and default settings (QeEF, 5z2G, QDG2, x49W). Comparable alternatives prompted questions about our specific design. We clarified that error differences target what the global predictor has not captured. We also explicitly stated that the defaults were prespecified. Comparable performance across weighting, MMP-priority, and anchor-count variants supports the molecular-relation approach across different component choices.

  • Similarity as a reliability signal (QeEF). The reviewer questioned the similarity gate given non-monotonic improvement frequency and competitive fixed coefficients. We clarified that similarity limits correction strength when close references are unavailable, without requiring monotonic improvement probability. Removing the gate increases mean MoleculeACE RMSE by 0.03443, supporting this control while not establishing superiority over every fixed coefficient.

  • Additional computation versus simpler configurations or fine-tuning (5z2G, jdcA). To test whether the gains require the full configuration, we jointly evaluated eight anchors and similarity-only weighting. This preserves the mean MoleculeACE RMSE gain while reducing measured relation-stage time by approximately 31%. We also tested whether comparable computation would be better spent on lightweight global fine-tuning. MolRIFT wins on 48/58 tasks, with mean MoleculeACE RMSE of 0.88509 versus 0.94115 for the tested fine-tuning configuration.

  • Possible fitting of optimistic OOF residual patterns (jdcA). To examine whether relationships learned from smaller OOF errors transfer to held-out molecules, we confirmed the OOF–test error gap and measured how predicted corrections track actual test errors. MolRIFT increases the correction–error Spearman correlation by 0.0407 over kNN residual transfer and 0.0415 over shuffled-label controls, with both confidence intervals excluding zero. This supports useful information beyond directly transferring neighbor residuals.

  • Uncertain benefits under severe structural and other distribution shifts (5z2G, QDG2, x49W). To characterize when correction helps, we measured net query-level MAE benefits across similarity groups. Benefits are evident at nearest-training similarity of 0.30 or higher, but not established below 0.30. Together with the inconsistent DrugOOD and chronological results, this identifies the demonstrated scope and where more reliable correction remains needed.

  • Missing reference baselines and positioning relative to activity-cliff research (x49W). We expanded the comparison to eight conventional and GNN baselines across all 30 MoleculeACE tasks under our structural split. MolRIFT achieves lower mean RMSE and MAE than each. We explained that the split tests unseen chemical series without close labeled analogues, added activity-cliff literature, and clarified that the cliff-subset comparison against the global predictor does not establish superiority over specialized activity-cliff methods.

The discussion strengthened the evidence for molecular-relation contexts while clarifying the magnitude and limits of their benefits. MolRIFT provides a practical way to improve strong frozen TFMs through molecular comparisons, with consistent gains across molecular representations and TFM backbones without task-specific weight updates.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.