acceptodds
Under review as a conference paper at ICLR 2027

MDG-IS: An Iterative Sampling Framework For Theory-Experiment Data Alignment In Artificial Intelligence Chemistry

Abstract

AI4Chemistry faces a fundamental bottlewneck: high-fidelity experimental data are scarce and heterogeneous, while large low-fidelity theoretical datasets are abundant but subdomain-biased. This mismatch leads to distribution misalignment and frequent negative transfer, limiting the effectiveness of existing transfer learning and multi-fidelity approaches that assume global theory–experiment consistency. We propose Multi-SubDomain GMM-guided Iterative Sampling (MDG-IS), a fidelity-aware data selection framework that identifies which theoretical samples are actually useful for learning from real experimental data. MDG-IS models experimental subdomains through a Gaussian Mixture Model and performs local KL-based alignment to iteratively extract a compact, well-aligned subset of theoretical data. Across molecular, reaction, and materials benchmarks, MDG-IS markedly reduces computational cost, mitigates negative transfer, and consistently improves prediction accuracy in real scenarios.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI4Chemistry, Machine learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_kwvoOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper presents a sampling framework for the connection between theoretical and experimental data, which are of different fidelities. This is a very widespread problem within chemistry, where are larger quantity of lower accuracy theoretical data is easier to access than experimental data, which is hence of lower quantity. Methods that provide some route to "correct" the lower fidelity data to higher fidelity data are therefore of significant interest in chemistry.

The paper focuses on a method to reduce the amount of the lower fidelity data that is required for the accurate prediction, also identifying data that is of lower quality. The approach is demonstrated across three different chemistry case studies. Ablation studies are carried out.

Critical strengths
  • Such an approach has potential application across a wide range of chemistry areas, being related to a general problem. This increases the audience for the contribution.
  • The authors have tested on three different case studies representing different sub-fields of chemistry, which again increases the audience for the contribution.
Critical weaknesses
  • The paper could be written to be easier for a reader to follow in places, there are a few typos, missing introductions of acronyms. Several of the captions could benefit from expansion for the readers' ease.
  • The three case studies are stated to relate to the combination of a larger computational dataset and a smaller experimental dataset. However, for the material band gap prediction, both datasets are computational I believe (assuming GWVD are high accuracy electronic structure calculations - the acronym is not introduced). The latter dataset however is referred to as "experimental", which is misleading.
Questions for authors

No further questions beyond the requests for clarifying edits to the paper made in the preceeding box.

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

The paper presents MDG-IS, a framework for selecting source data to improve prediction in a target domain with limited labeled data. It combines pretrained features, PCA, target subdomain modeling with Gaussian mixtures, KL-guided Thompson sampling, and iterative pruning. The selected subset is used for pretraining, followed by target-domain fine-tuning. Evaluation covers molecular formation enthalpy, drug–target affinity, solvation free energy, and material band gaps, with comparisons against data-selection baselines and component ablations.

Critical strengths
  1. A practically relevant problem. Selecting useful source data under distribution mismatch addresses a meaningful bottleneck in chemical modeling. Explicitly accounting for heterogeneous target subdomains provides a reasonable motivation for improving transfer with limited target labels.

  2. Promising empirical results against relevant selection baselines. The paper reports lower prediction errors while retaining 16.8%–63.3% of the source data. Comparisons include recent methods such as InfoBatch, MolPeg, and CSIE, providing useful evidence of potential benefits beyond random selection.

Critical weaknesses
  1. Unclear evaluation protocols and insufficient evidence of reliable improvements. The paper does not specify target training/validation/test splits or clarify whether target distribution estimation excludes held-out samples. Several target datasets are small, including only 52 GWVD entries, but actual test sizes are unreported. Only three repetitions are provided, and improvements on affinity, band-gap, and solvation prediction are small relative to the reported standard deviations. Consequently, test independence and the stability of the gains remain difficult to assess.

  2. The distinct methodological and practical contribution is insufficiently established. The method combines established density modeling, divergence matching, and bandit sampling, but the ablations change coreset sizes alongside selection mechanisms without documenting matched training budgets. This limits attribution of the gains to the proposed combination. Furthermore, the absence of a target-only baseline prevents establishing the claimed mitigation of negative transfer, while end-to-end computational savings are not demonstrated.

Questions for authors

mentioned in the weaknesses

3 comments

Thanks for your thoughtful and constructive feedback. We sincerely appreciate the time and effort you have invested in reviewing our manuscript. We are encouraged by your recognition of our work. At the same time, we acknowledge that some aspects require further clarification.


W1. Evaluation protocol and test independence

Target GMM, KL alignment, and iterative selection use only target train+validation embeddings; held-out test samples are never used for GMM fitting, KL scoring, threshold calibration, or coreset pruning. For NREL-GWVD (52), ATcT (358), and CASF-2016 (285), we use repeated stratified 5-fold cross-validation; for CombiSolv-Exp (8,780), we repeat experiments with three random seeds.

W1. Evidence of reliable improvements: Significance Analysis

Because comparisons are paired across splits, we report paired differences against the second-best method with 95% CIs. MDG-IS significantly outperforms the second-best method on all tasks.

methodEnthalpyDTACombiSolvBand Gap
second-best32.37$\pm_{1.49}$1.66$\pm_{0.14}$0.455$\pm_{0.057}$0.652$\pm_{0.028}$
ours30.36$\pm_{1.20}$1.61$\pm_{0.13}$0.436$\pm_{0.021}$0.628$\pm_{0.033}$
p-value7×10−47 \times 10^{-4}2×10−32 \times 10^{-3}0.0283×10−33 \times 10^{-3}

W2. Attribution analysis of the gains

Coreset size and selection mechanism vary together, so we add a fixed-size ablation keeping size consistent with the baseline. Under fixed size, full MDG-IS still outperforms all ablation variants on DTA and Band-gap, indicating component synergy; forcing a fixed size contradicts the auto-size design and degrades performance.

EnthalpyDTACombiSolvBand Gap
Ablationauto-sizefixed-sizeauto-sizefixed-sizeauto-sizefixed-sizeauto-sizefixed-size
w/o Subspace-weighted GMM32.54±1.2232.54\pm_{1.22}32.50±1.2432.50\pm_{1.24}1.83±0.081.83\pm_{0.08}1.98±0.091.98\pm_{0.09}0.467±0.0210.467\pm_{0.021}0.482±0.0210.482\pm_{0.021}0.761±0.0720.761\pm_{0.072}0.782±0.0700.782\pm_{0.070}
w/o KL-guided bandit selection31.79±1.3931.79\pm_{1.39}31.48±1.4231.48\pm_{1.42}1.69±0.091.69\pm_{0.09}1.71±0.091.71\pm_{0.09}0.452±0.0210.452\pm_{0.021}0.458±0.0210.458\pm_{0.021}0.732±0.0850.732\pm_{0.085}0.749±0.0810.749\pm_{0.081}
w/o Coreset purification34.36±2.0134.36\pm_{2.01}34.44±2.0434.44\pm_{2.04}1.89±0.121.89\pm_{0.12}2.06±0.122.06\pm_{0.12}0.515±0.0330.515\pm_{0.033}0.514±0.0350.514\pm_{0.035}0.951±0.1210.951\pm_{0.121}0.938±0.1320.938\pm_{0.132}
vanilla method30.36±1.2030.36\pm_{1.20}32.15±1.2632.15\pm_{1.26}1.61±0.131.61\pm_{0.13}1.68±0.141.68\pm_{0.14}0.436±0.0210.436\pm_{0.021}0.476±0.0200.476\pm_{0.020}0.663±0.0810.663\pm_{0.081}0.680±0.0800.680\pm_{0.080}

W2. Baseline for target only

A target-only baseline performs reasonably only on CombiSolv, where data are sufficient, and degrades sharply elsewhere.

methodEnthalpyDTACombiSolvBand Gap
target only model57.80$\pm_{8.61}$2.45$\pm_{0.196}$0.535$\pm_{0.074}$1.348$\pm_{0.265}$
ours30.36$\pm_{1.20}$1.61$\pm_{0.13}$0.436$\pm_{0.021}$0.628$\pm_{0.033}$

W2. End-to-End Computational Efficiency

We compared the training efficiency—defined as the sum of data selection time and training time—of training on the full dataset, using 2nd-best method, and employing our proposed approach. It can be seen that our method demonstrates the best performance.

total timeEnthalpyDTACombiSolvBand Gap
Full dataset77.6min4.1h124.3h3.1min
InfoBatch31.5min1.7h39.9h6.3min
MDG-IS (ours)26.7min1.5h32.0h1.6min

All of the aforementioned changes will be indicated in the revised PDF using highlighting (red text). Your feedback is very important for us to improve our work. If our response fully addresses your concerns, we sincerely hope that you will consider modifying your rating. We are also happy to provide further clarification or continue the discussion.

Dear Reviewer,

We sincerely appreciate the time and effort you dedicated to reviewing our manuscript. Your thoughtful and constructive comments significantly contributed to improving the clarity and rigor of our work.

In our recent response, we have carefully addressed your concerns. We would be grateful to know if our replies have addressed your concerns. If there are any remaining issues or areas where you feel further clarification or improvement is needed, we would be more than happy to provide additional details or make further revisions to strengthen the manuscript.

 Thank you once again for your valuable feedback and for helping us improve the quality and clarity of our work.

 Best regards,

 The authors

Dear Reviewer,

The window for reviewing rebuttals and exchanging feedback is about to close. We encourage you to take this final opportunity to consider the authors’ responses and contribute any remaining insights to help inform the final decision.

We sincerely appreciate your attention and timely participation.

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

This paper proposes MDG-IS, a target-aware coreset selection framework for aligning theoretical and experimental chemistry data. The method combines subdomain-aware Gaussian mixture modeling, KL-guided Thompson sampling, and iterative pruning to select informative theoretical samples. Experiments across four chemistry tasks demonstrate promising improvements in prediction accuracy while reducing the training data size.

Critical strengths
  1. The paper addresses an important challenge in AI for chemistry: leveraging abundant theoretical data under distribution shifts and limited experimental supervision.

  2. The method achieves competitive performance across four tasks, supported by comparisons with various coreset selection methods, ablation studies, and additional backbone evaluations.

  3. Automatically determining the coreset size is a useful feature that may improve practical applicability.

Critical weaknesses
  1. The method description could be clearer. Eq. (6) appears to select the same maximum-density point for all subdomains, making it unclear how distinct anchors are obtained. The pruning score and threshold in Eq. (17) also lack sufficient justification.

  2. Some experimental results are inconsistent. For example, Tables 2 and 3 report different RMSE values for CombiSolv at the same coreset ratio. Moreover, end-to-end runtime comparisons with baselines are missing, making the efficiency gains difficult to assess.

  3. The method focuses on feature distribution alignment, but it remains unclear whether this is sufficient to handle systematic label discrepancies between theoretical and experimental data.

Questions for authors

See the weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.