acceptodds
Under review as a conference paper at ICLR 2027

CB7Mol: A Physicochemically Grounded AI-driven Model for CB[7] Affinity Prediction

Abstract

AI-driven scientific discovery requires more than predictive accuracy: models must remain reliable under scarce supervision while respecting underlying physicochemical principles. We study this challenge through cucurbit[7]uril (CB[7])-based host–guest binding affinity prediction, an important problem in molecular recognition for supramolecular chemistry. A fundamental challenge is to jointly acquire the molecular, geometric, and affinity knowledge required for grounded prediction despite limited CB[7]-specific data. We propose CB7Mol, a physicochemically grounded AI-driven model that strategically transfers complementary knowledge from compatible sources to bridge these three gaps: molecular encoder post-training acquires CB[7]-relevant chemical semantics while preserving geometry; host–guest complex pre-training learns binding geometry and conformational flexibility from resolved crystal structures; and cross-host calibration transfers affinity supervision from mechanistically compatible macrocyclic hosts. CB7Mol achieves the best overall average ranking across three CB[7] affinity evaluation sets. More importantly, mechanistic analyses show that its predictions reflect key physicochemical determinants of CB[7] recognition rather than dataset-specific correlations, supporting reliable discovery beyond known guest chemistries. Finally, an AI-to-experiment pipeline screens 304k molecules in GEOM-Drugs and prioritizes candidates for wet-lab validation. CB7Mol recovers established strong-binding scaffolds, uncovers physicochemically consistent candidates with novel scaffolds, and yields high-affinity guests validated by competition NMR with up to approximately 12.6, expanding the experimentally validated CB[7] chemical space and molecular toolbox for supramolecular design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular Representation Learning, Supramolecular Chemistry, AI for Chemistry

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_qp1VOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper aims to predict binding affinities for CB[7]-based host-guest complexes, which is a relatively unexplored domain in ML, and introduces CB7Mol. CB7Mol is a binding affinity prediction model that considers (1) molecular encoder training with alignments with chemical knowledge, (2) complex-level pre-training with flexibility, and (3) cross-host calibration. The results show improvements over the considered baselines. The authors also conduct wet-lab validation using competition NMR.

Critical strengths
  • Well-motivated methodology for an unexplored task. This paper addresses binding affinity predictions for an under-explored domain in ML: CB[7]. This moves beyond conventional protein-ligand binding affinity prediction and considers binding affinity between non-polymer entities. The proposed components, e.g., chemical knowledge alignment and binding-related pre-training, are well-motivated.

  • Wet-lab results. The proposed method shows clear improvements over considered baselines, and the individual components are well-supported by ablations. Moreover, the paper validates their predictions using wet-lab experiments.

Critical weaknesses
  • Lack of comparison with MD-based methods. Since CB[7] host-guest complexes are relatively small, molecular dynamics (MD)-based binding affinity computations should be considered. Although MD cannot be widely used for huge systems such as protein-ligand complexes, MD provides accurate binding affinity estimates, making it a feasible option for CB[7] host-guest complexes, which are typically smaller than conventional protein-ligand complexes.
Questions for authors
  • I was wondering how the prediction heads of the baselines are implemented. Do they share an architecture across different models?
  • The proposed method seems applicable to general molecular binding affinity prediction. Why do the authors focus only on CB[7]? Does it also work well for other host-guest systems?
  • I was wondering whether the performance is sensitive to the set of cross-hosts in cross-host calibration.
5 comments
AuthorsOpenReview ↗

Reply to Reviewer qp1V (Part 1)

We thank you for your constructive evaluation, and for crediting our well-motivated methodology and wet-lab validation. We address your questions below:

Q1: Prediction head implementation for baselines

We adopt the same prediction head architecture for the same kind of methods. aISS is not a learnable baseline, while QSAR methods are not based on neural networks, so there is no prediction head for physics-driven baselines. For CB7Mol and all other baselines, a two-layer MLP is adopted for prediction, with the same architectural design and different dimensional settings. The main difference lies in the readout design, which depends on different types of input. GNN-based baselines adopt the concatenation of molecular graph representations, molecular relational learning-based baselines concatenate Set2Set molecular representations following their official implementation, while complex-based baselines use the same readout strategy as CB7Mol. We believe that such implementation can facilitate both fair comparison and input utilization.

Q2: Why only focus on CB[7] and performance on other host-guest systems

Thank you for raising this question. We focus on CB[7] for two reasons:

(1) CB[7] is a uniquely important and challenging recognition target. Among synthetic hosts, CB[7] uniquely combines water solubility, biocompatibility, and exceptionally strong and selective binding, underpinning its use in targeted drug delivery and displacement-based biosensing [1]. CB[7] also forms the strongest known synthetic host–guest complexes [2], giving an affinity range that spans many orders of magnitude and governed by subtle physicochemical effects, which require specialized, per-molecule quantum-chemistry-derived electrostatics to capture [3], and therefore remain prohibitively expensive for large-scale screening. CB[7] affinity prediction is thus both high-impact and scientifically hard.

(2) Other macrocyclic hosts have different binding mechanisms, so the task does not transfer by design. For example, CB[8] routinely forms 1:2 ternary complexes that co-encapsulate two guests via charge-transfer stabilization [1], and coordination cages encapsulate guests in self-assembled metal-organic cavities, usually binding multiple guests [4]. Many existing hosts mechanistically fall outside the representational scope of CB7Mol.

We also run additional experiments to test the ability of CB7Mol to generalize to unseen hosts. We adopt 16 OA/TEMOA host-guest pairs from SAMPL6 [11] and 16 TrimerTrip (an acyclic cucurbituril) host-guest pairs from SAMPL7 [8] for evaluation, and use the same model checkpoint in CB[7]-related experiments. Results are summarized in the table below:

MethodSAMPL6 ρSAMPL6 τSAMPL6 RMSESAMPL7 ρSAMPL7 τSAMPL7 RMSE
CB7Mol0.711 ± .060.523 ± .061.374 ± .230.306 ± .200.220 ± .182.571 ± .53
GCN0.625 ± .030.407 ± .032.002 ± .33−0.616 ± .08−0.463 ± .061.973 ± .20
GIN0.579 ± .030.373 ± .051.672 ± .43−0.569 ± .11−0.407 ± .101.973 ± .33
MPNN0.583 ± .030.373 ± .051.946 ± .25−0.589 ± .04−0.450 ± .032.683 ± .26
MMGNN0.658 ± .040.463 ± .061.513 ± .43−0.022 ± .28−0.013 ± .212.174 ± .49
I2MOLE0.701 ± .030.537 ± .041.991 ± .42−0.119 ± .35−0.087 ± .262.134 ± .65
CIGIN0.544 ± .120.380 ± .102.972 ± .80−0.144 ± .26−0.093 ± .202.657 ± .46
CGIB0.629 ± .150.467 ± .161.682 ± .980.195 ± .270.143 ± .212.098 ± .52
CMRL0.540 ± .290.377 ± .191.563 ± .90−0.070 ± .25−0.017 ± .182.690 ± 1.16
EPT0.594 ± .140.447 ± .110.868 ± .140.286 ± .220.177 ± .193.818 ± .29
GET0.526 ± .060.400 ± .062.181 ± .400.425 ± .030.327 ± .031.230 ± .07
aISS0.068 ± .000.033 ± .00–0.056 ± .000.033 ± .00–

Despite belonging to a different host family, OA/TEMOA recognize guests through the same dominant driver as CB[7], which is the hydrophobic encapsulation of the guest. Therefore, CB7Mol can achieve competitive performance on SAMPL6. TrimerTrip, by contrast, is an open, flexible acyclic host, whose very different geometry alters the binding mechanism, and CB7Mol is outperformed by general-purpose foundation model like GET.

AuthorsOpenReview ↗

Reply to Reviewer qp1V (Part 2)

Q3: Sensitivity to the set of cross-hosts

The performance is sensitive to the cross-host set, and the sensitivity depends on the binding mechanism compatibility of cross-hosts used during calibration. We have reported the sensitivity of CB7Mol to different sets of training hosts in Figure 3(c). Empirically, using the full set of compatible hosts is consistently and substantially the best configuration across all three CB[7] evaluation sets. Removal of β-CD and p-SC4 causes severe performance degradation, because they each cover a dominant determinant of CB[7] recognition. β-CD presents a hydrophobic macrocyclic cavity of dimensions comparable to CB[7], while p-SC4 can reproduce cationic recognition similar to CB[7]. This directly reflects the design rationale of cross-host calibration described in Section 3.3, which is to learn the binding determinants shared across mechanistically compatible host–guest systems. The observed sensitivity is therefore expected, where performance improves as the compatible-host set more completely covers recognition determinants of CB[7].

W1: Comparison with MD baselines

Thank you for raising this constructive question. We have run substantial additional experiments. In brief: (1) we add an endpoint MM/GBSA baseline across all three CB[7] evaluation sets; (2) we attempted rigorous MD free-energy calculations and found them to suffer from both systematic bias and numerical instability on CB[7]; and (3) making such calculations reliable would require QM-grade force-field re-parameterization, whose per-molecule cost is prohibitive and cannot be completed within an acceptable timeframe.

(1) Endpoint baseline: MM/GBSA

We add MM/GBSA [5], the cheapest MD-based estimator, as a physics-based endpoint baseline with GAFF2 force field, AM1-BCC charges, and igb=8 implicit solvent, which is evaluated under the same protocol as aISS. We will add the new results in Table 1, which are reported here for your convenience:

Evaluation setMethodρτRMSE
SAMPL4MM/GBSA0.6160.4201.322
CB7Mol0.741±.120.585±.121.864±.12
ExpertMM/GBSA0.1790.1341.973
CB7Mol0.558±.060.427±.051.729±.11
SupraBenchMM/GBSA0.487 ± .100.360 ± .082.106 ± .51
CB7Mol0.834±.120.691±.121.253±.35

On the small SAMPL4 set, MM/GBSA shows advantage in prediction precision. For other cases, CB7Mol consistently outperforms simple MD baseline.

(2) Rigorous MD free-energy attempts

We attempted absolute binding free-energy calculations using two standard approaches, both built on the standard small-molecule stack (GAFF2 typing, AM1-BCC charges via antechamber, TIP3P explicit water):

  • Attach-Pull-Release (APR) [6] with Boresch 6-DOF orientational restraints;
  • Alchemical double-decoupling (DDM) [7] with Hamiltonian replica exchange.

Testing on CB[7] guests from the SAMPL4 benchmark and high-affinity binders reported in our case study, we observed two failure modes:

(a)Systematic bias. Alchemical method tends to predict the wrong sign of binding free-energy. The converged binding affinity estimates ranged from +2.7 to +11.7 kcal/mol, while the experimental results lie between −6 and −17 kcal/mol. On the other hand, formal charge-related over-estimation of binding is observed in APR calculation. For example, the +2 adamantane-diammonium is estimated at ΔG_bind = −43.7 kcal/mol versus experimental −17.2, and the +1 azaspiro at −29.9 versus −14.2.

(b) Numerical instability and convergence failure. The APR calculations were numerically unstable. We summarize the failure statistics on SAMPL4:

OutcomeCountAnalysis
NaN5MD integrator blows up during the pull
Non-convergent pull5physically impossible ΔG_bind of −300 to −800 kcal/mol

Both failure modes trace to the inappropriate charge modeling. CB[7] recognition is dominated by ion-dipole interactions and the release of high-energy water, while fixed-charge force field with library charges (AM1-BCC) captures neither the polarization response of the portal carbonyls to the guest charges nor the delicate energetics of confined high-energy water [3,8]. This electrostatic mis-balance produces the systematic bias and numerical instability.

AuthorsOpenReview ↗

Reply to Reviewer qp1V (Part 3)

(3) A reliable fix requires QM-grade force fields with prohibitive per-molecule cost

Correcting the aforementioned failure modes requires deriving the electrostatic forces from per-molecule quantum mechanics, such as QM-density fixed charges (MBIS [3]) and Polarizable force field (AMOEBA [9]). However, these methods require per-molecule QM parameterization, where every new candidate needs its own QM calculation to obtain its parameters. We empirically measured the single-run AMOEBA parameterization cost:

MoleculeChargeTime CostDisk Requirement
Adamantan-1-amine+110–45 min~ 0.2 GB
1,6-Hexanediamine+28.6 h~ 1 GB
p-Xylylenediamine+221.2 h~ 1 GB
2-(2-Aminoethyl)piperazine+3> 48 h~ 1 GB
CB[7]0> 90 h> 150 GB

Cost grows exponentially with molecular size and charge/flexibility, which is driven by cubic time complexity, harder SCF convergence and per-rotatable-bond torsion scans [9,10]. A reliable MD pipeline for CB[7] would need this per-molecule QM parameterization plus a polarizable free-energy simulation for every candidate. It is infeasible to obtain even a handful of well-tuned results within the rebuttal period.

Finally, we would also like to clarify why we didn’t take MD baselines into account in the first place. CB7Mol targets large-scale guest screening, which does not require high prediction accuracy. In contrast, MD free-energy methods are computationally expensive, and are conventionally used to accurately calculate the binding free-energy of a small number of candidates after faster virtual screening. Our experiments further proved that rigorous MD is not as reliable as we have expected for CB[7]-based host-guest system. We therefore do not consider MD a baseline worth massive investment for the large-scale guest screening problem CB7Mol addresses.

We hope that our clarifications and additional experiments will lead to a reconsideration of your score.

Reference:

[1] Kraft, Albert, Connor R. Schmidt, Christian Akakpo, Frank Biedermann, Matthew J. Webber, and Lyle Isaacs. "Cucurbit [n] uril-type molecular containers." Nature Reviews Methods Primers 6, no. 1 (2026): 40.

[2] Alnajjar, Mohammad A., Werner M. Nau, and Andreas Hennig. "A reference scale of cucurbit [7] uril binding affinities." Organic & biomolecular chemistry 19, no. 39 (2021): 8521-8529.

[3] Gonzalez, Duvan, Luis Macaya, Carlos Castillo-Orellana, Toon Verstraelen, Stefan Vogt-Geisse, and Esteban Vohringer-Martinez. "Nonbonded force field parameters from minimal basis iterative stockholder partitioning of the molecular electron density improve CB7 host–guest affinity predictions." Journal of Chemical Information and Modeling 62, no. 17 (2022): 4162-4174.

[4] Rizzuto, Felix J., Larissa KS von Krbek, and Jonathan R. Nitschke. "Strategies for binding multiple guests in metal–organic cages." Nature Reviews Chemistry 3, no. 4 (2019): 204-222.

[5] Genheden, Samuel, and Ulf Ryde. "The MM/PBSA and MM/GBSA methods to estimate ligand-binding affinities." Expert opinion on drug discovery 10, no. 5 (2015): 449-461.

[6] Henriksen, Niel M., Andrew T. Fenley, and Michael K. Gilson. "Computational calorimetry: high-precision calculation of host–guest binding thermodynamics." Journal of chemical theory and computation 11, no. 9 (2015): 4377.

[7] Boresch, Stefan, Franz Tettinger, Martin Leitgeb, and Martin Karplus. "Absolute binding free energies: a quantitative approach for their calculation." The Journal of Physical Chemistry B 107, no. 35 (2003): 9535-9551.

[8] Amezcua, Martin, Léa El Khoury, and David L. Mobley. "SAMPL7 Host–Guest Challenge Overview: assessing the reliability of polarizable and non-polarizable methods for binding free energy calculations." Journal of computer-aided molecular design 35, no. 1 (2021): 1-35.

[9] Walker, Brandon, Chengwen Liu, Elizabeth Wait, and Pengyu Ren. "Automation of AMOEBA polarizable force field for small molecules: Poltype 2." Journal of computational chemistry 43, no. 23 (2022): 1530-1542.

[10] Smith, Daniel GA, Lori A. Burns, Andrew C. Simmonett, Robert M. Parrish, Matthew C. Schieber, Raimondas Galvelis, Peter Kraus et al. "PSI4 1.4: Open-source software for high-throughput quantum chemistry." The Journal of chemical physics 152, no. 18 (2020).

[11] Rizzi, Andrea, Steven Murkli, John N. McNeill, Wei Yao, Matthew Sullivan, Michael K. Gilson, Michael W. Chiu et al. "Overview of the SAMPL6 host–guest binding affinity prediction challenge." Journal of computer-aided molecular design 32, no. 10 (2018): 937-963.

Reviewer_tkLZOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper presents CB7Mol, a framework for predicting the binding affinity of guest molecules to CB[7], a synthetic host molecule. The framework has three stages: a molecular encoder, an atom-pair interaction encoder, and a cross-host adaptation.

The authors evaluate a number of different baselines and report different results.

They also test on 304,000 molecules from GEOM-Drugs, and test compounds using NMR experiments. They report success and failures.

Critical strengths

Combining multiple kinds of supervision together seems to be a good approach. The evaluation is extensive, though I qualify this in the Weaknesses. The experimental validation is strong, and the documentation of failures is appreciated.

Critical weaknesses

The primary weakness of this paper is lack of a comparison against modern NNP methods. Such methods have shown good results.

Théo Jaffrelot Inizan, Thomas Plé, Olivier Adjoua, Pengyu Ren, Hatice Gökcan, Olexandr Isayev, Louis Lagardère, Jean-Philip Piquemal; Scalable hybrid deep neural networks/polarizable potentials biomolecular simulations including long-range effects. Chem. Sci. 2023; 14 (20): 5438–5452. https://doi.org/10.1039/d2sc04815a

Includes a CB8 complex.

The authors make an effort to include many baselines, but some of the choices may not be that meaningful. For example, GCN and MPNN are architectures, not models. A given architecture can give rise to many different models, depending on how it's trained.

I found the paper to be somewhat hard to read.

Questions for authors
  1. How does CB7Mol compare against modern NNP-based binding affinity techniques, especially the existing ANI-2X/AMOEBA results on SAMPL4? The comparison can include both accuracy and computational cost.
  2. How were GCN, MPNN, and other general-purpose architectures trained? Were they sufficiently optimized, etc. so as to provide a meaningful comparison?
  3. Can the authors demonstrate that self-distillation actually preserves geometric information and that geometry pretraining improves physically meaningful binding representations, rather than merely acting as additional regularization? Comparisons against simpler regularization and representation-preservation alternatives would help to show this.

I rated this paper as a 1, but I feel more it is like a 1.5.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer tkLZ (Part 1)

We thank you for your careful evaluation, and we appreciate your recognition of our methodology design, the extensive and strong experimental validation, and the documentation of failure cases. We address your questions below:

Q1: Comparison with ANI-2X/AMOEBA

Thank you for raising this question. According to the published results, ANI-2X/AMOEBA achieved RMSE 0.94 kcal/mol, which is approximately 0.69 logKa. We fully acknowledge that NNP-based methods represent state-of-the-art methods for absolute binding free-energy computation. However, the SAMPL dataset used in ANI-2X/AMOEBA is not the same as the SAMPL4 dataset used in our paper, and local re-implementation is required for fair comparison.

However, ANI-2X/AMOEBA free-energy calculation requires per-molecule AMOEBA parameterization. Every host and every guest must first be parameterized with poltype2-style workflows: QM-derived distributed multipoles, atomic polarizabilities, and QM torsion scans. This is a quantum-chemical procedure whose cost scales exponentially with molecular size, flexibility and formal charges [1, 2]. Our empirical tests show that parameterizing a single charged and flexible guest can take 8-48 hours, and parameterizing the CB[7] host itself costs about one week. Crucially, these runtimes reflect only a single automated parameterization pass, and they do not yet include the additional cost of per-molecule parameter tuning [1]. The optimal parameters differ for every molecule, so each must be fit and validated individually. Taken together, the re-implementation over our evaluation sets is infeasible within the rebuttal period.

More fundamentally, ANI-2X/AMOEBA and broader MD-related methods address a different problem from the one CB7Mol targets. You may refer to another response of ours for more details: https://openreview.net/forum?id=jkI4d6JE11&noteId=6NOJO0TKfE. CB7Mol targets large-scale virtual screening, which aims to be scalable. In contrast, ANI-2X/AMOEBA and other MD-based methods aims to produce high-accuracy absolute ΔG for the input complex, which is not designed for efficient virtual screening. We therefore view ANI-2X/AMOEBA as complementary rather than competing, and we believe that reproducing this baseline is not a worthwhile investment of the substantial time and computational resources due to task mismatch.

Q2: How were GNN-based general architectures trained

We agree that a weak training setup would make the comparison with general architectures uninformative, and we clarify our training protocol and hyperparameter tuning process.

All general-purpose GNNs are trained and evaluated using the same data setting as CB7Mol calibration stage, and they are all equipped with the same prediction head architecture as CB7Mol. We use the standard PyG implementations of GCN, GIN and MPNN, where the input molecular graphs are reduced to heavy-atom graphs, and each atom node carries a 19-dimensional feature vector, including a 15-way element one-hot plus degree, hydrogen count, aromaticity, and ring membership. Molecular graph embeddings are derived from atom-wise average pooling. The number of GNN layers ({1,2,3}), the hidden layer dimension ({32,64,128}), and host-guest output fusion strategy ({concatenation, addition, MLP fusion}) are properly tuned. The GNN results reported in Table 1 are the best results we have tested.

Q3: Whether self-distillation actually preserves geometric information.

Thank you for raising this constructive question. We ran two additional experiments to address the two underlying concerns: (1) whether self-distillation (SD) preserves geometric information, and (2) whether it contributes beyond regularization.

(1) Self-distillation preserves geometric information.

We train a ridge regression model on the frozen mean-pooled Uni-Mol heavy-atom embeddings to recover five 3D-geometry descriptors, including radius of gyration, asphericity, acylindricity, relative shape anisotropy and span. We train the regressor on all the SupraBench data and test on Expert set. We report R2 and MAE in standardized units of CB7Mol, CB7Mol w/o chem align, CB7Mol w/o SD and publicly released Uni-Mol in the table below:

EncoderR² ↑MAE ↓
Public Uni-Mol0.6830.337
CB7Mol0.1080.570
w/o chem align0.6860.338
w/o SD−0.0140.669

Although chemical alignment is harmful for geometry preservation, self-distillation retains the molecular 3D-geometry information in CB7Mol.

AuthorsOpenReview ↗

Reply to Reviewer tkLZ (Part 2)

(2) Self-distillation is not merely regularization.

To test whether SD provides transferable representational value rather than a generic regularization, we replace SD with an L2-SP [3] penalty that anchors the encoder weights to the public Uni-Mol checkpoint, which is a simpler regularization term. We train this baseline under the identical protocol of CB7Mol and report the results of strongest L2-SP setting below:

DatasetMethodRMSEρτ
SupraBenchCB7Mol1.253±.350.834±.120.691±.12
SupraBenchw/ L2-SP1.407 ± .330.806 ± .070.655 ± .09
ExpertCB7Mol1.729±.110.558±.060.427±.05
Expertw/ L2-SP1.741 ± .050.473 ± .060.352 ± .05
SAMPL4CB7Mol1.864±.120.741±.120.585±.12
SAMPL4w/ L2-SP2.106 ± .300.490 ± .160.371 ± .11

We can see that weight-anchoring regularizer fails to generalize to unseen molecules and chemotypes. This shows that SD-based post-training facilitates the learning of transferable geometrical information.

We thank you again for pointing out potential issues in our paper writing, and we hope that our clarifications and additional experiments will lead to a reconsideration of your score.

Reference: [1] Walker, Brandon, Chengwen Liu, Elizabeth Wait, and Pengyu Ren. "Automation of AMOEBA polarizable force field for small molecules: Poltype 2." Journal of computational chemistry 43, no. 23 (2022): 1530-1542.

[2] Smith, Daniel GA, Lori A. Burns, Andrew C. Simmonett, Robert M. Parrish, Matthew C. Schieber, Raimondas Galvelis, Peter Kraus et al. "PSI4 1.4: Open-source software for high-throughput quantum chemistry." The Journal of chemical physics 152, no. 18 (2020).

[3] Xuhong, L. I., Yves Grandvalet, and Franck Davoine. "Explicit inductive bias for transfer learning with convolutional networks." In International conference on machine learning, pp. 2825-2834. PMLR, 2018.

AuthorsOpenReview ↗

Kind Reminder

Dear Reviewer,

As the discussion period is drawing to a close, we would like to check in on whether our response has resolved your concerns. We are happy to answer any further questions. If our response has addressed your concerns, we would be grateful if you would consider raising your rating. Thank you again for your time and thoughtful feedback.

Best Regards,

Authors

Reviewer_SPL2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces CB7Mol, a physicochemically grounded AI model for predicting cucurbit[7]uril (CB[7])–guest binding affinities under limited target-specific data. The authors develop a three-stage knowledge transfer framework comprising molecular encoder post-training to enhance chemical representations, host–guest complex pre-training to capture binding geometry and conformational flexibility, and cross-host affinity calibration to leverage binding data from mechanistically compatible macrocyclic hosts. CB7Mol achieves the best overall average ranking across three CB[7] affinity evaluation datasets. Furthermore, mechanistic analyses demonstrate that its predictions reflect key physicochemical determinants of molecular recognition rather than dataset-specific correlations. Finally, the authors screen approximately 304,000 molecules from GEOM-Drugs and experimentally validate promising candidates using competition NMR, identifying high-affinity guests with novel molecular scaffolds and logKa values up to approximately 12.6.

Critical strengths
  1. The paper proposes a systematic multi-stage knowledge transfer framework to address the scarcity of CB[7]-specific affinity data. By integrating molecular encoder post-training, host–guest complex pre-training, and cross-host affinity calibration, the framework incorporates complementary chemical, geometric, and affinity information from different data sources. This design provides a well-motivated strategy for improving host–guest affinity prediction under limited supervision.
  2. The proposed CB7Mol model demonstrates promising predictive performance and physicochemical interpretability. CB7Mol achieves the best overall average ranking across three CB[7] affinity evaluation datasets. Furthermore, the authors conduct mechanistic analyses to examine whether the model captures key physicochemical factors governing host–guest recognition. These analyses provide additional support for the chemical relevance of the learned representations beyond conventional predictive performance metrics.
  3. The study extends beyond computational benchmarking through prospective molecular screening and experimental validation. The authors screen approximately 304,000 molecules from GEOM-Drugs and prioritize promising candidates for competition NMR experiments. The identification of experimentally validated high-affinity guests, including candidates with novel molecular scaffolds and logKa values up to approximately 12.6, provides encouraging evidence of the framework's potential for discovering new CB[7]-binding molecules.
Critical weaknesses
  1. Insufficient transparency in benchmark construction. Although three CB[7] evaluation datasets are reported, their selection criteria, data distributions, and potential overlaps are not sufficiently analyzed.
  2. Uncertainty in dataset reliability. The Expert dataset contains 60 expert-assigned affinity labels rather than direct experimental measurements, raising concerns about label accuracy and uncertainty.
  3. Limited evaluation of generalization. The use of random five-fold cross-validation and a small external SAMPL4 dataset does not adequately establish generalization to structurally novel guests.
  4. Questionable necessity of a complex multi-stage framework. Despite the limited CB[7] affinity data, the proposed framework introduces multiple sophisticated training stages, including representation alignment, geometric pre-training, and cross-host calibration. Although these strategies may alleviate data scarcity, stronger comparisons with simpler, data-efficient models are needed to justify their necessity and additional complexity.
Questions for authors
  1. Dataset construction and reliability. Could the authors provide more details on the selection criteria and data distributions of the three CB[7] evaluation datasets? In particular, how was the reliability of the 60 expert-assigned affinity labels assessed?
  2. Generalization under limited data. Given the small number of CB[7] affinity measurements, could the authors evaluate model performance using scaffold-based splits or molecular similarity-controlled splits to better assess generalization to unseen chemical structures?
  3. Necessity of representation alignment. What is the specific motivation for aligning the 3D Uni-Mol representation with the 2D GROVER representation? Could the authors compare this strategy with simpler feature concatenation or fusion approaches?
  4. Model complexity and data efficiency. Considering the limited affinity data, could the authors compare CB7Mol with simpler models, such as frozen molecular encoders combined with shallow regression models, under identical training data conditions? Such comparisons would help justify the necessity of the multi-stage framework.
  5. Effectiveness of cross-host transfer. How was the mechanistic compatibility between CB[7] and other macrocyclic hosts established? Could the authors investigate the individual contributions of different host families and potential negative transfer effects?
  6. Validation of physicochemical knowledge. Could the authors further validate whether the learned geometric and flexibility representations correspond to experimentally observed host–guest interactions or binding thermodynamics?
  7. Interpretation of ablation results. Some ablated models achieve lower RMSE than the full model on certain evaluation sets. Could the authors explain these inconsistencies and clarify how the contributions of individual components should be interpreted?
  8. Experimental validation and screening efficiency. Could the authors report the total number of experimentally tested candidates, including unsuccessful cases, and compare the experimental hit rate with simpler computational screening approaches?
4 comments
AuthorsOpenReview ↗

Reply to Reviewer SPL2 (Part 1)

We thank you for your careful evaluation, and for crediting our well-motivated methodology design, promising predictive performance, physicochemical interpretability of our proposed framework and wet-lab evidence. We address your questions below:

W1 & W2 & Q1: Dataset construction and reliability assessment.

Thank you for raising this question. For benchmark selection criteria and reliability of Expert dataset, SAMPL4 and SupraBench are the only two existing benchmarks targeting CB[7] with proper scale, complete molecular profiles and experimentally validated binding affinity labels. To further enrich available data, domain experts curated 60 candidate guests from DrugBank with MDForge, selecting molecules that are (i) drug-like and biomedically relevant and (ii) potentially valuable targets for CB[7]-based downstream applications. All three evaluation sets have been de-duplicated by InChIKey. Because wet-lab validation of all 60 drug molecules is impractical, affinities were assigned by physically-grounded expert inference, anchored to structurally-related guests with measured CB[7] affinities and adjusted using established recognition patterns, including cavity fit, portal ion–dipole contact and steric penalties. For example, Adapromine carries an adamantane cage plus a protonated amine, which is a high-affinity CB[7] guest motif. However, its α-ethyl substituent imposes a geometric trade-off: the ammonium cannot engage the carbonyl portal while the adamantane remains optimally seated in the cavity. Relative to its parent adamantan-1-amine, one of the strongest CB[7] binders with logKa 14.2, this steric penalty lowers the expected affinity, and the experts accordingly assigned Adapromine a logKa of 11. Each expert label is derived from such anchored and mechanism-based reasoning, ensuring the reliability of Expert dataset.

As for data distribution, the three evaluation sets cover complementary regions of CB[7] chemical and affinity space. SupraBench is the broadest in both affinity range (logKa 2.3–15.1) and scaffold diversity, including 65 aliphatic, 49 aromatic, 8 adamantane and 3 bicyclooctane molecules. Expert shifts toward aromatic medicinal chemotypes while still including the canonical adamantane binders, with logKa ranging from 3.5 to 12.0. SAMPL4 concentrates on aliphatic and aminoadamantane-type guests, with a narrower affinity range (4.9–10.5).

W3 & Q2: Generalization under limited data.

We agree that random five-fold cross-validation alone does not test extrapolation to novel scaffolds, and this is exactly why our evaluation does not rely on SupraBench and SAMPL4 alone. Our Expert set is already a scaffold-disjoint generalization benchmark, where 88% of its 60 guests carry generic Murcko scaffolds absent from the other two evaluation sets. On this scaffold-novel drug-like set, CB7Mol attains competitive ranking quality and the best prediction accuracy. Our reported generalization therefore already reflects performance on structurally novel chemotypes, directly addressing your concern. We are happy to include additional results under scaffold split setting if you think it’s necessary.

Q3: Necessity of representation alignment.

For motivation, aligning the 3D Uni-Mol representation to the 2D GROVER representation distills GROVER's chemical knowledge into 3D molecular encoder. CB[7] recognition is governed by specific chemical motifs, such as an adamantane cage paired with a cationic group. GROVER is pre-trained with objectives that explicitly encode chemical knowledge, including contextual-property prediction and functional-group prediction [1]. Given limited labeled CB[7] affinity data, chemical prior injection via GROVER alignment can refer to the chemical patterns learned from large-scale datasets and facilitate the explicit encoding of the chemical motifs that drive CB[7] recognition. Our experiments in Figure 3(a) empirically demonstrate the contribution of chemical representation alignment to the learning of CB[7]-preferred chemical motifs.

We also compare our alignment strategy with simple concatenation of Uni-Mol and GROVER representations:

MethodSupraBench ρSupraBench τSupraBench RMSEExpert ρExpert τExpert RMSESAMPL4 ρSAMPL4 τSAMPL4 RMSE
CB7Mol0.834±.120.691±.121.253±.350.558±.060.427±.051.729±.110.741±.120.585±.121.864±.12
w/ concatenation0.726 ± .080.569 ± .081.579 ± .320.547 ± .030.418 ± .022.320 ± .190.648 ± .120.508 ± .132.432 ± .25
AuthorsOpenReview ↗

Reply to Reviewer SPL2 (Part 2)

Across all three evaluation sets, alignment exceeds simple concatenation on all metrics. Under alignment, GROVER's chemical knowledge learned from large-scale datasets is distilled into the 3D encoder during pre-training, which is integrated into molecular representations and can be transferred to novel guests independently of affinity labels. In contrast, concatenation-based method must learn how to exploit GROVER’s chemical knowledge from very limited affinity labels. This demonstrates the necessity of alignment-based methodology design.

Q4: Model complexity and data efficiency.

Thank you for raising this question. We take two frozen Uni-Mol encoders with public pre-trained weights to encode hosts and guests independently. Then, we concatenate their fixed embeddings, and fit a classic shallow regressor. We tested several regressor options under the identical cross-host training setting with CB7Mol, and find that SVR in sklearn achieved the best performance. The results are reported in the table below:

MethodSupraBench ρSupraBench τSupraBench RMSEExpert ρExpert τExpert RMSESAMPL4 ρSAMPL4 τSAMPL4 RMSE
CB7Mol0.834±.120.691±.121.253±.350.558±.060.427±.051.729±.110.741±.120.585±.121.864±.12
Frozen Uni-Mol + SVR0.550 ± .100.400 ± .081.997 ± .420.298 ± .020.216 ± .011.970 ± .020.489 ± .020.331 ± .021.971 ± .05

Across all three sets, CB7Mol substantially outperforms the best shallow model on all metrics. We have also compared CB7Mol with several other simpler baselines in Table 1, including QSAR-based methods and GNN-based methods. Taken together, the multi-stage framework design is necessary under data sparsity, because each stage is designed to recruit a distinct, complementary form of transferable supervision from large-scale data, in order to achieve physiochemically grounded affinity prediction when direct labels are limited.

Q5: Effectiveness of cross-host transfer.

How mechanistic compatibility was established: The hosts in our training corpus are all water-soluble macrocycles that recognize guests through the same core physical mechanism: inclusion of a guest moiety into a well-defined hydrophobic cavity, driven by the hydrophobic effect and the release of high-energy cavity water, and modulated by portal/rim electrostatics. The shared recognition mechanisms are the basis on which we selected the cross-hosts. You may find more details in our Appendix B.2.

Individual contribution of different hosts: We have empirically validated the contribution of each host in Figure 3(c). The transferable signal is strongest from hosts that share similar size and the same CB[7] recognition determinants. Removing β-CD or p-SC4 causes the largest performance degradation, where β-CD presents a hydrophobic macrocyclic cavity of dimensions comparable to CB[7], and p-SC4 can reproduce cationic recognition similar to CB[7]. In contrast, removing α-/γ-CD and CB[5] has less impact on guest ranking due to size mismatch. Notably, removing any host will cause performance degradation, and no negative transfer is observed. Hosts with compatible binding mechanisms and different geometrical features are not harmful.

Q6: Validation of physicochemical knowledge.

We agree that representation-level comparison against experimentally observed host–guest interactions and thermodynamic decompositions could be informative. However, such a validation requires experimental observables, including per-atom structural data (X-ray/NOE) and calorimetric thermodynamics (ITC ΔH/ΔS). Our wet-lab is based on competitive NMR, which can not directly reflect structural or thermodynamic decompositions. Acquiring these observables is itself a substantial experimental undertaking beyond what we are able to carry out in this work, so we are unfortunately not in a position to perform this validation, though we agree it is a promising direction for future study.

AuthorsOpenReview ↗

Reply to Reviewer SPL2 (Part 3)

Q7: Interpretation of ablation results.

Thank you for your careful observation. CB7Mol is mainly designed and evaluated for large-scale guest screening, where ranking quality is more important to prioritize promising candidates. Each component is included because of its contribution in guest ranking, which is not necessarily equivalent to prediction accuracy. On the other hand, a lower RMSE from an ablated model comes from the composition of Expert and SAMPL4. The Expert affinities are inferred by analogy to measured anchors, not direct experimental measurements, so their RMSE is less reliable than ranking-based metrics. SAMPL4 covers far fewer guest chemotypes than the other two evaluation sets as we have clarified in Q1, so any systematic bias CB7Mol carries for a given chemotype is shared coherently across most of the set. Because components such as RMSF flexibility and chemical-knowledge alignment model each guest more accurately and hence more consistently, they make this systematic bias more coherent on a chemotype-homogeneous set, which can amplify RMSE even as the ranking improves. Taken together, RMSE is a valuable reference instead of a decisive factor.

Q8: Experimental validation and screening efficiency.

Thank you for your constructive suggestion. We have reported all the 5 experimentally tested candidates in our paper. Two challenges faced by wet-lab experiments are: (1) Not all drugs can be purchased or synthesized in time; (2) Verifying the results of wet-lab experiments often takes more than two weeks. Therefore, it is not feasible to do large-scale wet-lab experiments and come up with a statistically meaningful experimental hit rate.

Thank you again for your detailed comments. We hope that our clarifications and additional experiments will lead to a reconsideration of your score.

Reference:

[1] Rong, Yu, Yatao Bian, Tingyang Xu, Weiyang Xie, Ying Wei, Wenbing Huang, and Junzhou Huang. "Self-supervised graph transformer on large-scale molecular data." Advances in neural information processing systems 33 (2020): 12559-12571.

AuthorsOpenReview ↗

Kind Reminder

Dear Reviewer,

As the discussion period is drawing to a close, we would like to check in on whether our response has resolved your concerns. We are happy to answer any further questions. If our response has addressed your concerns, we would be grateful if you would consider raising your rating. Thank you again for your time and thoughtful feedback.

Best Regards,

Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions for the Area Chair and Reviewers

Dear Area Chair and Reviewers,

Thank you for the detailed and constructive feedback, which helped us improve and clarify important details of our paper. We uploaded a revised PDF in which all changes are highlighted in blue. While all reviewer comments have been carefully considered and addressed in our responses, we have prioritized revisions that we believe are most important for improving the clarity and rigor of the manuscript. Below, we summarize the new evidence and clarifications:

MD Baseline Comparison

  • We add a simple MD baseline MM/GBSA as an additional baseline in Table 1.
  • Related work section is updated with additional MD method description.
  • Appendix A.2 further clarifies the reason why we don't adopt more advanced MD baselines for comparison, which is mainly about the numerical instability, systematic bias and extremely high computational cost.

Experimental Setup Details

  • Appendix A.1 adds more details about Expert dataset curation and the affinity distribution of three evaluation sets.
  • Appendix A.2 adds more details of GNN baseline implementation.
  • Appendix A.4 adds additional clarification about the baseline implementations.

Additional Experimental Results

  • We add new geometrical information preservation analysis in Section 4.4 and Figure 3c, in order to demonstrate that self-distillation based post-training in CB7Mol is able to preserve geometrical meaningful information. The corresponding experimental settings are summarized in Appendix A.9.
  • Appendix B.4 adds additional interpretation of ablation results, in order to address potential concerns about better RMSE results in ablated models.

Readability Improvement

  • We improved the writing of the third and fourth paragraph of introduction as advised by reviewers.

We look forward to your feedback and hope that our responses help clarify our contributions and that you will consider revising your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.