acceptodds
Under review as a conference paper at ICLR 2027

PolyBC: A Graph-Augmented Chemical Language Model for Complex Copolymer Learning

Abstract

Copolymers are macromolecules assembled from repeating structural units, arranged in random, alternating, or block architectures. Predicting and designing copolymer properties from molecular structure is of broad interest across energy storage, construction, medicine, and aerospace, yet remains fundamentally challenging because (1) there is no effective parameterization for the structural units or for the complex architectures connecting them, and (2) experimental measurements for any particular application are typically scarce and unbalanced. To resolve these challenges, we propose PolyBC, a physics-guided, graph-augmented chemical language model. In particular, we introduce a graph neural network over structural units at two levels of granularity — repeat units and sub-monomer fragments — whose nodes are encoded by Poly4mer, a chemical language model that supplies transferable latent embeddings of homopolymer segments. At either resolution, the GNN reasons over the polymer topology and closes the finite chain with architecture-matched boundary conditions: periodic for alternating copolymers, Neumann for block copolymers, and mean-field for random copolymers. This construction imposes a theoretically guaranteed, consistent bulk physics: training on one copolymer type and boundary condition generalizes to the others, and predictions are invariant to how many repeats are drawn. On both synthetic and experimental benchmarks, our model is repetition-invariant, data-efficient, and transferable across copolymer types at both granularities, outperforming state-of-the-art GNN and LLM baselines on prediction and design tasks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
polymer property prediction, chemical language model, graph neural networks, physics-guided machine learning, molecular representation learning

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_T9wfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

PolyBC represents copolymers as chains of monomer-level or fragment-level units rather than atom graphs. Architecture is encoded through boundary rules at the chain ends, while node features come from a frozen Poly4mer representation. The paper evaluates on Coley and PoLyInfo with cross-architecture transfer, repeat-number tests, small-data settings, and inverse design. My main concern is attribution. The empirical results, especially on PoLyInfo, are strong, but the experiments do not show that the proposed boundary-condition formulation is what produces them.

Critical strengths
  • The empirical evaluation is extensive.
  • The PoLyInfo transfer results are strong, particularly given the smaller model size and R=1-only training. The paper reports several results that do not favor its own method or interpretation, which makes the empirical section more credible.
  • Reproducibility details are relatively complete, with code, splits, checkpoints, and per-row predictions promised.
Critical weaknesses
  1. The source of the gains is unclear. The paper attributes zero-shot transfer to treating architecture as a boundary condition, but the ablations suggest that this component has only a small effect.
  • In Table 17, changing only the ghost rule changes R² by roughly 0.002. And the boundary-condition intervention is only tested in-domain, not on the transfer cells where the main claim matters.
  • The theoretical results also do not seem to depend on the nonlocal-operator formulation itself. The repetition result follows from repeated local structure and an invariant readout, while the shared-operator result largely follows from weight sharing and the absence of an architecture embedding.
  • A more plausible source of the large gains is the combination of unit-level graphs and strong pretrained unit features.

Three controls would help separate these effects: use directly computed ECFP as node features while keeping the rest fixed; remove the 1851 replicated homopolymer rows; run the boundary-condition intervention on transfer cells.

  1. Some theoretical claims are stronger than the evidence. Prop. 3 covers only environments that coincide with those seen in training. Corollary 1 notes that on monomer graphs at L=1 the environment sets of alternating and block chains are disjoint, so the hardest transfers depend entirely on generalization to unseen environments, which the proposition does not cover. The abstract’s “theoretically guaranteed” generalization across types should be qualified accordingly.

  2. The inverse-design evidence is limited. The comparison to MOSES VAE/AAE is difficult to interpret because those models are retrained on only 683 molecules, while the paper's decoder benefits from large-scale pretraining.

  3. Page 2 contains leftover editing text ("Here's the polished version, with your LaTeX kept:").

Questions for authors
  1. What happens if the distilled node features are replaced by directly computed ECFP?
  2. How much of the PoLyInfo transfer gain remains after removing the replicated homopolymer rows?
  3. Can the boundary-condition intervention be run on cross-architecture transfer cells?
  4. Why is wD-MPNN not included as a baseline?
  5. Does the nonlocal-operator interpretation impose any constraint beyond the choice of ghost rules?
1 comment

We thank the reviewer for the constructive feedback of our manuscript. Our response:

Source of the gains; boundary intervention on transfer cells: We ablate the boundary conditions on Coley dataset, where the accuracy of each model is relatively better than PolyInfo, so we believe that is the reason why the improvement of boundary condition is less. We compare our PolyBC (full) with the PolyBC (without boundary condition) on the small PolyInfo dataset both in-domain as well as transfer with 95%95\% confidence intervals. We find that our boundary condition improves the accuracy, especially on the transfer cells.

Model TgT_gIn-domain (R=1R=1)Transfer (R=1R=1)In-domain (R=20R=20)Transfer (R=20R=20)
PolyBC-F (full)0.751±0.0170.751\pm0.0170.699±0.0350.699\pm0.0350.749±0.0170.749\pm0.0170.699±0.0360.699\pm0.036
PolyBC-F (w.o. BC)0.723±0.0240.723\pm0.0240.646±0.0550.646\pm0.0550.715±0.0240.715\pm0.0240.638±0.0550.638\pm0.055
Improvement Δ\Delta (w.o. BC- full)−0.029-0.029 [−0.054-0.054, +0.011+0.011]−0.052-0.052 [−0.087-0.087, −0.017-0.017]−0.034-0.034 [−0.059-0.059, +0.002+0.002]−0.060-0.060 [−0.099-0.099, −0.023-0.023]

Directly computed ECFP instead of distilled features: We ran exactly this control: PolyBC-F with its distilled node features replaced by directly computed ECFP, everything else fixed, and a second arm concatenating both feature sets (PoLyInfo, 5 seeds):

Node featuresIn-domain R2R^2Transfer R2R^2
Distilled Poly4mer features0.827±0.0110.827\pm0.0110.809±0.0090.809\pm0.009
Direct ECFP0.795±0.0070.795\pm0.0070.808±0.0080.808\pm0.008
Distilled + ECFP0.807±0.0090.807\pm0.0090.808±0.0070.808\pm0.007

We find out that our distilled Poly4mer features performs slightly better than Direct ECFP, and by doing distillation, the features are therefore differentiable, enabling the possibility to implement inverse design using PolyBC.

Removing the 1,851 homopolymer rows: These rows are 617 measured homopolymer TgT_gs, entered once per sequence type (617×3617\times3), and they share no record with the test set. Removing them lowers PolyBC-F's transfer R2R^2 from 0.809±0.0090.809\pm0.009 to 0.574±0.0090.574\pm0.009, so we state their role prominently. Without them, PolyBC-F still leads the chemical-LM chain models on transfer (MoLFormer 0.487±0.0090.487\pm0.009, ΔR2=−0.087\Delta R^2=-0.087 [−0.184-0.184, −0.008-0.008]; TransPolymer 0.398±0.0280.398\pm0.028, ΔR2=−0.176\Delta R^2=-0.176 [−0.502-0.502, −0.047-0.047]).

Prop. 3 and Corollary 1: We removed "theoretically guaranteed generalization" from the abstract.

Fair MOSES baselines: We retrained the MOSES VAE and AAE on the same data as PolyBC’s decoder (1.73M polymer SMILES for pretraining, then the same PoLyInfo monomers). With equal data, PolyBC-M design with local random search outperforms MOSES baselines in terms of uniqueness and novelty, while MOSES reaches higher validity. The two methods do different jobs: MOSES samples monomers without a target, whereas PolyBC edits a given copolymer at fixed composition and architecture under the guidance of its accurate property predictions. Therefore, our PolyBC also benefits from designing functional materials with desired properties for downstream applications.

MethodValidity (%)Uniqueness (%)Novelty (%)
MOSES VAE, equal data99.9±0.199.9\pm0.14.3±0.44.3\pm0.472.2±2.772.2\pm2.7
MOSES AAE, equal data78.2±1.578.2\pm1.523.0±0.923.0\pm0.984.9±0.684.9\pm0.6
PolyBC-M design (gradient descent)11.6±3.511.6\pm3.526.1±6.726.1\pm6.772.8±7.972.8\pm7.9
PolyBC-M design (local random)51.4±2.451.4\pm2.435.7±4.835.7\pm4.892.0±1.592.0\pm1.5

Editing residue on page 2: Thank you; it is removed.

wD-MPNN: We added the official wD-MPNN on both datasets (details in our response to Reviewer BSeN). On Coley, PolyBC-M is more accurate under Coley's random CV (0.9986±0.00000.9986\pm0.0000 vs. 0.9965±0.00020.9965\pm0.0002) and on the pooled pair split (0.9962±0.00010.9962\pm0.0001 vs. 0.9936±0.00030.9936\pm0.0003). Under the manuscript's settings, PolyBC is more accurate on Coley IP (0.947±0.0010.947\pm0.001 vs. 0.810±0.0220.810\pm0.022), PoLyInfo-Full (0.815±0.0060.815\pm0.006 vs. 0.757±0.0350.757\pm0.035) and PoLyInfo-Small (0.727±0.0180.727\pm0.018 vs. 0.620±0.0480.620\pm0.048).

On nonlocal-operator: The nonlocal perspective comes from the fact that message passing graph can be seen as a discretized nonlocal operator with finite horizon, with the sum over neighbors acting as a quadrature of the integral (see., e.g., [1], [2]). Hence, there should be no additional constraint introduced from this perspective.

[1] N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart and A. Anandkumar, Neural operator: Learning maps between function spaces with applications to PDEs, Journal of Machine Learning Research 24 (2023), no. 89, 1–97.

[2] G. Gilboa and S. Osher, Nonlocal operators with applications to image processing, Multiscale Modeling & Simulation 7 (2008) 1005–1028.

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

The paper introduces PolyBC, which combines pretrained chemical embeddings with message passing for copolymer property prediction and inverse design. It reports strong results on computed EA/IP and experimental glass-transition (Tg) data.

However, the paper's central claim, that its "theoretical guarantees" ensure consistent bulk polymer physics, is wrong. The paper presents an invariance that holds by construction in its own neural representation as a “theoretical guarantee”, but real polymers behave the opposite way where their properties change with the number of repeat units, which PolyBC cannot capture. The benchmarks used (oligomer-based labels and Tg without chain length) cannot test them.

Critical strengths

The predictive results look strong. The empirical method may be useful.

Critical weaknesses

Severe problem

Their claimed central contribution (ii), lines 80–85, is a factual error in polymer science:

"for alternating and block copolymers, the predicted property is invariant to the number of repeat units used to represent the same copolymer. For random copolymers, this invariance holds asymptotically as the number of repeat units grows."

  • It means the model cannot distinguish chains of different length built from the same repeat unit. For example, pentane, decane and polyethylene share the same -CH₂- repeat unit. A model that is invariant to the number of repeat units must assign them the same property, although pentane is a liquid at room temperature and polyethylene melts at over 100 °C. Likewise, low-molecular-weight polystyrene has a Tg tens of degrees lower than high-molecular-weight polystyrene.

The whole of Section 3.5, "THEORETICAL GUARANTEES," is built on the same error.

  1. Lemma 1 and Theorem 1 are trivially true by construction. Although they are correct as mathematics, they are useless as a description of polymers. They may describe the receptive field of a network and the behavior of a periodic graph, but cannot describe a polymer ensemble which has different chain lengths.
  2. Polymer properties may depend heavily on chain length or degree of polymerization. Simple examples include Tg (which depends on Mn through Flory-Fox) and block morphology (which depends on χN). Chain-length invariance is an assumption that may hold in the long-chain limit for some cases; it is not a theorem (R=1 is clearly an oligomer; R=20 on the Coley data is also far from the long-chain limit).
  3. "Theoretically guaranteed, consistent bulk physics" is severely misleading, as the "guarantees" are quite against polymer physics.

The above leads to several follow-up mistakes:

  1. The Coley labels are used incorrectly. Aldeghi & Coley computed EA/IP on octamer (8-mer) sequences rather than a series of physical chain lengths. That means there are no labels for longer chains (larger R). Thus, testing different R against the same label is wrong. EA/IP of conjugated chains apparently change with length, so the claimed invariance up to R=20 contradicts how the labels were generated.
  2. The Neumann rule simply copies endpoint hidden states into ghost nodes. It neither takes a physical infinite-chain limit nor shows that the resulting representation reproduces the finite-oligomer target.

Other major issues

  1. The first dataset comes from Aldeghi & Coley, but their own model, wD-MPNN, is not included. Aldeghi & Coley already report R² = 1.00 on this dataset. The authors leave out the benchmark's own model but claim PolyBC is the best.
  2. Random/block/alternating describe sequence, not architecture. Architectures such as graft, branched and star are not covered in this work.
  3. The "soft-alternating" chain (AB)ₙXₘ is essentially an alternating–homopolymer diblock, not an alternating copolymer. It makes up over 70% of the alternating test set.
Questions for authors
  1. The Coley EA/IP labels were computed on octamers. Why should a prediction for an 8-mer property be invariant from R=1 to R=20, and what label is the R=20 prediction supposed to match?
  2. How were block copolymer records that report more than one Tg handled? Note that microphase separated blocks usually show two Tgs.
4 comments

We thank the reviewer for the constructive feedback of our manuscript. Our response:

Severe problem: We respectfully clarify that the repeat unit we used here may contain one or more monomers. For example, an ABAB cell has two physical building blocks, and four repetitions form an eight-block chain. On the other hand, an ABBBABBB cell has four blocks, and two repetitions form an eight-block chain. We denote the number of repetitions by RR. Examples of pentane, decane and polyethylene contain repeat units -CH$_2$-, which are not monomers. We have explained the terminology to avoid any misunderstandings.

On polymer physics and theoretical results: We agree with the reviewer that properties of copolymers may depend on the chain length or degree of polymerization. We want to clarify that the motivation here is how to learn the exact physics (repetition invariance) behind the data. In other words, the training dataset should contain the properties of ensemble, or fixed-length chains as long as the repetition invariance has already been incorporated in the dataset. Ideally, the training dataset may contain labels of chains long enough (effective conjugation length or large MnM_n) for EA/IP/Tg to guaranty repetition invariance in physics. In the previous study, GRIN allows identical labels to copolymers with different number of repeat units (R=1,3R=1, 3 for training and R=20R=20 for testing) on EA/IP. In fact, at least 20-mer can be considered as a long chain for electronic properties [1,2]. Additionally, we also test on R=100R=100 for EA/IP, and R=500R=500 for TgT_g. The results show that our PolyBC still outperforms GRIN consistently on longer chains.

ModelTargetIn-domain (R=1R=1)Transfer (R=1R=1)In-domain (R=100R=100)Transfer (R=100R=100)
PolyBC-FEA0.9902±0.00030.9902\pm0.00030.9380±0.00170.9380\pm0.00170.9887±0.00040.9887\pm0.00040.9383±0.00170.9383\pm0.0017
PolyBC-FIP0.9904±0.00010.9904\pm0.00010.9246±0.00120.9246\pm0.00120.9888±0.00020.9888\pm0.00020.9221±0.00130.9221\pm0.0013
PolyBC-MEA0.9954±0.00010.9954\pm0.00010.9287±0.00250.9287\pm0.00250.9941±0.00030.9941\pm0.00030.9252±0.00380.9252\pm0.0038
PolyBC-MIP0.9955±0.00020.9955\pm0.00020.9058±0.00130.9058\pm0.00130.9941±0.00020.9941\pm0.00020.9049±0.00230.9049\pm0.0023
GRINEA0.9537±0.00060.9537\pm0.00060.8934±0.00340.8934\pm0.00340.9368±0.00100.9368\pm0.00100.8775±0.00390.8775\pm0.0039
GRINIP0.9599±0.00080.9599\pm0.00080.8752±0.00300.8752\pm0.00300.9559±0.00090.9559\pm0.00090.8684±0.00610.8684\pm0.0061
ModelTargetIn-domain (R=1R=1)Transfer (R=1R=1)In-domain (R=500R=500)Transfer (R=500R=500)
PolyBC-FTgT_g0.8231±0.00210.8231\pm0.00210.8105±0.01340.8105\pm0.01340.8221±0.00230.8221\pm0.00230.8103±0.01340.8103\pm0.0134
PolyBC-MTgT_g0.7593±0.02840.7593\pm0.02840.7771±0.01400.7771\pm0.01400.7583±0.02870.7583\pm0.02870.7768±0.01400.7768\pm0.0140
GRINTgT_g0.7919±0.02130.7919\pm0.02130.7135±0.02200.7135\pm0.02200.7916±0.02140.7916\pm0.02140.7105±0.02140.7105\pm0.0214

Lemma 1 and Theorem 1 describe the structure of alternating copolymer −(AB)n−-(AB)_n- using periodic graphs, and show the strict repetition invariance can be achieved by this kind of representation incorporated in PolyBC.

[1] Zhang et al. (2013), Synthesis, Electronic Structure, Molecular Packing/Morphology Evolution, and Carrier Mobilities of Pure Oligo-/Poly(alkylthiophenes), Journal American Chemical Society

[2] Zade and Bendikov (2006), From Oligomers to Polymer: Convergence in the HOMO-LUMO Gaps of Conjugated Oligomers, Organic Letters

Coley dataset: Coley's dataset was generated by N=8N=8, but we think it is an approximation of repetition invariance based on four facts:

  1. Prior work uses the same setting. GRIN's dataset is a subset of Coley dataset, which was also generated from N=8N=8.

  2. The labels change little with chain length. Due to the time limit, we simulate EA/IP of 36 kinds of copolymers in Coley dataset with N=4,8,16N=4, 8, 16. We observed that the labels changes little, especially for IP dataset. Even generating Coley dataset will consume 4 to 5k hours on a single RTX5090. Thus, we use Coley dataset as an approximation for the exact repetition invariance like what GRIN does.

    Label used →\to target lengthEA, relative L2L_2 (%)IP, relative L2L_2 (%)
    N=4→8N=4\to810.710.7 [9.49.4, 12.212.2]4.74.7 [4.24.2, 5.35.3]
    N=8→16N=8\to167.37.3 [6.46.4, 8.18.1]3.03.0 [2.52.5, 3.53.5]
    N=4→16N=4\to1617.117.1 [15.315.3, 19.219.2]7.57.5 [6.56.5, 8.48.4]
  3. The literature predicts this behavior. For π\pi-conjugated chains, EA and IP converge smoothly with oligomer length towards a polymer limit, with the largest changes at the shortest lengths [3,4]. Extrapolations from the first oligomers of a series (N≤8N\le 8) already estimate the polymer-limit transition energy to within about 1.5% [4].

    [3] Gierschner, J., Cornil, J., & Egelhaaf, H.-J. (2007). Optical bandgaps of π\pi-conjugated organic materials at the polymer limit: experiment and theory. Advanced Materials, 19(2), 173–191.

    [4] Torras, J., Casanovas, J., & Alemán, C. (2012). Reviewing extrapolation procedures of the electronic properties on the π\pi-conjugated polymer limit. The Journal of Physical Chemistry A, 116(28), 7571–7583.

  4. Test on R=8R=8: We also found that PolyBC performs consistently on R=8R=8, where Coley labels are precise:

    ModelTargetIn-domain (R=1R=1)Transfer (R=1R=1)In-domain (R=8R=8)Transfer (R=8R=8)ΔR2\Delta R^2 all 9 cells
    PolyBC-MEA0.9954±0.00010.9954\pm0.00010.9287±0.00250.9287\pm0.00250.9952±0.00010.9952\pm0.00010.9283±0.00270.9283\pm0.0027−0.0003-0.0003 [−0.0004-0.0004, −0.0002-0.0002]
    PolyBC-MIP0.9955±0.00020.9955\pm0.00020.9058±0.00130.9058\pm0.00130.9953±0.00020.9953\pm0.00020.9058±0.00180.9058\pm0.0018−0.0001-0.0001 [−0.0003-0.0003, +0.0004+0.0004]
    PolyBC-FEA0.9902±0.00030.9902\pm0.00030.9380±0.00170.9380\pm0.00170.9890±0.00040.9890\pm0.00040.9387±0.00170.9387\pm0.0017−0.0000-0.0000 [−0.0001-0.0001, +0.0001+0.0001]
    PolyBC-FIP0.9904±0.00010.9904\pm0.00010.9246±0.00120.9246\pm0.00120.9890±0.00020.9890\pm0.00020.9229±0.00130.9229\pm0.0013−0.0016-0.0016 [−0.0018-0.0018, −0.0014-0.0014]
    GRINEA0.9730±0.00060.9730\pm0.00060.9118±0.00380.9118\pm0.00380.9726±0.00040.9726\pm0.00040.9139±0.00400.9139\pm0.0040+0.0012+0.0012 [−0.0000-0.0000, +0.0024+0.0024]
    GRINIP0.9635±0.00080.9635\pm0.00080.8790±0.00300.8790\pm0.00300.9663±0.00120.9663\pm0.00120.8802±0.00460.8802\pm0.0046+0.0018+0.0018 [−0.0000-0.0000, +0.0035+0.0035]

The Neumann rule: The Neumann boundary condition is implemented as an approximation of the constant extension of feature h1(k)h_1^{(k)} and hN(k)h_{N}^{(k)} to the infinite-chain. It was based on a classical extension theorem in real analysis. Taking a function u(x)u(x) on domain Ω=[0,1]\Omega=[0,1] for illustration. The boundary condition in our model is a constant extension of the solution beyond the boundary, u~(x)=u(0)\tilde u(x)=u(0) for x<0x<0 and u~(x)=u(1)\tilde u(x)=u(1) for x>1x>1. The homogeneous Neumann conditions u′(0)=0u'(0)=0 and u′(1)=0u'(1)=0 used in the implementation are their leading-order approximation, in the following sense. By the standard gluing argument for one-sided derivatives (see, e.g., Ch. 11, Thm. 7 of [5]), the constant extension u~\tilde u is CkC^k across the boundary if and only if

u(i)(0)=0,i=1,…,k.u^{(i)}(0)=0,\qquad i=1,\dots,k .

A smooth solution with constant extension therefore satisfies this hierarchy for every kk, and the Neumann condition is its first member (k=1k=1). Moreover, if the first kk derivatives vanish, Taylor's theorem shows that the constant extension differs from any smooth continuation EuEu of uu only at order k+1k+1,

∣u~(x)−Eu(x)∣≤Mk+1(k+1)! ∣x∣k+1,x<0,|\tilde u(x)-Eu(x)|\le \frac{M_{k+1}}{(k+1)!}\,|x|^{k+1},\qquad x<0,

where Mk+1M_{k+1} bounds ∣(Eu)(k+1)∣|(Eu)^{(k+1)}| near the boundary. Hence the Neumann condition alone yields an O(∣x∣2)O(|x|^{2}) approximation of the constant extension near the boundary, and each additional vanishing higher-order derivative improves the accuracy by one order.

To further verify if this first-order approximation is sufficient, we also implemented second-order approximation (u′′(0)=0u''(0)=0 and u′′(1)=0u''(1)=0), and found that the performance is still comparable, so we adopt the first-order approximation.

ModelIn-domain (R=1R=1)Transfer (R=1R=1)In-domain (R=20R=20)Transfer (R=20R=20)
Neumann BC0.8270±0.01070.8270\pm0.01070.8094±0.00860.8094\pm0.00860.8267±0.01040.8267\pm0.01040.8092±0.00880.8092\pm0.0088
2nd order BC0.8253±0.01470.8253\pm0.01470.8139±0.01100.8139\pm0.01100.8255±0.01450.8255\pm0.01450.8140±0.01120.8140\pm0.0112

[5] M. Spivak, Calculus, 4th ed., Publish or Perish, Houston, TX, 2008.

wD-MPNN: We added wD-MPNN as an additional baseline both on Coley’s released 10-fold random CV (setting A) and our pair split with all sequence types pooled (setting B). ΔR2\Delta R^2 = wD-MPNN −- PolyBC (negative favors PolyBC) with 95% intervals. We respectfully note that the "R2=1.00R^2 = 1.00" reported by Aldeghi & Coley is a rounded value: from their released predictions it is 0.9966 (EA) and 0.9968 (IP), and PolyBC-M outperforms it.

SettingTargetwD-MPNNPolyBC-MPolyBC-FΔR2\Delta R^2 vs. PolyBC-M
AEA0.9965±0.00020.9965\pm0.00020.9986±0.0000\mathbf{0.9986\pm0.0000}0.9949±0.00000.9949\pm0.0000−0.0020-0.0020 [−0.0022-0.0022, −0.0019-0.0019]
AIP0.9968±0.00010.9968\pm0.00010.9987±0.0000\mathbf{0.9987\pm0.0000}0.9953±0.00010.9953\pm0.0001−0.0018-0.0018 [−0.0020-0.0020, −0.0017-0.0017]
BEA0.9936±0.00030.9936\pm0.00030.9962±0.0001\mathbf{0.9962\pm0.0001}0.9906±0.00010.9906\pm0.0001−0.0026-0.0026 [−0.0031-0.0031, −0.0021-0.0021]
BIP0.9946±0.00040.9946\pm0.00040.9963±0.0001\mathbf{0.9963\pm0.0001}0.9906±0.00020.9906\pm0.0002−0.0017-0.0017 [−0.0021-0.0021, −0.0013-0.0013]

Branched, star and graft topologies: We clarify that our terminology architecture follows Aldeghi & Coley, where they called Random/block/alternating chain architecture. We also applied PolyBC to the branched and star topologies.

We used the public ToPoRg-18k benchmark [6]: 18,150 coarse-grained bead–spring polymers in six topology classes (linear, cyclic, star, branched, comb, dendrimer), with two bead types arranged in 25 sequence patterns at three sizes (40–60, 90–110 and 190–210 beads). The target is the MD mean-squared radius of gyration ⟨Rg2⟩\langle R_g^2\rangle (in σ2\sigma^2), which at fixed size depends strongly on topology. All splits were fixed before training, with isomorphic graphs always on one side: A (in distribution; 20% of shapes held out, stratified by class and size, three replicates A_s0–A_s2), P (five bead patterns never seen in training, on unseen shapes), and S (train on 40–110 beads, test on every 190–210-bead polymer).

Each bead is a unit node on the true bond graph, and a branch point is simply a unit with more than two neighbours. PolyBC keeps its shared local update and boundary condition. Here, we change the readout to a length-aware Gaussian-network readout, Rg2=tr⁡(Lw+)/NR_g^2=\operatorname{tr}(L_w^{+})/N, where LwL_w is the Laplacian of the bond graph weighted by learned per-unit bond lengths. We observe that PolyBC can handle complex topologies including branched, star, comb and dendrimer. Results R2R^2 of Rg2R_g^2 compared to GRIN:

Test setClassesPolyBC-MGRIN
In distribution, A_s0all six0.997±0.000\mathbf{0.997\pm0.000}0.996±0.0010.996\pm0.001
In distribution, A_s1all six0.997±0.000\mathbf{0.997\pm0.000}0.995±0.0030.995\pm0.003
In distribution, A_s2all six0.997±0.000\mathbf{0.997\pm0.000}0.996±0.0000.996\pm0.000
Longer chains, Sall six0.987±0.002\mathbf{0.987\pm0.002}0.985±0.0010.985\pm0.001
In distribution, A_s0branched0.997±0.000\mathbf{0.997\pm0.000}0.995±0.0020.995\pm0.002
In distribution, A_s1branched0.997±0.001\mathbf{0.997\pm0.001}0.994±0.0040.994\pm0.004
In distribution, A_s2branched0.996±0.001\mathbf{0.996\pm0.001}0.995±0.0010.995\pm0.001
In distribution, A_s1star0.997±0.000\mathbf{0.997\pm0.000}0.993±0.0030.993\pm0.003
Unseen patterns, Pstar0.992±0.000\mathbf{0.992\pm0.000}0.989±0.0040.989\pm0.004
Longer chains, Sstar0.984±0.004\mathbf{0.984\pm0.004}0.966±0.0060.966\pm0.006

[6] Jiang & Webb (2025), Physics-Guided Neural Networks for Transferable Prediction of Polymer Properties, Macromolecules.

Soft-alternating chains: The soft-alternating sequence at RR is (AB)mRX∣nA−nB∣R(AB)^{mR}X^{|n_A-n_B|R}: 254 of the 330 alternating test rows are unequal-composition soft-alternating, and 67 are strict 1:1. We relabel these rows as "alternating–homopolymer diblock (soft-alternating)" and report strict 1:1 alternating separately.

Block copolymer records with more than one TgT_g: We used the mean of all TgT_g values listed in a record, and we now state this explicitly. We retrained PolyBC-F without the two-transition records and found the performance is still comparable (0.809±0.0090.809\pm0.009 →\to 0.816±0.0080.816\pm0.008), so the averaged labels do not affect our performance. In the revision, we report the results on lower TgT_g which confirms the occurrence of microphase separation and the formation of an independent soft domain.

R2R^2 and MAE (∘^\circC) are averaged over the three training architectures; mean ±\pm SD over seeds:

PredictionR2R^2MAE (∘^\circC)
PolyBC-F0.775±0.0200.775\pm0.02036.7±1.736.7\pm1.7
GRIN0.548±0.0570.548\pm0.05748.7±3.648.7\pm3.6
Reviewer_WL7qOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PolyBC uses a frozen Poly4mer LM to embed monomers or fragments, then a GNN over the chain. Copolymer architecture is encoded as a boundary condition: periodic for alternating, Neumann for block, mean-field for random. The paper claims repeat invariance, zero-shot transfer across architectures, and gradient-based inverse design. The idea is neat, but most of the headline wins come from invariance built into the design, benchmarks that are already saturated, and computed labels. The real-data and design results are weak.

Critical strengths
  • Separating unit chemistry from chain architecture is clean, and the boundary-condition framing is easy to follow.

  • Every cell of the 3×3 tables is reported with seeds, and the appendix openly admits several weaknesses.

Critical weaknesses
  • Repeat invariance is guaranteed by the architecture, and the baselines can't have it. Baselines are trained at R=1 and tested at R=20. polyBERT and TransPolymer get truncated at 512 tokens, and on PoLyInfo some strings already exceed the window at R=1 (1,313 of 2,907 for MoLFormer). So the Test20 gap mostly shows that the baselines can't read the input. There are no fair invariant baselines, like a monomer-set or DeepSets model, or GRIN given the same unit-level input.

  • The inverse-design results are weak. The Tg designs have no oracle at all. Validity is lower than the baselines'. Only 15–29 novel units come out per 1,000 samples. The authors also concede that the latent–decoder mismatch, not the optimizer, is what limits design, and that re-encoding adds a spurious shift of "tens of degrees." That doesn't support the claim that the model is "differentiable for design." The zero-shot transfer story is incomplete. a→b and b→a remain the weakest cells (0.85–0.88 on Coley EA). There's no analysis of why a shared interior operator should transfer, beyond the fact that it's shared.

  • The ± values are population standard deviations over 5 seeds (GRIN gets 3 seeds and different augmentation), and there are no significance tests.

  • There are a lot of acronyms (Test1/Test20, M/F, r/a/b), and the main tables are hard to read.

  • Missing the 3D/conformational line of work (e.g., PolyConf), which weakens the "physics-guided" claim.

Questions for authors
  • Tg depends heavily on chain stiffness, torsions, and tacticity. Have you tried adding conformation-aware features, e.g., from PolyConf-generated or MD conformers? Is repeat invariance the wrong bias for short-block copolymers, where chain folding matters?

  • Can you validate the Tg designs with an independent oracle (MD or another model), or with even a few known polymers? How big is the shift introduced by re-encoding decoded designs, and how does it affect the claim that the model is useful for design?

  • Why 5 seeds for PolyBC but 3 for GRIN, with different augmentation? Can you report sample standard deviations or confidence intervals and do paired significance tests?

2 comments

We thank the reviewer for the positive assessment of our manuscript. Our response:

Repeat invariance and fair invariant baselines: We agree that the only GRIN among the baselines has repetition invariance. We added DeepSets and unit GNN with identical node features as additional baselines tested on R=1R=1. As shown in the table below, PolyBC still outperforms these baselines.

ModelEA transfer R2R^2EA ΔR2\Delta R^2IP transfer R2R^2IP ΔR2\Delta R^2
PolyBC-M0.930±0.0020.930\pm0.002–0.907±0.0020.907\pm0.002–
DeepSets0.878±0.0000.878\pm0.000−0.052-0.052 [−0.057-0.057, −0.047-0.047]0.794±0.0000.794\pm0.000−0.113-0.113 [−0.121-0.121, −0.105-0.105]
Mean-pool unit GNN0.907±0.0000.907\pm0.000−0.023-0.023 [−0.026-0.026, −0.020-0.020]−0.399±0.001-0.399\pm0.001−1.306-1.306 [−1.397-1.397, −1.222-1.222]

Why a$\to$b and b$\to$a are the hardest cells: We agree with the reviewer that an alternating$\leftrightarrow$block cell is the worst transfer cell, and would like to point out that it is the hardest for every model we evaluated on the Coley dataset. This is because physics-wise the differences of properties (EA/IP) between alternating and block copolymers are lager than others, which makes the transfer harder in nature. For a$\leftrightarrow$b, the relative errors of EA and IP are 7.5% and 3.5%, while the relative errors for the other copolymers are only about half as large (a$\leftrightarrow$r: 3.5% and 1.6%, b$\leftrightarrow$r: 3.2% and 1.6%).

Seeds, sample standard deviation, and significance tests: We report the sample SD (ddof =1=1) and 95% tt-intervals. GRIN was rerun with 5 seeds. The Test1 results with 95% tt-intervals are as follows (ΔR2\Delta R^2= GRIN −- PolyBC, negative shows PolyBC better):

SettingGRIN R2R^2PolyBC-F R2R^2PolyBC-M R2R^2ΔR2\Delta R^2 vs. PolyBC-FΔR2\Delta R^2 vs. PolyBC-M
PoLyInfo-Full0.714±0.0220.714\pm0.0220.811±0.0130.811\pm0.0130.777±0.0140.777\pm0.014−0.097-0.097 [−0.279-0.279, −0.035-0.035]−0.064-0.064 [−0.260-0.260, −0.001-0.001]
PoLyInfo-Small0.519±0.0920.519\pm0.0920.682±0.0250.682\pm0.0250.737±0.0220.737\pm0.022−0.163-0.163 [−0.262-0.262, −0.045-0.045]−0.218-0.218 [−0.306-0.306, −0.108-0.108]
Coley EA0.893±0.0030.893\pm0.0030.938±0.0020.938\pm0.0020.929±0.0030.929\pm0.003−0.045-0.045 [−0.063-0.063, −0.029-0.029]−0.035-0.035

3D and conformational information: We added a related-work paragraph on the 3D line of work including PolyConf and MMPolymer. Since the link of PolyConf model weights is not valid, per the reviewer's suggestion we added MD-derived stiffness features. We simulated a 10-mer chain of every monomer that could be parameterized (895 monomers; OpenFF Sage + NAGL; 2 ns at 500 K) and extracted dihedral entropy, circular variance, trans fraction, torsion transition rate, a CNC_N proxy, RgR_g per unit, and side-chain size and reach. We found that adding MD descriptors on PolyBC does not improve the accuracy. Therefore, we believe that the pretrained latent embeddings of PolyBC already capture most of the stiffness information here. The results on PolyInfo dataset are reported as follows.

ModelIn-domain R2R^2Transfer R2R^2
PolyBC-F0.827±0.0110.827\pm0.0110.809±0.0090.809\pm0.009
PolyBC-F + MD descriptors0.817±0.0190.817\pm0.0190.791±0.0120.791\pm0.012
MD descriptors only0.735±0.0060.735\pm0.0060.553±0.0160.553\pm0.016

For short-block copolymers, repetition invariance is appropriate when their sequence, block length, degree of polymerization, and end groups are preserved. However, this invariance becomes inappropriate when it conceals physical differences that affect target properties. As demonstrated in experiments on oxyethylene/oxybutylene diblock copolymers, altering the actual block length can change how the chain folds [1]. In the previous study, GRIN allows identical labels to copolymers with different number of repeat units on EA/IP, and we follow the same settings here.

[1] Yang et al. (1995), Structures of Oxyethylene/Oxybutylene Diblock Copolymers in Their Solid and Liquid States, Macromolecules.

Independent TgT_g oracle to evaluate designs: We built an all-atom MD to simulate TgT_g and used it to verify our designs. Each structure is a cell of 10 atactic linear chains of at least 600 atoms (25–88 repeat units). The chains are assembled unit by unit with the copolymer's sequence type and composition and capped with hydrogen. Parameters come from OpenFF Sage 2.2.1 with NAGL partial charges. The cell is compressed stepwise at 700 K to 0.8 g cm$^{-3}$ and equilibrated for 4.1 ns (NVT, then NPT at 1,000 and at 1 bar). It is then cooled from 700 to 150 K in 25 K steps of 2 ns NPT each, about 51 ns per run (OpenMM; Langevin dynamics; 3 fs step with hydrogen mass repartitioning; PME electrostatics). TgT_g is the breakpoint of a continuous two-segment linear fit of specific volume against temperature, and each structure is simulated from 3 independent builds. We simulated the 16 designed copolymers. Of these, 11 move MD TgT_g in the intended direction. We list several examples here and our design aligns with MD well:

Starting copolymerDesigned repeat unitObjectivePredicted (∘^\circC)MD (∘^\circC)
caprolactone-rich random copolymer (6% ester-amide)caprolactone →\to *CCC=NOP(N)N*raise+57.6+57.6+73.0±3.1+73.0\pm3.1
caprolactone-rich random copolymer (5% ester-amide)caprolactone →\to *CCC=NP(N*)NOraise+59.4+59.4+74.3±12.7+74.3\pm12.7
poly(trimethylene oxide)-bb-polycaprolactone (9% TMO)TMO →\to *CCNN*, caprolactone →\to *CCN=NP(N*)NOraise+81.8+81.8+62.7±11.8+62.7\pm11.8
caprolactone-rich random copolymer (4% hydroxy-ester)caprolactone →\to *CCCCCP(N*)NOraise+57.9+57.9+38.0±5.0+38.0\pm5.0
MMA-rich random copolymer (3% methacrylic acid)MMA →\to *CC(*)(C)C(=C)OClower−49.9-49.9−32.7±22.7-32.7\pm22.7

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global response

We are grateful to the reviewers for their constructive feedback and for recognizing the strengths of our work, including: (i) the clean separation of unit chemistry from chain architecture, with architecture encoded as a boundary condition (WL7q), (ii) the strong predictive results, especially the PoLyInfo transfer achieved with a small model trained only at R=1R=1 (BSeN, T9wf), (iii) the extensive and transparent evaluation, with every cell reported over seeds and with results that do not favor our own method (WL7q, T9wf), and (iv) the reproducibility, with code, splits, checkpoints and per-row predictions (T9wf).

A revised manuscript has been uploaded, where major revisions are highlighted in blue. We have clarified the terminology and the theoretical claims, added new baselines and statistical tests, tested physical chain lengths and much longer chains, built an independent MD oracle for TgT_g, and performed new attribution and topology experiments. Key contributions clarified during rebuttal:

  1. Stronger and fairer baselines. We added invariant baselines (DeepSets and a unit GNN with identical node features), reran GRIN with 5 seeds, and included wD-MPNN, the model of Aldeghi & Coley, both under their own random CV and under our splits. We now report sample standard deviations, 95% intervals and paired significance tests. PolyBC is more accurate than GRIN on cross-architecture transfer in every setting, with all intervals excluding 0, and more accurate than wD-MPNN under Coley's own protocol (0.99860.9986 vs. 0.99650.9965 on EA).

    Setting (transfer R2R^2)GRINPolyBC-FPolyBC-M
    PoLyInfo-Full0.714±0.0220.714\pm0.0220.811±0.013\mathbf{0.811\pm0.013}0.777±0.0140.777\pm0.014
    PoLyInfo-Small0.519±0.0920.519\pm0.0920.682±0.0250.682\pm0.0250.737±0.022\mathbf{0.737\pm0.022}
    Coley EA0.893±0.0030.893\pm0.0030.938±0.002\mathbf{0.938\pm0.002}0.929±0.0030.929\pm0.003
    Coley IP0.875±0.0030.875\pm0.0030.925±0.001\mathbf{0.925\pm0.001}0.906±0.0010.906\pm0.001
  2. Repetition invariance at physical and long chain lengths. We clarified that a repeat unit may contain several monomers. We computed new xTB labels at N=4,8,16N=4,8,16 for 36 Coley copolymers: the labels change little with length (relative L2L_2 change from N=8N=8 to 1616: 7.3% for EA and 3.0% for IP). PolyBC stays stable at R=8R=8, R=100R=100 (EA/IP) and R=500R=500 (TgT_g) and outperforms GRIN at every length (e.g., TgT_g transfer R2R^2 at R=500R=500: 0.8100.810 vs. 0.7110.711).

  3. Independent TgT_g oracle. We built an all-atom MD protocol for TgT_g and simulated the designed copolymers together with their starting copolymers: 11 out of 16 designs move the MD TgT_g in the intended direction.

  4. Attribution of the gains. Removing the boundary condition lowers transfer R2R^2 on PoLyInfo-Small from 0.6990.699 to 0.6460.646 (Δ=−0.052\Delta=-0.052 [−0.087-0.087, −0.017-0.017]). Replacing the distilled features by direct ECFP gives the same transfer accuracy (0.8080.808 vs. 0.8090.809), while only the distilled features are differentiable for design. Without the homopolymer rows, PolyBC-F still leads the chemical language models. We also tested a second-order end condition, which performs comparably to the Neumann rule.

  5. Beyond linear chains. On the ToPoRg-18k benchmark of branched, star, comb, cyclic and dendrimer polymers, PolyBC reaches R2=0.997R^2=0.997 in distribution and 0.9870.987 on longer chains, ahead of GRIN with the same readout.

  6. Real-data clarifications. We now state that multi-TgT_g records are averaged, and show that retraining without the two-transition records leaves transfer R2R^2 unchanged (0.809→0.8160.809\to0.816). PolyBC also predicts the lower TgT_g of these records (R2=0.886R^2=0.886). We relabeled the soft-alternating chains, defined all acronyms, and showcased our PolyBC contains enough conformational information.

  7. Fair generative baselines. We retrained MOSES VAE and AAE on the same data as PolyBC's decoder and report validity, uniqueness and novelty under identical scoring.

Overall: PolyBC separates unit chemistry from chain architecture by encoding architecture as a boundary condition of a shared operator. With more fair baselines, tests at physical and much longer chain lengths, an independent MD oracle, and attribution and topology experiments, the rebuttal shows that PolyBC is more accurate than GRIN and wD-MPNN on cross-architecture transfer, stable from R=1R=1 to R=500R=500, and applicable beyond linear chains. We believe these contributions make PolyBC a useful and transparent tool for copolymer property prediction and meet the scientific bar of the conference.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.