- 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
- 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.
-
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.
-
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.
-
Page 2 contains leftover editing text ("Here's the polished version, with your LaTeX kept:").
Questions for authors
- What happens if the distilled node features are replaced by directly computed ECFP?
- How much of the PoLyInfo transfer gain remains after removing the replicated homopolymer rows?
- Can the boundary-condition intervention be run on cross-architecture transfer cells?
- Why is wD-MPNN not included as a baseline?
- Does the nonlocal-operator interpretation impose any constraint beyond the choice of ghost rules?
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% confidence intervals. We find that our boundary condition improves the accuracy, especially on the transfer cells.
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):
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 Tgs, entered once per sequence type (617×3), and they share no record with the test set. Removing them lowers PolyBC-F's transfer R2 from 0.809±0.009 to 0.574±0.009, so we state their role prominently. Without them, PolyBC-F still leads the chemical-LM chain models on transfer (MoLFormer 0.487±0.009, ΔR2=−0.087 [−0.184, −0.008]; TransPolymer 0.398±0.028, ΔR2=−0.176 [−0.502, −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.
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±0.0000 vs. 0.9965±0.0002) and on the pooled pair split (0.9962±0.0001 vs. 0.9936±0.0003). Under the manuscript's settings, PolyBC is more accurate on Coley IP (0.947±0.001 vs. 0.810±0.022), PoLyInfo-Full (0.815±0.006 vs. 0.757±0.035) and PoLyInfo-Small (0.727±0.018 vs. 0.620±0.048).
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.