acceptodds
Under review as a conference paper at ICLR 2027

Functional Group-Guided Tokenization for Molecular Representation Learning

Abstract

As a foundation for molecular property prediction and drug discovery, molecular representation learning has attracted considerable attention. Molecules are naturally represented as graphs composed of atoms and bonds. So, most existing methods employ graph neural networks to learn molecular representations. However, these methods ignore functional groups, which play a critical role in determining molecular properties. As a result, the learned representations may fail to capture the intrinsic information that determines molecular properties, thereby limiting their out-of-distribution generalization. To address this issue, we propose a functional group-guided tokenization method that transforms a molecular into a token graph to capture both functional groups and their interactions. This information is then integrated to obtain the final molecular representation. Extensive experiments on benchmark data sets demonstrate that our method effectively improves the out-of-distribution generalization, showing its ability to learn intrinsic molecular representations that remain invariant across distributions.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Molecular representation learning, Functional group, Out-of-distribution generalization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pzJhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes functional-group-guided soft tokenization for molecular property prediction. A predefined set of 53 functional-group patterns reweights molecular bonds in a graph-cut regularizer, guiding atom-to-token assignments between two graph encoders. The method reports improved performance across GOOD and DrugOOD distribution-shift settings.

Critical strengths

Using chemical substructures to guide learned graph coarsening is a reasonable inductive bias for molecular OOD prediction. The reported improvements cover multiple shift settings across GOOD and DrugOOD. The same-architecture comparison with zero cut-loss weight in Table 5 also provides useful evidence that the regularizer improves prediction on GOOD-ZINC, although it does not isolate the contribution of functional-group-specific weighting.

Critical weaknesses
  1. There are inconsistencies in the chemical definitions underlying the method. In Tables 1 and 6, for example, benzene is associated with C1CCCCC1, a cyclohexane pattern, rather than an aromatic pattern such as c1ccccc1. If these are limited to the tables, please clarify the patterns used in the implementation.

  2. The cut objective admits a degenerate solution, and it is unclear how this is avoided. For any strictly positive probability vector cc, let M=1nc⊤M=\mathbf{1}_n c^\top. This satisfies the row-wise softmax constraints. Since L1n=0L\mathbf{1}_n=0,

    Lcut=tr⁡(M⊤LM)=0\mathcal{L}_{\mathrm{cut}}=\operatorname{tr}(M^\top LM)=0.

    Moreover, Eq. (9) gives Xit=ci∑pXˉpnci=1n∑pXˉpX_i^t=\frac{c_i\sum_p\bar X_p}{nc_i}=\frac{1}{n}\sum_p\bar X_p for every token, so all pooled token node features are identical. Equation (13) includes only the task and cut losses and does not specify an anti-collapse term.

Questions for authors
  1. Could the authors clarify the patterns actually used for the entries in Table 1 and whether correcting the inconsistencies affects the reported results?

  2. Is there an anti-collapse mechanism not described in the formulation? If the task loss alone is relied upon, could the authors provide quantitative evidence on assignment diversity and token redundancy to show that constant assignments are avoided in practice?

  3. How are the 53 patterns selected, and what fraction of atoms and bonds receive effective FG guidance in training and test molecules? Table 9 reports occurrence counts, not coverage or performance on poorly covered molecules. Do the benefits persist for structures outside the predefined patterns? Have alternative priors been considered, such as BRICS-derived fragments or the ring/FG decomposition used in ReLMole?

  4. Are ID/no-shift results available to assess whether the OOD gains incur a trade-off in ordinary predictive performance? The reported experiments cover GOOD/DrugOOD shift settings, without a separate conventional MoleculeNet evaluation beyond the HIV-derived GOOD setting.

  5. The reference cited for GOOD-ZINC appears to omit several authors. Could the authors check the author list against the full citation below?

Rafael Gómez-Bombarelli, Jennifer N. Wei, David Duvenaud, José Miguel Hernández-Lobato, Benjamín Sánchez-Lengeling, Dennis Sheberla, Jorge Aguilera-Iparraguirre, Timothy D. Hirzel, Ryan P. Adams, Alán Aspuru-Guzik; Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules. ACS Cent. Sci. 28 February 2018; 4 (2): 268–276.

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

The paper proposes a functional group-guided tokenization method that transforms the molecule into a token graph to capture both functional groups and their interactions, which are integrated to get the final molecular representation. The authors compare the proposed method with other baselines on GOOD-HIV, GOOD-ZINC, and DrugOOD datasets.

Critical strengths

None

Critical weaknesses
  • The paper claims to be the first to inject FG domain knowledge into MRL (lines 088–090), but it doesnt discusses or compares against established FG/motif-aware methods (e.g., GROVER[1], Mole-BERT[2], KANO[3], and KPGT, which is cited but not compared). Also it does not compare against the graph-pooling methods its tokenization module closely resembles (DiffPool[5], MinCutPool[6]). Evaluating at least a subset of these under the GOOD/DrugOOD splits is necessary to show that the gains come from the proposed FG-guided tokenization rather than from FG information or hierarchical pooling in general.

  • The selected hyperparameters (γ\gamma, KK, kk, d1′d_1' and learning rate) differ for each of the 14 evaluation settings (Table 8). The best values in the sensitivity analyses (Tables 4–5) are exactly the values used for the main results. For example,$\gamma = 1.0 $,for scaffold, 0.1 for size/covariate and 0.001 for size/concept on GOOD-ZINC. The paper never states that these were selected on the validation split. In OOD benchmarks the validation and test distributions differ, so selecting on test leaks information.

[1] Self-Supervised Graph Transformer on Large-Scale Molecular Data, Rong at al. NeurIPS 2020,

[2] Mole-BERT: Rethinking Pre-training Graph Neural Networks for Molecules, Xia et al., ICLR 2021,

[3] Knowledge graph-enhanced molecular contrastive learning with functional prompt, Fang et al., Nature Machine Intelligence 2023,

[4] A knowledge-guided pre-training framework for improving molecular representation learning, Li et al., Nature Communications 2023,

[5] Hierarchical Graph Representation Learning with Differentiable Pooling, Ying et al., NeurIPS 2018,

[6] Spectral Clustering with Graph Neural Networks for Graph Pooling, Bianchi et al., ICML2020

Questions for authors
  1. What is MPL?
  2. What is bound?
  3. Why is GOOD-PCBA missing from benchmark evaluations? In line 329 the authors mention that they "adopt the experimental protocol of iMoLD" but they do not include GOOD-PCBA which is the hardest of the three.
  4. Why DrugOOD is limited to LBAP-core IC50/EC50, excluding the SBAP (protein/target-shift) tasks and the refined/general noise levels?

Suggestions

  • MILI is listed as a baseline (line 362) but is missing from Table 2

  • The legends in Figures 3 and 4 can’t be read

  • In Table 1, C1CCCCC1 is cyclohexane, not benzene (c1ccccc1)

  • In Table 6, N1CCCCC1 is piperidine, not pyridine C1CCOC1 is tetrahydrofuran, not furan [NH]1CCCC1 is pyrrolidine, not pyrrole

  • “Ccid_halide” in Table 6 should be “Acid_halide”

Reviewer_5UFLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a functional group-guided tokenization framework for molecular representation learning under distribution shifts. The method uses predefined functional groups to construct a weighted molecular graph and guides soft atom clustering through a graph cut regularization objective. The resulting token graph is processed by a second graph neural network for molecular property prediction. Experiments on GOOD and DrugOOD report improvements over existing OOD generalization baselines.

Critical strengths
  1. Incorporating chemical domain knowledge into molecular representation learning is a meaningful and intuitive direction.

  2. The method reports consistent improvements across multiple molecular OOD evaluation settings.

Critical weaknesses

The proposed regularization loss admits trivial solutions where all atoms have identical soft assignments, resulting in zero graph cut loss. Without explicit constraints encouraging diverse or meaningful tokens, it is unclear how the method avoids degenerate tokenization.

The qualitative visualizations are encouraging, but they do not establish that the learned tokens consistently correspond to meaningful functional groups. Quantitative evaluations of token quality would strengthen this claim.

Comparing against zero graph cut regularization does not distinguish the contribution of functional group guidance from generic graph smoothing. Additional comparisons using the original adjacency matrix or alternative edge weighting schemes would be helpful.

The methodological novelty is limited. The main components, including soft graph partitioning, Laplacian regularization, graph pooling, and GPS-based representation learning, are largely established techniques. The contribution appears to lie primarily in their combination with predefined functional group information.

Questions for authors

Please refer to the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.