acceptodds
Under review as a conference paper at ICLR 2027

Synthon-CGR: Co-crystal Formation Prediction with Crystal-Supervised Synthon Reasoning

Abstract

Co-crystallization offers a practical way to tune the physicochemical properties of molecular compounds without changing their covalent structure, but finding compatible molecular pairs still requires substantial experimental effort. Most machine-learning approaches treat this problem as binary classification and are trained only on pair-level formation labels. As a result, they provide limited information about the intermolecular motifs that support a predicted co-crystal. We introduce Synthon Candidate Graph Reasoner (Synthon-CGR), a graph neural network that predicts co-crystal formation together with a ranked set of candidate supramolecular synthons. The model combines molecular graph representations, functional motifs, and crystal-derived supervision so that intermolecular interactions become explicit prediction targets rather than post-hoc explanations. We curate literature and crystallographic data, augment the training pool to 27,718 molecular pairs to address the scarcity of reported negatives, and derive synthon labels from resolved crystal structures. On a pair-disjoint held-out test set, Synthon-CGR reaches F1 = 0.893 for co-crystal formation and HitRate@5 = 0.935 on the synthon-ranking subset. We also evaluate the model on an independent dataset obtained from co-crystallization experiments conducted in our laboratory and compare predicted interaction motifs with expert annotations for experimentally confirmed co-crystals. Together, these experiments show that crystal-derived interaction supervision can extend pair-level co-crystal prediction with explicit and experimentally testable hypotheses about molecular recognition.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
graph neural network, transfer learning, co-crystal prediction, supramolecular synthon, interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_nqCTOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Synthon-CGR, a GATv2-based model that predicts co-crystal formation and ranks supramolecular synthons (interacting functional groups of the two molecules). The authors curate a dataset of molecular pairs and derive synthon labels from resolved crystal structures. They show that the full model outperforms its ablations on a held-out test set, and report results on internal co-crystallization experiments, where it performs similarly to the strongest baseline.

I want to preface the review by stating that my knowledge of crystal engineering and molecular chemistry is very limited, and that my comments on the chemical or physical aspects of the work are, at most, educated guesses. However, I have a background in geometric deep learning; therefore, I am quite certain about my assessment of the machine learning parts of the review.

Critical strengths
  • The paper is well motivated: the introduction does a good job of conveying why this setup is important and the current limitations of the datasets (most datasets only record whether a pair formed a co-crystal, and they are biased) and of the models (which only predict whether two molecules form a co-crystal).
  • The authors propose a dataset creation and synthon annotation pipeline, which could be very useful from a practical standpoint for practitioners interested in co-crystal engineering.
  • The method, while relatively simple from an ML perspective, seems to be effective.
  • I appreciated the inclusion of experimentally validated results from the authors' laboratory.
  • The paper is relatively easy to follow and understand.
Critical weaknesses
  • (This is not greatly influencing my score.) The paper's aesthetics and writing could be improved:
    • Aesthetics: I do like Figure 1, but I would recommend that the authors improve the other figures (e.g., Figure 2 and the figures in the appendix), for instance, by increasing the font size and including them as PDF/SVG so that they are rendered as vector graphics.
    • Writing: I am not a fan of including so many numbers in the prose when discussing the experiments and results. For instance, when describing the dataset in Section 5.1, I would recommend using a table (e.g., Table 7 from the appendix) instead of listing the number of samples. The same holds for Section 6: referring to the tables and discussing the results at a higher level would help the reader understand them better.
  • From an ML perspective, the experiments and models are quite simple, and I am unsure whether they are conclusive enough. The simplicity of the pipeline is a strength, but the experimental evaluation has some flaws:
    • As far as I can tell, the number of seeds is not reported. The differences are generally large enough to be somewhat convincing, but since both the dataset and the models are small, I believe that a mean and std over at least 3 seeds should be reported for all experiments.
    • No hyperparameter search is reported (except for the backbone comparison in Table 10). At this (small) scale, the hyperparameters could be crucial.
    • The results reported in Table 3 on the experimental validation set seem relatively weak: Synthon-CGR is essentially tied with DeepCocrystal (balanced accuracy 0.636 vs. 0.632), and with only 17 positives the difference is within noise.
    • The ablation study is confusing and hard to follow; several changes are made at once between the variants in Table 1. As the authors themselves note, "The comparison between V2 and V3 should not be read as isolating a single architectural factor because the contextual reasoner and the training regime change together." A V3 variant fine-tuned without the synthon loss would isolate the paper's central claim.
    • For the @K synthon-ranking metrics (Table 2), it would be important to include baselines (e.g., a random predictor): with a median of six candidates per pair, a random ranking already reaches HitRate@5 = 5/6 ≈\approx 0.83 on a six-candidate pair with a single positive.
    • Some numbers do not add up: the Table 1 metrics are inconsistent with a balanced 350/350 test set (for V3, precision 0.884 and recall 0.902 imply a balanced accuracy of ≈\approx 0.89, not 0.873), and 245 masked test pairs plus 729 masked Stage 2 pairs give 974, not 916 (Section 3.2).

Overall, I appreciate the motivation behind the proposed method and analysis --- the overall narrative and intuitions are strong, with the model architecture and dataset construction pipeline being reasonable. However, I believe the writing can be improved and that the paper has limitations in its empirical evaluation. I recommend rejection, but I encourage the authors to iterate on the manuscript a few more times and strengthen their empirical evaluation; the work has the potential to be a good paper.

Questions for authors
  • Have the authors performed any hyperparameter search? How is the validation set constructed?
  • Have the authors used multiple seeds for their experiments?
  • Could any other meaningful baselines be added for the experimental evaluation?
  • Will the curated dataset, the synthon annotations, and the laboratory dataset be released? This might be a major contribution of the paper.
Reviewer_qZXuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Synthon-GCR predicts co-crystal formation from a synthons that two molecules might form. A synthon is a combination of atoms from each of two molecules that are likely to interact in a co-crystal structure. Their model is trained on publically available co-crystal formation data from the Crystallography Open Database.

Critical strengths

The paper presents an interesting approach for predicting co-crystallisation from synthons rather than just a representation of the molecule itself. Synthons capture atoms that could interact between the two molecules, thus, capturing the interactions of the molecules rather than just using their identity.

The authors train their model on a large, curated dataset of 15,606 unique molecular pairs that they curated across 11 literature sources and the Crystallography Open Database.

Critical weaknesses

Performance metrics lack confidence intervals, making it difficult to determine if improvements are meaningful. The authors should include confidence intervals, or variance across training seeds to demonstrate if their model is actually a meaningful improvement over their ablations and baselines.

The validation of the synthons used is also limited to 17 co-crystals and only 6 matched the highest-ranked model prediction. More extensive validation of the synthons used by the model is necessary to determine if the synthons ranked by the model are accurate. The paper could also benefit from comparing with other baselines for chemical interactions on synthon identification.

Limited quantitative analysis of model performance, the authors only show improvements on one dataset with baselines DeepCocrystal and CCGNet. How do these models perform against the dataset used in table 1?

Authors ablate overall architecture in Table 1 but do not adequately ablate components of their synthon ranking performance. Do the contextual attention, crystal-derived supervision and statistical interaction priors improve synthon ranking metrics? Since synthon reasoning is a key contribution of the paper more detailed ablations are necessary to understand what components of the synthon ranking are most useful.

Questions for authors

Please evaluate against Synthon-CGR against additional co-crystal prediction methods against the dataset reported in table 1.

Please evaluate Synthon-CGR synthon ranking capabilities against additional baselines such as random ranking, a simple interaction prior based on common synthons found in the training set, simple heuristics-based approach for identifying synthons such as hydrogen bond donor-acceptor compatibility.

Furthermore, also include ablations of key components of the synthon ranking module and evaluate its performance on synthon ranking.

Report confidence intervals and variation across training seeds for the main results especially for table 3 where there are small improvements compared to DeepCocrystal.

Include further analysis of the synthons being ranked in terms of performance across interaction types, functional groups involved, molecules being co crystalised. This will help readers understand the quality of synthon ranking.

Reviewer_2QTAOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose Synthon-CGR, which jointly predicts co-crystal formation together with candidate synthons in co-crystal structures. The model combines molecular representations with contextual reasoning over candidate synthons and is trained with synthon annotations extracted from crystal structures. The authors evaluate co-crystal formation prediction and synthon ranking. The authors also consider experiments with their laboratory data.

Critical strengths
  • External experimental validation. The authors consider external evaluations on their laboratory data for both co-crystal formation and synthon ranking predictions.
Critical weaknesses
  • Lack of comparison with structure prediction methods. Existing cutting-edge models, e.g., OXtal and Packora, can directly predict co-crystal structures, addressing a more comprehensive task than synthon prediction. Without comparison, I was wondering whether focusing on synthon prediction offers meaningful benefits over directly predicting the structures.

  • Unclear significance of synthon prediction. The paper only reports the proposed model's synthon ranking performance. It remains unclear whether the predicted synthons provide meaningful benefits for downstream tasks, e.g., co-crystal design, compared to existing structure prediction methods.

Questions for authors
  • Can the authors compare synthon predictions with motifs extracted from structures generated by OXtal or Packora? I was also wondering whether synthon prediction can provide further utility compared to co-crystal structure predictions.

  • Can the authors experimentally show how their synthon predictions benefit meaningful downstream tasks, such as co-crystal design?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.