acceptodds
Under review as a conference paper at ICLR 2027

Shared Motifs, Specific Matches: Learning Protein–Glycan Compatibility

Abstract

Predicting protein–glycan binding for unseen molecules requires recognition patterns that transfer across measured pairs. Recent approaches provide increasingly detailed glycan representations, but models must still learn which local structures matter for each protein from labels assigned to complete molecular pairs. We organize this supervision around recurring motifs: an unseen glycan can share local patterns with measured glycans, while their contribution to binding depends on the protein partner. We introduce **MotifMatch**, which learns one shared parameter vector per motif from measurements on all glycans carrying it. Explicit, linkage-aware motif counts combine these vectors into glycan factors, which sequence-derived protein factors rescale through Hadamard matching. We show that gradient coupling through the shared motif parameters factorizes into motif overlap and the alignment of pair-conditioned responses. Fixed-budget experiments show that measurements on glycans with shared motifs improve prediction for held-out glycans. Controlled comparisons further show that the benefit of Hadamard matching depends on the glycan representation. On GlycanML, MotifMatch achieves the highest mean ranking scores for unseen proteins, unseen glycans, and pairs where both partners are unseen. With protein features cached for all methods, MotifMatch trains and infers faster than the fastest baseline in each stage.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein–glycan binding prediction, Molecular representation learning, Interaction modeling, Cold-start generalization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dv87OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces MotifMatch for protein–glycan binding response prediction. It combines explicit, linkage-aware glycan motif counts with frozen ESM2 protein features through projected factors and Hadamard matching. The authors derive an overlap–alignment factorization of gradient coupling and investigate cross-glycan transfer using fixed-budget support experiments. On GlycanML, the method achieves the highest mean ranking scores across protein-cold, glycan-cold, and double-cold settings, with lower training and inference times under the reported implementation.

Critical strengths

The controlled representation–scoring comparisons provide a useful empirical insight. The paper goes beyond comparing glycan encoders under a single prediction head: it crosses motif and circular representations with concatenation and Hadamard scoring, and includes multiplicity and bilinear controls. On the official protein-cold split, Hadamard scoring improves NDCG@20 from 0.4330 to 0.4708 with motif counts, whereas it provides no corresponding improvement with circular features. This supports the consequential observation that the effectiveness of a molecular representation depends on how the interaction between partners is modeled.

The fixed-budget support experiment meaningfully strengthens the empirical investigation. Replacing motif-negative observations with motif-containing support observations while preserving the total measurement budget and each protein’s observation count is a thoughtful way to investigate cross-glycan transfer. Together with the competitive protein-cold ranking results, this makes the proposed approach a useful and computationally lightweight contribution, beyond a straightforward benchmark comparison.

Critical weaknesses

The evidence for an advantage in the additional cold-start settings is not sufficiently conclusive. Although MotifMatch achieves the highest mean ranking scores, several margins are very small relative to the reported variability. For example, glycan-cold Protein Spearman is 0.1971 versus 0.1967 for GIFFLAR, and double-cold NDCG@20 is 0.4751 versus 0.4702 for LectinOracle. Only three runs are reported, with split and training seeds varied together in these settings, and no paired uncertainty analysis is provided. Moreover, double-cold NMSE is 1.0030 and Protein Spearman is 0.0991, indicating limited absolute predictive performance. These results establish competitiveness more clearly than a reliable generalization advantage.

The mechanistic contribution is less established than the paper’s framing suggests. The overlap–alignment factorization follows directly from the outer-product gradient of the shared linear count projection and, as Appendix A.2 acknowledges, also applies to more general scoring functions. The support experiments do not directly measure response alignment or establish its relationship with observed transfer. Adding motif-containing glycans also changes coverage of other co-occurring structures, while the reported support gains have substantial seed variability. Consequently, the experiments support the usefulness of the proposed combination, but do not yet convincingly isolate the proposed mechanism. This matters because the architecture otherwise combines established components, making the strength of the explanatory contribution important to the paper’s novelty.

Questions for authors

Statistical clarification: Can you provide paired differences and confidence intervals for the principal ranking metrics against the strongest baselines in the glycan-cold and double-cold settings? Please specify the independent evaluation units and account for dependence among observations sharing proteins or glycans. This would help determine whether the small mean improvements are reliable.

Mechanism clarification: Is there direct evidence connecting the response-alignment term in Proposition 1 to the observed support-transfer gains? In particular, do overlap and alignment predict actual target-loss changes, and how does the support-selection protocol account for changes in co-occurring motifs? Please distinguish what the existing experiments establish from what remains an interpretation.

Experimental fairness clarification: What hyperparameter-search budgets and pretrained components were used for the baselines? For the efficiency comparison, were independently computed entity representations cached and reused for baselines wherever their architectures permit this? These details would materially affect my interpretation of both the accuracy comparisons and the reported speedups.

3 comments

Thank you for reading our work so carefully and for suggesting concrete ways to strengthen the evidence. We appreciate your recognition of the insight from the controlled representation–scoring comparisons, the value of the fixed-budget support design, and the practical contribution of a lightweight method with competitive protein-cold ranking. Your questions helped guide our expanded cold-start evaluation with paired uncertainty analyses (§§4.1–4.2; Appendix C.3, Table 10) and the new controlled response-alignment experiments (§4.3, Figure 4; Appendix C.7); we also provide clearer explanations of the training, pretraining, caching, and timing protocols below (§4.5, Table 5; Appendices C.2 and C.5). These additions and clarifications are included in the revised manuscript and marked in blue, so that you can readily check them against your concerns. We hope these additional results and clarifications help address your concerns about the reliability of the ranking gains, the connection between response alignment and transfer, and the fairness of the comparisons.

Small glycan-cold and double-cold gains relative to run variation, together with limited double-cold accuracy, do not yet establish a reliable advantage. Please provide paired differences and intervals against the strongest baselines, clarify evaluation units, and account for shared proteins and glycans. (W1, Q1)

We understand your concern. In these challenging cold-start settings, small mean margins need closer examination in light of the variation across runs. Double-cold response prediction is particularly difficult: all evaluated predictive models have mean NMSE close to the constant-reference value of 1 (Table 9). Following your suggestion, we therefore examined whether the ranking gains persist across data splits and remain supported after accounting for dependence among observations sharing proteins or glycans.

We expanded the evaluation to four models × five splits × three training seeds, totaling 60 model runs. Comparators were fixed before the expansion using the strongest submitted baseline for each metric, and paired methods were evaluated on identical test observations. Protein Spearman and NDCG@20 are averaged over eligible proteins, while Global Spearman uses all test pairs. We average the three training seeds within each split, then use the five paired split differences to construct a t-based 95% interval. The split is the replication unit for this interval.

As you noted, shared entities introduce a further source of dependence. We therefore also resample protein similarity clusters and glycan similarity clusters jointly, recomputing the metrics in each replicate. The same cluster weights are shared across methods and runs, preserving the links among repeated entities. These intervals assess sensitivity to test-entity composition conditional on the fitted models, complementing the split-level analysis. The full six comparisons are reported below.

SettingMetricBaselinePaired ΔSplit-level 95% CICluster 95% CIPositive splits
Glycan-coldProtein SpearmanGIFFLAR+0.0077[−0.0067, 0.0221][−0.0170, 0.0259]3/5
Glycan-coldNDCG@20LectinOracle+0.0035[−0.0172, 0.0242][−0.0150, 0.0216]3/5
Glycan-coldGlobal SpearmanGIFFLAR+0.0039[−0.0103, 0.0181][−0.0125, 0.0197]4/5
Double-coldProtein SpearmanBiLSTM+0.0045[−0.0225, 0.0314][−0.0187, 0.0326]3/5
Double-coldNDCG@20LectinOracle+0.0147[0.0066, 0.0229][0.0028, 0.0268]4/5
Double-coldGlobal SpearmanLectinOracle+0.0097[−0.0129, 0.0324][−0.0111, 0.0309]3/5

Δ denotes MotifMatch minus the baseline, computed before rounding. Intervals are pointwise 95% intervals without multiplicity adjustment; positive splits are counted after averaging training seeds. Computational details are in Appendix C.3.

The additional analysis gives us a clearer answer for the hardest setting: the double-cold NDCG@20 gain is +0.0147, positive in four of five splits, with both intervals above zero. This supports a modest but robust improvement in prioritizing top candidates. The submitted double-cold NDCG@20 also exceeds the constant reference (0.4751 versus 0.3928, Table 9), illustrating the practical ranking value alongside the response-accuracy evaluation. We hope these experiments help address your concerns!

The factorization alone does not establish the mechanism behind variable support gains. Can response alignment predict actual loss changes beyond shared structure, while accounting for co-occurring motifs? Please distinguish direct evidence from interpretation. (W2, Q2)

Our theoretical analysis links motif overlap and protein-conditioned response alignment to local target-loss changes (Proposition 1), but as you point out, we needed to connect this theoretical account more directly to the experimental transfer results. We therefore added a controlled local-update test and an analysis of the existing complete-training gains.

Across 36 motif–run conditions (12 motifs × three seeds), we compared small updates from the same k = 0 checkpoint using the original training set or its k = 16 support replacement, and measured the difference in held-out target loss. Only shared motif-projection columns were updated, with total measurements and per-protein observation counts fixed. We included added-minus-removed contributions across all motif coordinates, so changes in co-occurring structures entered the loss-change prediction.

To isolate the contribution of response direction in the diagnostic, the control preserves motif overlaps, residuals, and sensitivity magnitudes, replacing only response inner products with products of norms. Both scores use the same support and target observations and structural information. Retaining direction reduces relative L1 error from 32.05% to 0.212%. This supports the distinction made by our theory: motif overlap identifies shared-parameter connections, while response alignment and residuals characterize how updates through those connections affect target loss.

EvaluationFull scoreWithout directional alignment
Local update: Spearman with measured gain ↑1.0000.891
Local update: correct change direction ↑36/3630/36
Local update: aggregate relative L1 error ↓0.212%32.05%
Training endpoint: Spearman with standardized MSE gain ↑0.9300.846

Endpoint gain is the target MSE reduction normalized by the k = 0 training-label variance; definitions are in Appendix C.7.

We then asked whether initial coupling relates to the k = 0→16 complete-training gains. Across 12 motif means, each averaging three seeds, the correlation is 0.930 for the full score versus 0.846 without direction. These endpoints come from separately trained models, connecting the initial diagnostic to differences in the eventual benefit of shared-motif support. The local experiment directly tests the predicted loss change; the endpoint analysis establishes an association with complete-training transfer. Together, they provide complementary empirical support for our theoretical account of how supervision is shared across glycans in a protein-dependent manner. Thank you for helping us make this experimental validation more complete!

What search budgets and pretrained components were used for the baselines? Were entity representations cached and reused wherever their architectures permit in the efficiency comparison? (Q3)

We apologize for not making these protocols sufficiently clear in the submission. Thank you for highlighting how they affect the interpretation of both accuracy and speed.

Our guiding principle was to follow the original authors’ released implementations and published experimental settings as closely as possible when adapting each baseline to the common protein–glycan regression task. We evaluated each baseline using its adopted configuration, without conducting additional hyperparameter search. Table 6 in Appendix C.2 lists the adopted configurations and per-run training limits for the official protein-cold benchmark. Each configuration was evaluated over three training seeds, with validation-based checkpoint selection.

To maintain comparability on the shared task, all benchmark methods use the same frozen, mean-pooled ESM2-650M protein features. Glycan encoders and prediction heads are trained on the binding task without loading additional pretrained glycan weights.

We also use the entity-reuse mechanisms available in the evaluated implementations. All methods use precomputed protein features and prepared glycan inputs. During training, learned encodings are recomputed as parameters change. The per-method caching procedures are detailed in Appendix C.5.

Timing begins after input preparation and model setup, and excludes ESM2 feature extraction. Training includes optimization, validation, and early stopping; inference uses prepared inputs and loaded models over the same 63,364 test pairs. These measurements therefore compare the evaluated implementations with their actual reuse procedures. Under this protocol, MotifMatch requires 25.63 s for training and 0.291 s for inference, with 1.016 M trainable parameters. We hope these details provide the context needed to assess the accuracy and efficiency comparisons.

Dear Reviewer dv87,

Thank you again for your detailed feedback. We have shared our responses along with the additional statistical and mechanistic analyses above. As the discussion deadline is approaching, we would be very grateful to hear whether our replies have addressed your concerns or if any questions remain.

Best regards,

The Authors

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

This paper introduces MotifMatch, which is a model predicting glycan recognition by proteins. It uses linkage-aware motifs of monosaccharide motifs with terminal context to represent glycans, protein representations from ESM2, and protein sequence-conditioned pooling with a Hadamard product of these motifs into prediction for what protein binds to a glycan. The authors analyze performance of their model on unseen proteins, unseen glycans, and both unseen proteins and glycans.

Critical strengths

The empirical performance in Table 1 shows that the model achieves the best ranking of glycans for the unseen protein benchmark. The architecture is also faster to train and run inference on compared to baselines. Their model performance also scales with increasing distinct training glycans that contain the target motif.

Critical weaknesses

Limited novelty in the approach. The contributions in the method to me do not seem substantial enough to warrant acceptance. The Hadamard fusion is an established approach, their pooling of the protein and glycans is also limited in methodological contribution. Mean pooling of the protein results in limited representation of the localized interactions between proteins and glycans. Given that the glycans have a branched structure, authors could have also explored better representations of the glycans which capture the global branched nature than as a count of motifs. The motifs are also defined statically, given the diversity of glycans further innovation could be made in how the motif vocabulary is formed.

Limited comparison to other fusion approaches. The authors compare Hadamard fusion with concatenation and scalar bilinear scoring, but they should also consider FiLM and protein-conditioned motif cross attention.

Lack of improvements for generalizing to new glycans. The improvements over basleines in table 2 are limited, with many within the sample SD over training seeds. Notably under the double-cold evaluation, the NMSE is greater than 1 and so worse than predicting the mean. This gives me limited confidence that the model has learnt generalizable features of protein glycan recognition. Unable to determine the contributions of the motif features. The authors make claims about the contributions of the terminal aware encoding of glycans, and use of motif counts. This should be supported with ablations to determine if these actually lead to model improvements.

Unable to determine the contributions of the motif encoder. The authors make claims about the contributions of the terminal aware encoding of glycans, and use of motif counts. This should be supported with ablations to determine if these actually lead to model improvements.

Questions for authors

What is the maximum sequence identity between a training and test set protein for the held-out protein benchmark in table 1? Clarification on how distinct are these proteins would be useful for understanding generalization capabilities of the model.

What is the contribution of key components of your motif encoder? Please include ablations for key components of the motif encoder, e.g. motif counts, terminal context, etc.

How much does the Hadamard fusion improve over potentially stronger fusion baselines?

Why the choice of mean pooling a protein when glycan binding is localized to carbohydrate recognition domains? Can you compare your model to mean pooling over the carbohydrate recognition domains only?

6 comments

Thank you for your careful review and the concrete experiments you suggested. We are glad you recognized the best glycan ranking on the unseen-protein benchmark, faster training and inference, and improving performance as more distinct training glycans contain the target motif. Your suggestions helped us examine our design choices more directly. The revised manuscript includes the sequence and generalization analyses (§§4.1–4.2; Appendix C.3), mechanism diagnostics (§4.3, Figure 4), component comparisons (§4.4; Appendix C.8, Table 16), and future directions (§5). We have marked these changes in blue so that you can check them easily alongside our responses. We also report the CRD pooling comparison below, and sincerely hope these additional results and explanations address your concerns about the contribution, representation choices, and generalization.

Methodological novelty is limited. Hadamard fusion is established, and the protein/glycan pooling offers little methodological contribution. Global branching representations and innovations in motif-vocabulary construction warrant further exploration. (W1)

We agree that Hadamard fusion and pooling are established operations. We would like to explain what led us to bring them together in this particular way. Our starting observation was that richer glycan encodings did not consistently improve cold-start predictions. This made us look more closely at the supervision available to the model. A binding measurement tells us about a complete protein–glycan pair, but does not identify which local structures contribute to that interaction. The question we pursued was therefore how measurements on different glycans could jointly teach recognition patterns that remain useful for an unseen glycan and depend on its protein partner.

Our methodological contribution is to design the shared learning units and their protein-conditioned use together. In MotifMatch, a motif coordinate identifies a structural pattern and specifies which observations train the same parameter vector. A measurement on one glycan can therefore train parameters that another glycan will use. Protein-conditioned matching then determines how this shared representation contributes to each pair’s prediction. The vocabulary fixes the available structural units, while their vectors and protein-dependent effects are learned from binding data. The overlap–alignment factorization (Proposition 1) characterizes the learning interactions that follow from these choices. It separates structural overlap from pair-conditioned response alignment and, together with prediction residuals, characterizes how a shared-parameter update affects a target locally. This connects the representation and matching design to an explicit, testable account of how supervision transfers across glycans.

The controlled comparisons helped us see why this joint design matters. On the official protein-cold split, motif and circular representations perform similarly under concatenation; with Hadamard matching, motif counts improve from 0.4330 to 0.4708 in mean NDCG@20, while circular representations show no corresponding ranking gain (Table 4). Thus, the benefit depends on which structural units are shared and how the protein-conditioned scorer uses them. The fixed-budget transfer results (§4.3) further support the intended consequence of this design: measurements on other glycans improve predictions for held-out targets without increasing the measurement budget.

We also see your suggestions on global branching representations and vocabulary construction as promising extensions of this framework. Global context could capture relationships among shared local units, while vocabulary learning could adapt the units themselves to glycan diversity. We have added both directions to the Discussion. Thank you for these valuable suggestions.

What is the maximum train–test sequence identity on the held-out-protein benchmark, and how distinct are the test proteins? (Q1)

The maximum train–test global sequence identity is 52.49% on the official GlycanML protein-cold split used in Table 1. To characterize how distinct the held-out proteins are, we exhaustively compared all 139 test proteins against the 1,111 proteins with training binding observations. For each test protein, we recorded its highest global identity to any training protein.

The median of these nearest-training identities is 43.47%, and 131/139 test proteins (94.24%) have their highest training identity below 50%. Identity was calculated as exact non-X residue matches divided by the full global alignment length, including gaps; the alignment procedure is detailed in Appendix C.3. These results quantify the sequence separation between the held-out and training proteins, providing context for the generalization results.

The contributions of motif counts and terminal context are unclear. Please provide component ablations to establish whether they improve performance. (W4, Q2)

Thank you for suggesting these component-level ablations. We agree that they provide a direct way to assess what counts and terminal context each contribute. We therefore compared the two choices in a 2×2 ablation on the official protein-cold split, using counts or binary presence, each with and without terminal context. Protein features, Hadamard matching, and training settings are held fixed, and all variants retain glycan size and branching descriptors.

EncodingNMSE ↓Global Spearman ↑Protein Spearman ↑NDCG@20 ↑
Counts + terminal context0.91600.29980.35340.4708
Binary + terminal context0.91900.29520.34850.4610
Counts, without terminal context0.92520.27750.32790.4607
Binary, without terminal context0.92230.28340.33840.4603

Means over three training seeds; complete-model values follow Table 1.

We found a clear benefit from terminal context. It lets the encoder distinguish the same linked pattern occurring at a glycan terminus from one occurring in the interior. With counts retained, adding this information improves Protein Spearman by 0.0255. The benefit extends across all four metrics in each of the three paired runs, supporting our choice to preserve this positional distinction.

Counts provide a further benefit when terminal context is retained. They preserve how often a motif occurs, beyond whether it is present. In this comparison, counts improve all four mean metrics over binary presence. These results give us greater confidence in retaining both types of information, so that the encoder captures where a pattern occurs as well as how often it appears.

Fusion comparisons are limited. How does Hadamard perform against FiLM and protein-conditioned motif cross-attention? (W2, Q3)

Thank you for suggesting FiLM and protein-conditioned motif cross-attention. We agree that they broaden the comparison to other protein-conditioned mechanisms for using the same motif information. We evaluated both on the official protein-cold split. FiLM learns protein-conditioned scale and shift, while cross-attention uses protein features to query motif tokens and retains motif multiplicity. All three methods share the same input information, main-effect branches, and training protocol.

FusionNMSE ↓Global Spearman ↑Protein Spearman ↑NDCG@20 ↑
Hadamard0.91600.29980.35340.4708
FiLM0.92530.29720.35100.4535
Protein-conditioned motif cross-attention0.93530.28420.34010.4309

Means over three training seeds; Hadamard values follow Table 1. Alternative-head specifications are given in Appendix C.8.

Hadamard achieves the best mean performance on all four metrics, with NDCG@20 higher than FiLM and cross-attention by 0.0173 and 0.0399, respectively. Both NDCG@20 and NMSE improve in all three paired runs against each alternative. These comparisons show that, under the same motif inputs and training protocol, Hadamard’s gains over the two alternatives extend to both response accuracy and top-candidate ranking.

Full-sequence mean pooling may limit the representation of localized protein–glycan interactions. Why use it, and how does it compare with pooling over carbohydrate recognition domains only? (W1, Q4)

To be candid, we initially chose full-sequence mean pooling because ESM2 already provides contextualized residue embeddings, and averaging gives a uniform, low-overhead protein input without requiring CRD annotations. Your suggestion encouraged us to test a biologically focused alternative. We compared the two readouts on a common InterPro/Pfam-annotated subset, using the same frozen ESM2 embeddings, downstream model, and training protocol. The initial comparison favored full-sequence pooling, with 0.0120 higher NDCG@20 and 0.0063 lower NMSE, both in all three paired runs. This supported our original choice under the existing training configuration.

PoolingNMSE ↓Global Spearman ↑Protein Spearman ↑NDCG@20 ↑
Full-sequence mean0.92590.23270.23830.4180
CRD-only mean0.93220.21860.23190.4061

Both tables use the same annotated subset of 153 training, 27 validation, and 5 test proteins (2,253 test pairs), and report means over three training seeds. Paired differences use unrounded values.

Still, CRD-focused pooling seemed promising to us, and we were puzzled by its lack of improvement in the initial comparison. Pooling over a different residue region also changes the numerical distribution of the protein input, so we wanted to understand whether the initial gap partly reflected how that input worked with downstream training. We therefore applied the same fixed rescaling and shift to both readouts, with parameters determined only from training features, allowing us to examine the potential of CRD pooling under a common adaptation.

Pooling with common adaptationNMSE ↓Global Spearman ↑Protein Spearman ↑NDCG@20 ↑
Full-sequence mean0.91320.25070.25220.4343
CRD-only mean0.91030.24950.25400.4525

This time, both readouts improved, and adapted CRD pooling achieved 0.0181 higher NDCG@20 than equally adapted full-sequence pooling, with positive differences in all three paired runs. We are glad we followed your suggestion beyond the initial comparison. It uncovered a better-ranking CRD variant on this annotated subset and clarified how residue selection and readout adaptation interact. Thank you for encouraging this investigation; it has made our evaluation more complete and revealed a concrete improvement to the protein readout.

Improvements on unseen glycans are limited and often within the reported seed variation. Double-cold NMSE exceeds 1, weakening confidence that the model learns generalizable protein–glycan recognition features. (W3)

Thank you for raising this concern. Generalizing to unseen glycans is demanding, especially in double-cold, where both partners are unseen and mean NMSE values across the compared predictive models remain close to the constant-mean reference (Table 9). Response prediction is challenging across methods, with substantial run-to-run variability. We therefore understand your concern about whether the small margins are reliable.

To investigate this, we expanded evaluation to four models × five splits × three training seeds (60 runs). On identical evaluation observations, the double-cold NDCG@20 paired gain over LectinOracle is +0.0147, positive in four of five splits after averaging training seeds. The split-level and protein-cluster × glycan-cluster pointwise 95% intervals are [0.0066, 0.0229] and [0.0028, 0.0268], supporting a modest but robust improvement in top-candidate ranking.

Complementary controls help clarify what contributes to these predictions (Table 4). On unseen proteins, additive-only scoring achieves mean NDCG@20 of 0.1951; the joint branch alone reaches 0.4664, close to the full model’s 0.4708. This supports a substantial contribution from joint protein–glycan matching beyond the single-molecule tendencies captured by the additive branches.

The fixed-budget experiment directly tests cross-glycan reuse (§4.3). Target glycans remain held out as support increases from 0 to 16 distinct motif-containing glycans, keeping total measurements and per-protein observation counts unchanged. All 12 tested motifs show seed-averaged target NMSE reductions. The equal-motif mean reduction is 0.0677 versus 0.0126 using the same motif counts with concatenation. Supervision on other glycans therefore improves unseen-target predictions, directly supporting the cross-glycan reuse of learned recognition information. We hope these experimental results can address your concerns.

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

The paper introduces MotifMatch, which predicts protein–glycan binding by sharing parameters across recurring glycan motifs while using protein-derived factors to modulate those motifs through Hadamard matching. Its key conceptual contribution is showing that transfer between pairs depends on both motif overlap and protein-conditioned response alignment.

On GlycanML, MotifMatch achieves the best mean ranking performance for unseen proteins, unseen glycans, and double-cold pairs. Fixed-budget experiments show that measurements on glycans sharing a motif improve prediction on held-out glycans carrying that motif, while ablations support the specific combination of motif counts and Hadamard matching. It is also substantially faster than the baselines.

Critical strengths
  1. Well-motivated modeling idea with a clear mechanism for transfer. MotifMatch explicitly shares parameters across recurring glycan motifs while conditioning their use on the protein partner. This gives a principled answer to how sparse pair-level supervision can transfer to unseen glycans, and the accompanying overlap–alignment analysis makes the mechanism more interpretable than a purely black-box encoder.
  2. Strong and targeted empirical validation. The paper evaluates protein-cold, glycan-cold, and double-cold generalization, and complements benchmark results with fixed-budget support experiments and representation/scoring ablations. The fact that shared-motif support improves held-out motif-containing glycans across all 12 tested motifs is particularly consequential because it directly tests the paper’s central claim rather than only reporting aggregate benchmark gains.
Critical weaknesses
  1. Practical gains are relatively modest and uneven. MotifMatch’s strongest advantage is in ranking rather than absolute response prediction, and in the double-cold setting its NMSE is essentially at the constant-mean baseline. In the motif-support experiments, gains also vary substantially by motif, with some near zero. This limits how strongly the results establish broad practical superiority.

  2. The representation is intentionally simple but loses important structural information. Motif counts preserve local linkage and terminal context but not motif arrangement, branch placement, or spacing, so structurally distinct glycans can become indistinguishable. This could constrain performance in settings where higher-order topology matters.

Questions for authors
  1. The double-cold setting shows only modest ranking gains and essentially no improvement in NMSE. Can the authors clarify how statistically robust the ranking advantage is across the different split seeds, and whether the improvement is consistent across proteins rather than driven by a subset? This would affect how strongly I interpret the claimed generalization benefit.

  2. In the motif-support experiments, transfer gains vary substantially across motifs, with some near zero. Can the authors provide a clearer analysis of what properties of a motif or its supporting glycans predict successful transfer? This would help determine whether the proposed mechanism is broadly reliable or only effective in favorable cases.

5 comments

Thank you for your careful reading and constructive feedback. We appreciate your encouraging comments on the motivation, the transfer mechanism, and the targeted empirical validation, particularly the shared-motif support gains across all 12 tested motifs. Below we address your questions through expanded paired evaluation (§§4.1–4.2; Appendix C.3, Table 10), an analysis of structural context and transfer (§4.3; Appendix C.4), and a quantification of structural collisions (Appendix C.8). These analyses are now included in the revised manuscript, with changes highlighted in blue for ease of review. We sincerely hope that the additional evidence and explanations below help address your concerns.

Double-cold ranking gains are modest, with little NMSE improvement. Are the gains robust across split seeds and broadly distributed across proteins, rather than driven by a subset? (W1, Q1)

Thank you for this thoughtful comment and for suggesting how we could better establish the reliability of the generalization results. Double-cold requires simultaneous generalization to unseen proteins and glycans, and mean NMSE values across methods remain close to the constant-mean reference (Table 9). MotifMatch nevertheless achieves the highest mean ranking scores, including NDCG@20 of 0.4751 versus 0.4702 for LectinOracle (Table 2). Your question motivated us to examine both the robustness of this modest ranking improvement and its distribution across proteins.

Following your suggestion, we expanded the evaluation to four models × five data splits × three training seeds, totaling 60 model runs. Comparators were fixed from the submitted per-metric baseline results before expansion, and each paired comparison uses identical evaluation observations.

Double-cold metricBaselinePaired ΔSplit-level 95% CICluster 95% CIPositive splits
NDCG@20LectinOracle+0.0147[0.0066, 0.0229][0.0028, 0.0268]4/5
Protein SpearmanBiLSTM+0.0045[−0.0225, 0.0314][−0.0187, 0.0326]3/5
Global SpearmanLectinOracle+0.0097[−0.0129, 0.0324][−0.0111, 0.0309]3/5

Δ = MotifMatch − baseline. Split-level intervals use five paired split means, each averaging three training seeds; cluster intervals use paired resampling of protein and glycan similarity clusters. All intervals are pointwise 95%; details are in Appendix C.3.

Both NDCG@20 intervals are above zero, supporting a modest but robust improvement in top-candidate ranking.

To examine how improvements are distributed across proteins, we analyzed the original three runs, each containing 208 eligible test proteins. NDCG@20 improves over LectinOracle for 55.8%, 51.3%, and 50.5% of proteins; Protein Spearman improves over BiLSTM for 52.7%, 49.5%, and 60.6%.

On the official protein-cold split, MotifMatch’s mean Protein Spearman gain over IUPAC n-gram, the strongest baseline for this metric, is +0.03106, positive in all three training seeds. After averaging seeds within each protein, 110 of 139 test proteins (79.1%) improve. These results show larger, broadly distributed gains on the official benchmark, complemented by a retained NDCG@20 advantage when both partners are unseen.

Transfer gains vary substantially across motifs. Which properties of motifs or supporting glycans predict successful transfer, and is the mechanism broadly reliable or effective only in favorable cases? (W1, Q2)

We all found this question both interesting and important! We therefore explored the patterns of transfer through shared motifs in more detail. Our analysis links larger gains to greater structural-context similarity between supporting and target glycans. Although both groups contain the selected motif, they differ in the other local structures they carry.

Across the 12 tested motifs, we measured the cosine similarity between support and target motif-prevalence vectors, excluding the selected motif itself. Using one three-seed mean per motif, similarity correlates with target and control-adjusted gains at Spearman ρ = 0.455 and 0.490, respectively. The adjusted-gain association remains positive in each seed and when any single motif is excluded.

The near-zero-gain case illustrates this relationship. Galα1–3Gal has the lowest context similarity and smallest mean target gain in the panel; Fucα1–2Gal ranks second-lowest on both. Manα1–3Man provides a comparison with higher similarity and gain.

MotifSupport–target context similarityTarget NMSE gain
Galα1–3Gal0.8446+0.0018
Fucα1–2Gal0.8678+0.0145
Manα1–3Man0.9450+0.0771

Three-seed means. Target gain is the NMSE reduction from k = 0 to 16 distinct supporting glycans; adjusted gain subtracts the motif-negative control gain. Total measurements and per-protein observation counts remain fixed.

At k = 16, MotifMatch achieves positive seed-averaged target gains for all 12 motifs and the lowest mean endpoint target NMSE for every motif among the four representation–scoring combinations.

These results show improvements across the tested panel while linking their magnitude to the structural relationship between supports and targets. When choosing supports containing a shared motif, the target glycans’ broader structural context is therefore a useful consideration. Thank you for your insight, which helped us arrive at this understanding!

Motif counts omit arrangement, branch placement, and spacing, so distinct glycans can share a representation. This may limit performance when higher-order topology matters. (W2)

Your comment struck a chord with us: we had the very same concern when choosing this representation. Sharing local motifs allows measurements on different glycans to train reusable parameters, while offering substantial computational savings. But we also wanted to understand how much structural detail we were giving up in return. We therefore checked which distinct structures become indistinguishable and how much error those collisions introduce.

The results were reassuring. Structurally distinct glycans with identical full encodings form three groups and account for 0.78% of official test observations. These collisions involve the very kinds of differences you mention, including branch placement and the distribution of sialylated termini across branches. As the comparison below shows, local motifs distinguish many structures that composition alone cannot separate, and terminal context resolves additional collisions.

RepresentationGlycan entries involved in structural collisions
Monosaccharide composition546/927 (58.90%)
Local motif counts without terminal context14/927 (1.51%)
Local motif counts with terminal context8/927 (0.86%)

All encodings retain size and branching descriptors.

We also wanted to know whether these few collisions could nevertheless carry a substantial error cost. Within each protein assay row, forcing feature-identical glycans to share a prediction increases the minimum empirical test MSE by 0.0030, normalized by the training-label variance, relative to allowing separate predictions for distinct structures. This gives us greater confidence in the trade-off: on this benchmark, the representation retains substantial structural discrimination while allowing supervision to be shared across glycans, and the direct error cost of exact collisions is small.

Dear Reviewer ZVvG,

Thank you again for your helpful review. We have posted our responses and additional analyses above, and with the discussion period nearing its end, we would greatly appreciate it if you could let us know whether they address your concerns. Please let us know if anything else needs clarification.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.