acceptodds
Under review as a conference paper at ICLR 2027

Refining Olfactory Receptor Representations from Binding Graph

Abstract

Predicting which olfactory receptors a given odorant molecule activates is a central challenge in computational olfaction. Current methods treat each molecule-receptor pair in isolation, representing receptors with frozen embeddings from general-purpose protein language models. The measured response profile of a receptor across many odorants, which directly encodes its functional selectivity, is never explicitly incorporated into the receptor representation. We introduce Olfactory Binding Graph (OlfaGraph), a three-stage procedure that explicitly models this relational evidence. First, experimentally confirmed receptor responses and non-responses are organized into a signed bipartite graph over odorants and receptors. Second, a two-layer graph neural network propagates information across these relations to produce response-aware receptor embeddings that encode the binding evidence observed during training. Third, the adapted embeddings are combined with the original protein and molecular features in a downstream predictor that generalizes to entirely novel odorants. Across three datasets spanning mammalian and insect olfaction, OlfaGraph ranks among the strongest methods in both seen- and cold-molecule settings, using only 1.6 million trainable parameters. Ablations reveal a domain-specific limitation: general-purpose protein embeddings offer little advantage over far simpler descriptors. When general-purpose encoders fall short for a task-relevant protein family, the accumulated experimental receptor-odorant measurements themselves can ground a more effective representation. The code for OlfaGraph is available at https://anonymous.4open.science/r/Olfaction-Binding-5FE7.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
olfaction, protein representation, graph learning

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_iD49OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper deals with the task of predicting odorant–receptor responses (binary receptor activation or continuous firing rate change). The baselines use frozen molecular (like ChemBERTa) and protein language model embeddings (like ESM3), optionally with cross-attention to produce an interaction vector, followed by a XGBoost predictor. The authors propose using a two-layer GNN (GraphSAGE) that feeds into the XGBoost predictor.

Critical strengths
  1. Outperforms comparatively expensive models based on atom-to-residue attention, like LORAX (LoRA fine-tuned ChemBERTa + cross-attention), ProSmith (multimodal transformer, 43.8M parameters), and MolOR
Critical weaknesses
  1. The novelty is incremental at best
  2. The gains over a simple, off-the-shelf gradient-boosted tree (XGBoost) using pooled embeddings are neglectable
Questions for authors
  • You compared two and one message passing layers, but why not more?
1 comment

We thank the reviewer for the recognition of OlfaGraph’s competitiveness against more expensive models. We address the concerns below.

W1. The novelty of our work is twofold. First, we show that simple receptor descriptors and even receptor identities can compete with pretrained protein language model embeddings, motivating closer examination of representation choices in computational olfaction. Second, we demonstrate the value of learning receptor representations from their experimentally observed responses, beyond sequence alone. OlfaGraph achieves the best overall mean rank across datasets and evaluation settings, establishing response-informed representation learning as a promising direction for the field.

W2. OlfaGraph achieves the best overall mean rank in Table 1 (2.35 versus 2.96 for XGBoost-base), ranking first or second in five of six settings. Additional experiments demonstrate statistically significant gains under harder cold-molecule evaluation.

On M2OR, a Hladiš-style HDBSCAN cluster-based split reduces the median maximum train–test Tanimoto similarity from 0.59–0.65 to 0.47–0.60, increasing molecular dissimilarity between training and test sets:

MethodAUROC ↑AUPRC ↑MCC ↑F1 ↑
XGBoost-base0.79080.63390.48880.5474
OlfaGraph (ours)0.8050*0.6543*0.52170.5867

Results are means across five partitions. * denotes a significant improvement over XGBoost-base in paired tests (p<0.05p<0.05).

OlfaGraph improves all four mean metrics, with significant gains in AUROC (p=0.0018p=0.0018) and AUPRC (p=0.0419p=0.0419), demonstrating benefits over pooled embeddings even when test odorants are less similar to training molecules.

Q1. Two layers are the minimum depth needed for receptor–odorant–receptor information exchange; one layer aggregates only directly connected odorant features. To assess whether greater depth helps, we additionally tested three layers on M2OR in the cold-molecule setting:

EncoderAUROC ↑AUPRC ↑MCC ↑F1 ↑
Two layers (default)0.83170.66840.56250.6566
Three layers0.83120.66510.56060.6562

Across five partitions, the third layer improves none of the four mean metrics, with no statistically significant differences (all paired-test p≥0.160p\geq0.160). These results support two layers as a sufficient choice for the proposed information exchange.

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

This paper introduces OlfaGraph, a method for odorant-receptor response prediction by adapting receptor representations using the measured response relationships. Specifically, the experimentally measured receptor responses and non-responses are organized into a signed bipartite graph over odorants and receptors. Afterward, OlfaGraph employs two-layer GNNs (GraphSAGE) to learn the receptor embeddings. These embeddings are further fed into a predictor (XGBoost) for downstream prediction tasks.

Critical strengths
  • A practical, response-aware representation with a clearly disclosed deployment setting.
  • The ablation studies substantially improve the scientific contribution beyond the performance gain.
  • The source code is provided to ensure reproducibility.
Critical weaknesses
  1. What is the rationale for choosing two layers for GNNs? In the ablation study, the authors compare against no message passing and a single message-passing layer, while they do not test whether adding more message-passing layers beyond two improves or degrades performance.
  2. What is the molecule distribution of each employed dataset? The current split is reasonable to some extent, but unseen molecules may still be very similar to the training molecules. Molecule-disjoint does not establish scaffold-disjoint or chemically distant generalization. If possible, I recommend that the authors report results under scaffold- or fingerprint-similarity split settings to strengthen the paper's claims.
  3. The experiments do not yet show a clear and consistent advantage over existing methods. The authors modify the baseline methods by replacing their protein encoders (ProtBERT, ESM-2, etc., with ESM-3) and removing some tuning and ensembling steps, so the results do not represent a comparison against the original complete methods. The claim of performance gains for the proposed methods needs additional experiments for validation.
Questions for authors

See above

2 comments

We sincerely thank the reviewer for the constructive feedback. Below, we address each concern. Our central contribution is to examine what makes receptor representations effective: strong, simple baselines can match complex interaction models, while incorporating measured response relations provides a promising route to better receptor representations. Our additional experiments examine this finding under deeper encoders, harder molecular splits, and tuned prediction pipelines.

W1. Two layers are the minimum needed for receptor–odorant–receptor information exchange: one layer aggregates features from directly connected odorants, while the second incorporates information from receptors sharing those odorants.

Following your suggestion, we evaluated a three-layer encoder on M2OR in the cold-molecule setting, using the full prediction pipeline from Table 1. Results are averaged across five partitions.

MethodAUROC ↑AUPRC ↑MCC ↑F1 ↑
OlfaGraph (2 layers, default)0.83170.66840.56250.6566
OlfaGraph (3 layers)0.83120.66510.56060.6562

Adding a third layer provides no improvement on any metric, with no statistically significant differences between configurations (all paired-test p≥0.160p\geq0.160). These results support two layers as a sufficient choice for the evaluated setting.

W2. We agree that molecule-disjoint splits do not guarantee chemically distant generalization. M2OR contains 596 odorants, whereas Mosquito and Fly each contain 110; we therefore focus our additional analysis on M2OR, unseen molecule setting.

We evaluated all methods using a Hladiš-style HDBSCAN cluster-based molecule split. This groups odorants by fingerprint-based Tanimoto distance and uses the resulting clusters to select test molecules, reducing train–test chemical similarity compared with random molecule-disjoint splitting.

Across five partitions, the median of each test odorant’s maximum Tanimoto similarity to training odorants decreases from 0.59–0.65 to 0.47–0.60. This provides a harder cold-molecule evaluation, although close analogs can still occur across splits.

MethodAUROC ↑AUPRC ↑MCC ↑F1 ↑
LORAX0.76730.58810.42980.4854
ProSmith0.78520.62070.47960.5579
Hladiš0.79980.64320.49070.5501
MolOR0.80670.63610.49960.5639
XGBoost-base0.79080.63390.48880.5474
OlfaGraph (ours)0.8050*0.6543*0.52170.5867

Results are means across five partitions. Bold marks the highest mean; * denotes a significant improvement over XGBoost-base in paired tests (p<0.05p<0.05). We use XGBoost-base as the statistical reference because it is the strongest baseline by overall mean rank in Table 1.

OlfaGraph achieves the highest mean AUPRC, MCC, and F1, with significant improvements over XGBoost-base in AUROC (p=0.0018p=0.0018) and AUPRC (p=0.0419p=0.0419). These results support its competitiveness against existing approaches and its benefit over the strongest aggregate baseline under greater train–test molecular dissimilarity.

W3. We deliberately use the same ESM3 protein representations to isolate differences in interaction modeling from differences in pretrained protein encoders. Likewise, a common fixed XGBoost predictor separates the contribution of learned representations from gains due to tuning and ensembling. These conditions apply equally to OlfaGraph. We will clarify that this is a controlled comparison, rather than a reproduction of each original complete pipeline.

To address the omitted optimization steps, we additionally applied tuning and boosting ensembling following the strategy used in ProSmith and LORAX, with 50 tuning trials per prediction head, to all methods. We focus on M2OR in the seen-molecule setting, the standard setting for most baselines, retaining ESM3 throughout. Results are means across five partitions.

MethodPrediction configurationAUROC ↑AUPRC ↑MCC ↑F1 ↑
OlfaGraph (ours)Single fixed head0.88990.74590.62970.7122
Tuned ensemble0.89820.76080.63450.7154
LORAXSingle fixed head0.89140.75350.62910.7122
Tuned ensemble0.90090.76670.64050.7210
ProSmithSingle fixed head0.87710.74190.61750.7044
Tuned ensemble0.89880.76220.63670.7180
MolORSingle fixed head0.89080.76120.62970.7113
Tuned ensemble0.90120.76610.64170.7218
HladišSingle fixed head0.90210.78090.65450.7319
Tuned ensemble0.90740.78210.65930.7329

Tuning and ensembling improve all methods, including OlfaGraph, whose AUROC increases from 0.8899 to 0.8982. Performance across methods remains positively correlated before and after these additions (Pearson r=0.82r=0.82–0.950.95), with Hladiš remaining the strongest method in this setting, consistent with Table 1.

To conclude, the main claim stated in the paper is consistent competitiveness across evaluation settings, rather than superiority on every benchmark, with controlled experiments designed to isolate the contribution of response-aware receptor representations from differences in protein encoders and downstream optimization.

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

The paper proposes OlfaGraph, an odorant-receptor interaction prediction pipeline that derives receptor representations from measured response profiles rather than from protein sequence embedding alone. Experimental evidence is first organized into a signed bipartite molecule-receptor graph, and a two-layer signed GraphSAGE encoder then produces a response-aware receptor embedding. This embedding is concatenated with the original ESM3 and ChemBERTa embeddings and passed to XGBoost. Evaluated on M2OR, Mosquito and Fly in random split and cold molecule settings, OlfaGraph is reported to achieve the best overall mean rank compared to odorant-receptor baselines and a plain concatenation XGBoost. The authors further report that interaction modelling gives limited gains over simple concatenation and that one-hot receptor identities and simple sequence descriptors are competitive with pre-trained protein language model embeddings.

Critical strengths
  • The core intuition is reasonable and well motivated. A receptor’s measured response profile across many odorants says something about its selectivity that sequence embeddings do not necessarily capture, and in the cold molecule setting, the receptors are fixed, so that information is available.

  • Well-designed and thorough ablations testing the benefit of various receptor representations and message passing designs under the same XGBoost model.

  • The negative finding that interaction baselines barely beat a concatenation + XGBoost baseline is useful to the field in its own right.

Critical weaknesses
  • The performance of nearly all tested ablations is not significantly worse that the full pipeline (Table 2). One possible explanation that is not tested for is that the model actually learns that similar molecules activate the same protein. In that setting, the model only needs a receptor id and a similarity measure between molecules, both of which are already provided with the combination of one-hot + ChemBERTa. This would also explain the strong performance of one-hot encoding on Mosquito and Fly datasets which have only 50 and 24 unique proteins, respectively, and also the strong performance of simple protein features. Note that the value of protein language models usually comes from the ability to measure the similarity of novel proteins to the training-set ones, which enables novel protein generalization. However, in the current setting, expanding OlfaGrpah to unseen proteins is not possible (the novel protein embedding would simply be unchenged and remain the initial ESM-3).

  • Based on the reported confidence intervals, OlfaGraph’s performance gains over the baselines seem often not significant. Even in cold molecule settings, the differences between means are usually on the third decimal place.

  • A dedicated limitations discussion is missing. One important limitation is that the proposed approach cannot be applied to novel proteins which is a desirable property given the large number of orphan olfactory receptors (e.g., in the online version of M2OR, there are approx. 1400 protein sequences and only 53000 experiments, with the majority being inactive).

  • (Minor) The introduction says that in the cold molecule setting, prediction “depends entirely on the receptor representation, since no odorant-specific information is available.” (line 90). This is not accurate since the models can exploit similarity to the training set molecules, which is available via ChemBERTa embedding here. After all, this is why the early drug design models developed for a single target worked even though they could not infer any protein-molecule interaction.

  • (Minor) In the introduction, the authors state that olfactory receptors (ORs) are a family of G-protein coupled receptors (GPCRs) (line 37). In fact, this only holds for mammalian ORs in M2OR database whereas insect receptors in Mosquito and Fly datasets are ligand-gated ion channels.

Questions for authors
  • OlfaGraph is using the response of the receptor to derive a representation for the olfactory receptors. An alternative strategy would be to drop GNN and describe each receptor directly by its receptor’s raw response vector over the chosen set of odorants. This should be possibly embedded to the same dimension as OlfaGraph embedding and fed to XGBoost. Can the authors evaluate such an alternative strategy?

  • As pointed out by Hladiš et al., M2OR contains data with different quality, with the majority of the samples measured in primary screening rounds, which has precision and recall of only 56% and 40%, respectively (Table 3 in Lalis et al., 2024). How is this data quality reflected in the graph construction?

  • The experiment on M2OR in Figure 2 is undermining the main narrative since the full information about the interaction is worse than using 6 molecules. Which 6 molecules were chosen? And can the authors provide more in-depth analysis of what is special about these 6 molecules out of approx. 600 others in M2OR? Related to the previous question, are these molecules tested in screening experiments? There are several molecules in M2OR that are tested and even responsive for many receptors in primary screening but have few experiments done in confirmatory assays, or their response changed significantly (see for example Figure 7 in Hladis et al.). Finally, why would having only a few anchor odorants be beneficial compared to the full graph?

  • To further probe the hypothesis that the model simply learns that "similar molecules activate the same protein" (see weaknesses), I suggest the following: 1. For cold molecule experiments, report the results stratified by each test molecule’s maximum Tanimoto similarity to the training odorants. This would reveal whether molecular similar is indeed a driving force for the performance. 2. On top of this, odorant molecules are relatively small volatiles, and experiments usually test molecules that are close analogs. Therefore, a random cold molecule split can put one analog in the training set and the other in the test set. To assess the performance in a real-world generalization task, please report performance under Butina-style clustering split (an analogy to Cluster-molecule evaluation setting of Hladis et al.). Since Mosquito and Fly datasets do not have enough molecules to create a meaningful clustering, evaluate the models at least on M2OR. Note that the negative results would not rule out the hypothesis, but the positive results would strengthen it.

  • Can the authors provide statistical tests to support the main claims? Considering the reported confidence intervals, the majority of values in Tables 1 and 2 seem to not be significantly different.

4 comments

We sincerely thank the reviewer for the thoughtful assessment, constructive suggestions, and time and effort devoted to our work. Below, we address the questions and identified weaknesses.

Before doing so, we would like to highlight the main motivation and value of our work. Computational olfaction is a developing field, and our results show that existing approaches warrant closer examination: simpler baselines can match or outperform more complex methods. We demonstrate this finding and propose a direction for advancing the field by learning receptor representations from measured odorant–receptor responses. This approach achieves consistently competitive results across datasets and evaluation settings.

Q1. We evaluated a raw response-profile (RP) baseline on M2OR in the cold-molecule setting, using the same experimental setup as Table 1. Each receptor’s RPRP is a vector over the 508 training odorants: 0 denotes a measured non-response, 1 a measured response, and 2 an unmeasured pair. We replace OlfaGraph’s learned receptor embedding with RPRP, retaining the original ESM3 and ChemBERTa embeddings: [xprot∥xmol∥RP][x_{\mathrm{prot}}\|x_{\mathrm{mol}}\|RP]. We report means across five partitions:

MethodTotal feature dimensionAUROC ↑AUPRC ↑MCC ↑F1 ↑
XGBoost-base19200.81960.66090.52740.6182
OlfaGraph (ours)21760.83170.66840.56250.6566
RP + original embeddings24280.83830.66720.55530.6506

RP significantly improves AUROC, MCC, and F1 over XGBoost-base (p=0.0211p=0.0211, 0.00050.0005, and 0.00780.0078, respectively; two-sided paired tt-tests), confirming that raw response profiles provide a strong baseline. Meanwhile, RP does not surpass OlfaGraph in mean AUPRC, MCC, or F1, despite using a larger receptor representation (508 versus 256 dimensions).

Q2. Measurement quality is considered during split preparation: the five cross-validation partitions inherited from LORAX reserve higher-quality dose-response measurements for testing, while training combines primary screening (74%), secondary screening (18%), and dose-response measurements (8%) in similar proportions across partitions. This combination provides the graph with broader receptor–odorant coverage from screening alongside more reliable dose-response evidence. Thus, the graph edges reflect the available mixed-quality measurements.

We additionally tested Hladiš-style weighting of the link-prediction loss to account for differences in assay reliability. Results are averaged across five partitions.

SettingGraph lossAUROC ↑AUPRC ↑MCC ↑F1 ↑
Seen moleculesDefault0.88870.74620.62910.7119
Weighted0.88950.74500.62700.7101
Cold moleculesDefault0.83170.66840.56250.6566
Weighted0.83220.66910.55900.6541

To conclude, weighting yields no statistically significant improvement in either setting (all paired-test p≥0.420p\geq0.420).

Q3. As described in Appendix C, we select anchor odorants greedily to maximize receptor-pair coverage: each selected odorant connects as many previously uncovered receptor pairs as possible, regardless of whether their measured responses are positive or negative. These odorants serve as bridges for receptor–odorant–receptor message passing. Appendix C also evaluates the full graph and progressively smaller graphs: at q=0q=0, approximately 580 training odorants participate in message passing; at q=0.99q=0.99, approximately six remain. The smaller graph performs better in both seen- and cold-molecule settings, empirically supporting the six-anchor configuration. Importantly, all training pairs still contribute to the loss; only message passing is restricted. These results suggest that a compact set of shared bridges supports receptor representation learning more effectively than including every measured edge. Selection is based on coverage, rather than assay reliability.

To clarify what distinguishes the selected anchors, we examined the six chosen in the first partition. In the full benchmark pool, each was tested against 435–694 receptors, with only 0.17–11.5% positive responses: they are broadly tested odorants, not universal activators. Their measurements predominantly come from primary screening, with only 1–88 dose-response measurements per odorant. Their value for graph construction therefore lies in extensive receptor coverage, including measured non-responses, rather than high assay reliability. Pairwise Tanimoto similarities are low (≤0.27), indicating that the selected anchors span diverse chemical structures.

Q4.

  1. We stratified M2OR cold-molecule predictions by each test odorant’s maximum Tanimoto similarity to training odorants in the same partition, using ECFP4 fingerprints. Higher Tanimoto values indicate greater similarity to the closest training odorant.
Maximum TanimotoMethodAUROC ↑AUPRC ↑MCC ↑F1 ↑
≥0.8\geq 0.8XGBoost-base0.86200.65340.53680.6272
OlfaGraph (ours)0.86460.65280.5892*0.6780*
[0.6,0.8)[0.6,0.8)XGBoost-base0.84030.73200.59050.6440
OlfaGraph (ours)0.85880.7482*0.6156*0.6662*
<0.6<0.6XGBoost-base0.77510.70340.49170.5962
OlfaGraph (ours)0.79240.7218*0.50940.6199

Bold * denotes a statistically significant improvement over XGBoost-base (p<0.05p<0.05, paired tests across five partitions).

OlfaGraph achieves statistically significant improvements across all three similarity bins: MCC and F1 at high similarity, AUPRC, MCC, and F1 at intermediate similarity, and AUPRC at low similarity (p=0.041p=0.041). Its gains are therefore not confined to close analogs, although both methods’ lower AUROC on less similar odorants indicates that molecular similarity remains an important factor.

  1. We evaluated all methods on M2OR using a Hladiš-style HDBSCAN cluster-based molecule split, with clustering based on Tanimoto distances. This produces a harder cold-molecule evaluation: the median of each test odorant’s maximum Tanimoto similarity to training odorants decreases from 0.59–0.65 with random splitting to 0.47–0.60 with cluster-based splitting. Results are averaged across five partitions.
MethodAUROC ↑AUPRC ↑MCC ↑F1 ↑
LORAX0.76730.58810.42980.4854
ProSmith0.78520.62070.47960.5579
Hladiš0.79980.64320.49070.5501
MolOR0.80670.63610.49960.5639
XGBoost-base0.79080.63390.48880.5474
OlfaGraph (ours)0.8050*0.6543*0.52170.5867

* denotes a significant improvement over XGBoost-base in paired tests across five partitions (p<0.05p<0.05). We use XGBoost-base as the reference because it is the strongest baseline by overall mean rank in the paper's Table 1.

OlfaGraph achieves the highest mean AUPRC, MCC, and F1, with significant improvements over XGBoost-base in AUROC (p=0.0018p=0.0018) and AUPRC (p=0.0419p=0.0419). These results support the benefit of response-aware receptor representations under a harder cluster-based evaluation with greater typical train–test molecular dissimilarity.

Q5. In paper's Table 1 and 2, we focus on ranks to summarize performance across metrics with different scales, datasets, and evaluation settings. OlfaGraph achieves the best overall mean rank (2.35 versus 2.96 for XGBoost-base in Table 1), ranking first or second in five of six dataset–setting combinations. We agree that this evidence supports only consistent competitiveness rather than statistically significant superiority. However, the additional experiment reported in our response to Q4 demonstrates statistically significant improvements in AUROC and AUPRC over XGBoost-base under the harder cluster-based cold-molecule evaluation.

W1, W3. Receptor identity alone does not always provide an equally effective representation. In Table 2, randomly initialized OlfaGraph preserves receptor identity without sequence information, yet outperforms direct one-hot encoding on M2OR (AUROC: 0.794→0.840 for seen molecules) and seen-molecule Mosquito (R2R^2: 0.544→0.667). Since both representations identify receptors, these gains support OlfaGraph’s underlying principle: organizing receptor information through measured response relations improves its utility for prediction beyond directly supplying receptor identity.

We focus on known receptors to investigate whether their existing response measurements enable better receptor representations for predicting responses to unseen odorants. We agree that generalization to unseen receptors is important, particularly for receptor deorphanization, but it falls outside the scope of this study. We will add a dedicated limitations section stating: “Our evaluation considers unseen odorants paired with receptors observed during training; generalization to unseen receptors remains untested. For receptors without experimental response measurements, OlfaGraph cannot refine their representations through response-based message passing, limiting its applicability to receptors lacking such data.”

W2. See Q5.

W4, W5. We thank the reviewer for pointing out these inaccuracies. In the revised paper, we will clarify that cold odorants lack training response measurements, but their molecular embeddings still provide information for prediction. We will also distinguish mammalian ORs, which are GPCRs, from insect ORs, which are ligand-gated ion channels.

I would like to thank the authors for their commitment and for running additional experiments.

Regarding W3, W4 and W5, I appreciate the correction and the acknowledgement of the limitations. I also appreciate additional experiment and clarification for Q2. I consider all these points to be resolved to my satisfaction.

W1, Q4: As for receptor identity hypothesis, it is expected that using more complex features than one-hot encoding improves the performance. After all, these features still carry receptor identity, but also enables to measure receptor similarity. However, the results in Table 2 show that in 2 out of 6 scenarios, one-hot encoding outperforms OlfaGraph (ESM3 and random init.). Although further experiments are probably needed to fully test the hypothesis, the results in the answer to Q4 suggest that indeed molecular similarity is an important factor driving the performance. What still remains unresolved even after rebuttal is why arguably the simplest possible features match the performance of the embedding derived from two complex models (PLM + OlfaGraph).

Q5: As for the statistical testing, I appreciate the authors honesty that the current tables support only consistent competitiveness rather than statistically significant superiority. On the other hand these results are undermining the novelty of the paper which is an important evaluation criterion for acceptance in a venue like ICLR.

Q3: A particularly important point is in the last question in Q3, which is why such a small set of molecules is beneficial and this question is not answered in the rebuttal. I understand that empirically this set worked better, even though I must note that the differences are small and possibly insignificant. However, if empirical performance is the only criterion then the randomly selected control in Figure 2 achieves the best performance. Thus, following your response, why receptor coverage should be a criterion for choing a molecule? Coverage does not play any role in the biochemical interaction between proteins and molecules and it is more reflecting selection bias of the experiments. Therefore, the selected 6 odorants can simply exploit this selection bias and the model is not learning a generalisable information about the protein-molecule interaction. Note that the fact that the majority of the responses of these molecules is negative, indicates that the response information on the edges in the graph is not crucial for the performance.

Q1: The most informative result from the rebuttal is the Table in the answer to Q1. According to the results, RP and OlfaGraph are hardly distinguishable (author did not state the significance tests for OlfaGraph vs RP, but the differences are up to 0.0072 across metrics where the confidence intervals of OlfaGrpah are around 0.05 in Table 1 cold molecule M2OR). Therefore, a parameter-free method reached the same performance as the proposed OlfaGraph, which undermines one of the main contribution of the paper. Note also that XGBoost has native missing-value handling that would handle unmeasured pairs, so the chosen encoding (0 = non-response, 1 = response, 2 = unmeasured) even plays against RP. Despite this, it is indistinguishable from OlfaGraph.

Overall, I appreciate the authors’ thorough engagement during the rebuttal. The paper’s negative result, that off-the-shelf embeddings with a standard predictor are competitive with dedicated interaction architectures, is a genuine contribution, and the RP experiment added during the rebuttal reinforces that accumulated experimental measurements are an underused source of receptor information. My remaining concern is that OlfaGraph itself, the paper’s stated contribution, has not been shown to improve significantly over either the concatenation baseline or the parameter-free response profile. I would encourage the authors to frame the paper around the conclusion their results support, but I do not think the present evidence meets the bar for acceptance and I maintain my rating.

Thank you very much for reading our paper so carefully. Your comments and suggestions helped us rethink several points and improve the paper. We sincerely appreciate the time and effort you put into this review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.