acceptodds
Under review as a conference paper at ICLR 2027

KiMChEE: Reaction Context from Co-Substrates and Environment Improves Kinetic Prediction for Uncharacterized Enzymes

Abstract

A growing number of machine learning models predict enzyme kinetic parameters such as turnover number () and Michaelis constant (), along with their ratio (), from sequence and substrate information for enzyme engineering and metabolic studies. However, their accuracy degrades sharply on enzyme sequences that have never been characterized. We hypothesize that one reason for this failure lies in the input representation. Most existing methods describe a reaction by an enzyme sequence and a main substrate, discarding the remaining reactants and the assay conditions that together determine the kinetic parameter values. Consequently, records sharing an enzyme and a main substrate but differing in reaction context (e.g., different co-substrates and reaction conditions) map to identical inputs with conflicting labels. We therefore build a multimodal kinetic predictor, KiMChEE (Kinetic parameter prediction Model for Chemicals, Enzymes, and Environment). KiMChEE augments the conventional enzyme-and-substrate input with the reaction's other reactants, the co-substrates, and with the assay pH and temperature. It also pools residue-level embeddings with higher weight on substituted positions, keeping variants (e.g., engineered enzymes) separable from their wild types. An ablation isolates the contribution of the co-substrate input, showing that it improves predictions on records whose inputs would otherwise be indistinguishable despite conflicting labels while leaving the remaining records essentially unchanged. We compare methods on an out-of-distribution benchmark constructed from records absent from the public repositories (BRENDA and SABIO-RK) used to train the compared models, so that none of them can have seen these records during training. KiMChEE attains the highest on all three kinetic parameters under this benchmark, with a margin of 0.128 on over the second-best model. These results support our hypothesis that the reaction context, rather than an enzyme and a single substrate alone, belongs in the input for kinetic prediction on novel enzymes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Enzyme Kinetics, Kinetic Parameter Prediction, Multimodal Representation Learning, Protein Language Models, Out-of-Distribution Generalization, Generalization to Unseen Proteins, AI for Science

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_26g7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

KiMChEE predicts enzyme kinetic parameters using enzyme sequences, main substrates, co-substrates, and assay conditions. It combines adapted pretrained encoders, mutation-aware pooling, and ExtraTrees regression, and evaluates performance on literature-curated external measurements.

Critical strengths
  1. Separating the main substrate from co-substrates addresses a clear limitation of conventional inputs.
  2. The collision-focused analysis helps explain where additional reaction context contributes.
  3. The manually curated external benchmark is valuable, and the paper openly reports remaining prediction failures.
Critical weaknesses
  1. Sequence overlap and random-split ablations weaken the evidence for novel-enzyme generalization.
  2. Mutation-aware pooling lacks direct validation, and mutation-direction prediction remains weak.
  3. Uncertainty estimation and baseline calibration need further checks.
Questions for authors
  1. The core co-substrate ablation uses a random record-level split, whereas the main claim concerns generalization to new enzymes. Please repeat this ablation with homology-disjoint folds and on the external benchmark, keeping encoder adaptation and ensemble settings matched.
  2. Mutation-aware pooling lacks a direct ablation, and Table 5 shows no above-chance mutation-direction prediction for kcatk_{\mathrm{cat}} or KmK_m. Please compare Eq. (5) with uniform pooling and examine sensitivity to the mutation weight α\alpha.
  3. Record-level bootstrapping may underestimate uncertainty when measurements share an enzyme or source article. Please add enzyme- or article-clustered confidence intervals, particularly for the smaller improvements on KmK_m and kcat/Kmk_{\mathrm{cat}}/K_m.
  4. DLTKcat’s PCC of 0.480 alongside an R2R^2 of −2.234 raises a calibration question. Please verify checkpoint-specific units and input conventions, and provide the baseline inference code so this discrepancy can be checked.
  5. Appendix D shows that 3.8–8.0% of external records contain sequences already present in training. Please report performance by sequence-identity range and on a strictly sequence-disjoint subset to support the claims about uncharacterized enzymes.
1 comment
AuthorsOpenReview ↗

Response to Reviewer 26g7

We thank the reviewer for the helpful suggestions on controlled ablations, uncertainty estimation, and baseline evaluation.

W1 & Q1 & Q5. Co-substrate ablation and sequence overlap

Item 2 of our common official comment repeats the co-substrate ablation under identity-clustered folds with matched encoder adaptation and ensemble settings. The channel contributes significantly, most on collided records, and the external benchmark does not distinguish this effect from zero (Table 3, Appendix D). Table F1 reports performance by maximum sequence identity for all three parameters; below 30% identity, KiMChEE ranks first for KmK_m and kcat/Kmk_{cat}/K_m and ties UniKP for kcatk_{cat}. Excluding the ≥99%\geq 99\% identity bin, which removes every record with an identical or near-identical training sequence, the kcatk_{cat} margin over UniKP remains significant (ΔR2=0.138\Delta R^2 = 0.138 [0.083, 0.202]), while the KmK_m and kcat/Kmk_{cat}/K_m margins remain non-significant, as on the full set (Table F2).

W2 & Q2. Uniform-pooling control and sensitivity to α\alpha

For the uniform-pooling control described in Item 1 of our common official comment, the table below gives the changes at α=0\alpha = 0 relative to α=10\alpha = 10 (Table 5).

ParameterOverall ΔR2\Delta R^2 [95% CI]Mutant ΔR2\Delta R^2 [95% CI]
kcatk_{cat}−0.015-0.015 [-0.055, +0.029]−0.019-0.019 [-0.087, +0.068]
KmK_m−0.001-0.001 [-0.015, +0.013]+0.002+0.002 [-0.026, +0.029]
kcat/Kmk_{cat}/K_m−0.018-0.018 [-0.054, +0.008]−0.003-0.003 [-0.061, +0.037]

No interval in the table excludes zero for any of the three parameters, so we do not claim a contribution from up-weighting substituted positions. Item 1 of our common official comment explains how this comparison addresses sensitivity to α\alpha and compares sign agreement on matched wild-type and mutant pairs between the two settings.

W3 (uncertainty) & Q3. Enzyme-clustered uncertainty

The co-substrate effect on collided records described in Item 2 of our common official comment remains supported after accounting for shared enzymes (Table 3). For the comparison with baselines, the KmK_m and kcat/Kmk_{cat}/K_m margins over the second-best model are not significant even with records resampled (Table F2), so we do not claim them under either resampling unit; with enzymes resampled, the kcatk_{cat} margin over UniKP is 0.1280.128 [0.007, 0.247], which still excludes zero.

ComparisonΔR2\Delta R^2 [95% CI]Resampling unitManuscript table
Co-substrate removal, collided records−0.035-0.035 [-0.055, -0.016]Enzyme3
Co-substrate removal, external set−0.002-0.002 [-0.016, +0.011]Enzyme3

W3 (baseline calibration) & Q4. Corrected DLTKcat evaluation

DLTKcat's negative R2R^2 came from a temperature-scaling mismatch in our evaluation, and correcting it raises its kcatk_{cat} R2R^2 on the external benchmark from −2.234-2.234 to 0.0250.025 (table below). The checkpoint expects temperature and inverse temperature to be min-max scaled with training-data statistics, whereas our original evaluation applied z-score normalization with fixed constants, so the same assay temperature reached the checkpoint on a different scale from its training inputs. We applied the training-data min-max transformation to both channels and reran inference on the full external benchmark. Output units matched the checkpoint, which predicts log⁡10kcat\log_{10} k_{cat}.

EvaluationR2R^2PCCSCCRMSE (log⁡10\log_{10} units)
Submitted−2.234-2.2340.4800.4692.519
Corrected0.0250.0250.4080.3811.383

The high PCC alongside a strongly negative R2R^2 in the submitted evaluation is consistent with this mismatch, which distorted the scale of the predictions more than their ordering. DLTKcat moves from eighth to fifth in Table 4. The corrected wild-type and mutant results are reported in Table 6, and the Discussion now states that three of seven released predictors fall below zero R2R^2, where the submitted version stated four of seven. The DLTKcat inference code is in the repository described in Item 4 of our common official comment.

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

The paper argues that existing kinetic-prediction models ignore assay context variables, including co-substrates, pH, and temperature, although these factors substantially affect the measured kcatk_{cat}, KmK_m, and kcat/Kmk_{cat}/K_m. For the same enzyme and reactant, catalytic rates can differ considerably under different reaction conditions. The proposed method, KiMChEE, therefore jointly models the enzyme sequence, reactant, co-substrates, pH, and temperature. It combines pretrained protein and chemical encoders, mutation-weighted residue pooling, supervised encoder adaptation, and an extremely randomized trees read-out. Separate models predict log-transformed turnover number, Michaelis constant, and catalytic efficiency.

The paper provides a targeted input-channel ablation and a manually curated external literature benchmark excluding source articles already cited by BRENDA or SABIO-RK. KiMChEE achieves the highest external R-squared for all three parameters among the evaluated released predictors, although absolute performance and mutation-effect direction remain limited.

Critical strengths
  1. A meaningful problem formulation that jointly models the enzyme sequence, reactant, co-substrates, pH, and temperature. Kinetic parameters measured for the same enzyme and main substrate can differ substantially across co-substrates and assay conditions. Methods that ignore these factors map such records to the same input despite different targets, which can make their predictions unstable. Incorporating reaction context into the input is therefore a well-motivated design choice. Appendix Table A1 shows that the issue is not marginal: 12.8% of kcatk_{cat} records collide under the conventional enzyme–main-substrate representation.

  2. A targeted ablation that confirms the contribution of the co-substrate channel. The ablation distinguishes records that collide under the enzyme–main-substrate representation from those that do not. Removing the co-substrate channel reduces R-squared by 0.122 on 345 collided held-out records, but only 0.001 on the remaining 2,194. This concentration of the benefit supports the proposed input-collision explanation more directly than an aggregate performance improvement alone would.

Critical weaknesses
  1. The experiments do not show that reaction context improves generalization to uncharacterized enzymes. The co-substrate ablation is run on a single random 90/10 split of the kcatk_{cat} dataset, where removing co-substrates lowers R-squared only from 0.739 to 0.723. Even the larger effect on collided records is measured under this random split. This shows that extra context helps distinguish repeated enzyme–substrate inputs, but not that the benefit carries over to unfamiliar enzymes.

    The external benchmark does not settle this either. It compares the full system with baselines that differ in training data, representations, and fitting procedures, and KiMChEE itself also changes encoder adaptation, mutation pooling, and the read-out; Appendix C even identifies the tree read-out as a dominant factor. Without context-channel ablations on the external or low-homology set, the external advantage cannot be attributed to the input representation, which is the paper's main claim. Mutation-weighted pooling likewise lacks a uniform-pooling control.

  2. The external benchmark contains unseen measurements, not necessarily unseen enzymes, and the application claims are under-supported. Appendix D reports that 7.6%, 8.0%, and 3.8% of external records for the three targets have identical sequences in training, and 80/485, 98/502, and 41/470 records have at least 99% identity. The authors disclose this, and it does not invalidate the benchmark, but excluding source articles is not the same as testing on uncharacterized enzymes. The term "out-of-distribution" should therefore be used more carefully, and results should be reported on lower-identity subsets.

    Practical utility also remains limited. External R-squared is 0.242 for kcatk_{cat} and 0.052 for kcat/Kmk_{cat}/K_m. On matched wild-type–mutant pairs, sign agreement is 0.488 for kcatk_{cat} and 0.433 for KmK_m, no better than chance. Although the Discussion acknowledges this, the suggested use in library ranking or lab-in-the-loop design is not supported by the current results.

Questions for authors
  1. Does the context contribution hold on external and low-homology data? Can the authors report co-substrate and environment ablations on the external set, stratified by maximum sequence identity? Does the full model keep its advantage after excluding identical or near-identical training sequences?
  2. What does mutation-weighted pooling contribute on its own? How does the pipeline perform with uniform pooling, especially on wild-type–mutant differences? How are substituted positions supplied at inference?
  3. What does "identify the assayed reaction" cover? The input includes co-substrate identities but not concentrations, stoichiometry, or other assay components. How many collisions remain among records that share the encoded inputs but differ in these conditions?
2 comments
AuthorsOpenReview ↗

Response to Reviewer TYaY

We thank the reviewer for the detailed feedback on the contribution of reaction context and the scope of our application claims.

W1 & Q1. Co-substrate contribution under identity-clustered folds and on the external benchmark

The co-substrate ablation summarized in Item 2 of our common official comment shows a statistically significant contribution under identity-clustered folds, with a larger effect on collided records (Table 3).

The external benchmark results are summarized below (Table D2).

EvaluationΔR2\Delta R^2 without co-substrates [95% CI]
External benchmark, all records−0.002-0.002 [-0.016, +0.011]
Below 30% maximum sequence identity−0.028-0.028 [-0.075, +0.008]

Both intervals include zero and the cross-validation effect of −0.011-0.011, so, as stated in Item 2, we attribute KiMChEE's external lead to the model as a whole rather than to the co-substrate channel alone. After excluding records with ≥99%\geq 99\% training-sequence identity, KiMChEE retains a significant kcatk_{cat} advantage over UniKP, and its KmK_m and kcat/Kmk_{cat}/K_m margins remain non-significant, as on the full set (Table F2). The environment channel was ablated on the random split, where removing it lowers R2R^2 on both collided and other records (Tables 1 and 2); the identity-clustered and external ablations above were run for the co-substrate channel, which carries the paper's main claim.

W2. Benchmark scope and application claims

Following the reviewer's suggestion, the revision uses "external benchmark" and reports sequence overlap and identity-stratified performance separately (Appendices E and F).

We agree that the Discussion's suggestions of library selection and lab-in-the-loop design were not supported by direct application results, and the revision removes them from the Discussion. Nevertheless, the ablation summarized in Item 2 of our common official comment shows that explicitly representing co-substrate identities improves prediction under identity-clustered folds, particularly for reactions that share an enzyme and main substrate (Table 3). Future work will test whether these gains improve candidate selection and whether retraining improves outcomes in later experimental rounds.

W1 (pooling) & Q2. Uniform-pooling control and substituted positions at inference

Under the uniform-pooling control in Item 1 of our common official comment, R2R^2 does not change significantly for kcatk_{cat} or KmK_m in any subset or for kcat/Kmk_{cat}/K_m overall or on mutant records (Table 5), so we do not claim a contribution from up-weighting substituted positions; Item 1 also compares sign agreement on matched wild-type and mutant pairs between the two settings.

Substituted positions are supplied with the record. Every measurement in the training corpus and in the external benchmark carries the mutation annotation of its source, and the substituted residue positions are parsed from that annotation and mapped onto the tokenized sequence, so the model reads these positions rather than inferring them from the sequence. A record with no annotated position is pooled with uniform weights, which is the α=0\alpha = 0 case above. For a designed variant, the positions are those of the substitutions introduced, so they are known at the time of prediction.

Q3. Assayed-reaction scope and other assay factors

Section 3.2 changes "the input identifies the assayed reaction" to "the input identifies the reactants of the assayed reaction", clarifying that the inputs do not specify a complete assay protocol. Stoichiometry is not encoded; co-substrates enter as an unordered set whose embeddings are averaged (Eq. 3). The percentages use 96,632 available BRENDA and SABIO-RK records after removing duplicate measurements.

Difference within (E,smain)(E, s_{main}) groupsReported share
Co-substrate identity10.18%
Native/recombinant annotation4.07%
Buffer identity/composition/presence3.54%
Co-substrate concentration2.76%
Isoform/isozyme annotation0.77%
Metal identity0.56%
Any of the five external factors8.72%

Co-substrate identities can be encoded directly as a set of molecular structures. Concentrations additionally require observations covering continuous ranges and their combinations, placing greater demands on condition-annotated data. We therefore model reactant identities, pH, and temperature, and excluded records with identified conflicts involving factors outside this scope during preprocessing of both the training and external datasets. In the final training set, no conflicting kinetic values remained among records with identical model inputs, including the mutation-position annotations.

Thank you for the response. I appreciate the motivation behind this work. Although adding co-substrates and environmental factors yields modest gains, enzyme kinetics prediction is inherently challenging, and large, robust improvements on existing public datasets may be difficult to achieve. That said, there remains substantial room for improvement in the technical novelty, representation design, and figures, particularly the clarity of the figures. I therefore lean toward maintaining my original score.

Reviewer_2AWUOpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
1(−1)
Presentation
1(−1)
Soundness
2
Read review
Paper summary

This paper introduces KiMChEE, a multimodal model for predicting enzyme kinetic parameters, including kcat, Km, and kcat/Km. The main motivation is that conventional kinetic prediction models often represent a reaction using only the enzyme sequence and its main substrate. This can cause different reactions to share identical model inputs even when they have different co-substrates and kinetic measurements. KiMChEE addresses this problem by introducing four input channels: enzyme sequence, main substrate, co-substrates, and assay environment. It also incorporates mutation-aware protein embedding pooling to distinguish engineered variants from their wild-type sequences. The authors evaluate the model on BRENDA and SABIO-RK training data and construct an external benchmark from primary literature. KiMChEE achieves the highest R2 among seven evaluated predictors for all three kinetic parameters. The co-substrate ablation further shows that this channel is particularly useful for records that would otherwise have identical inputs.

Critical strengths
  1. A clear and biologically meaningful problem formulation The paper identifies an important limitation in conventional enzyme kinetic prediction: a single-substrate representation does not necessarily describe the complete reaction context. The separation of main substrates and co-substrates is a meaningful design choice. It preserves the distinction between the substrate being varied in an assay and the other reactants involved in the reaction. Figure 1 provides a particularly clear illustration of this input-collision problem.
  2. Convincing ablation and external evaluation The co-substrate ablation is one of the strongest parts of the paper. Removing this channel decreases R2 by 0.122 on collided records but only 0.001 on non-collided records. This provides direct evidence that the additional information is useful in the situations the authors intended to address. The independently curated external benchmark is also valuable because it avoids relying entirely on the individual train-test partitions used by previously published predictors.
Critical weaknesses
  1. Record-level independence does not establish enzyme-level generalization The external benchmark contains measurements absent from BRENDA and SABIO-RK, but this does not mean that all test enzymes are sequence-disjoint from the training data. The authors' own sequence-similarity analysis shows that 80 of 485 kcat records, 98 of 502 Km records, and 41 of 470 kcat/Km records have at least 99% maximum sequence identity to training sequences. Therefore, the benchmark supports generalization to previously unseen kinetic records, but does not fully establish generalization to genuinely novel enzyme sequences. This distinction is particularly important given the paper's emphasis on uncharacterized enzymes.
  2. Absolute predictive accuracy and reproducibility remain concerns Although KiMChEE ranks first among the evaluated models, its absolute performance is still limited. The external benchmark reports R2 values of 0.242 for kcat, 0.443 for Km, and only 0.052 for kcat/Km. Furthermore, mutation-direction prediction remains weak. KiMChEE achieves only 0.488 agreement for kcat, which is not better than chance. The configuration search also evaluates 32 combinations using a single random split. In addition, the training-data assembly code, adapted encoder weights, and baseline execution code are not included in the stated reproducibility package.
Questions for authors
  1. Can the authors report performance separately for enzymes with less than 30% sequence identity to the training set? Such an evaluation would help distinguish generalization to new measurements from generalization to genuinely uncharacterized enzymes.
  2. Can the authors release the complete training-data construction pipeline, adapted encoder weights, and baseline evaluation scripts? Multiple-seed configuration selection would also help establish the robustness of the reported architecture.
1 comment
AuthorsOpenReview ↗

Response to Reviewer 2AWU

We thank the reviewer for the constructive feedback on generalization, predictive accuracy, and reproducibility.

W1 & Q1. Generalization below 30% sequence identity

Below 30% maximum sequence identity, KiMChEE ranks first for KmK_m and kcat/Kmk_\mathrm{cat}/K_m and ties UniKP at the reported precision for kcatk_\mathrm{cat} (Table F1).

ParameterExternal recordsKiMChEE R2R^2Highest baseline R2R^2 in this bin
kcatk_\mathrm{cat}159−0.095-0.095UniKP: −0.095-0.095
KmK_m800.5320.532CatPred: 0.3600.360
kcat/Kmk_\mathrm{cat}/K_m164−0.048-0.048UniKP: −0.186-0.186

Performance is parameter-dependent: KmK_m retains positive predictive performance, while kcatk_\mathrm{cat} and kcat/Kmk_\mathrm{cat}/K_m remain challenging at low identity. The revision reports these subsets separately from the overall external benchmark results (Appendix F). We agree that unseen records are not the same as novel sequences, and the revision now calls this set the external benchmark rather than out-of-distribution; after excluding the ≥99%\geq 99\% identity bin, the kcatk_\mathrm{cat} margin over UniKP remains significant (ΔR2=0.138\Delta R^2 = 0.138 [0.083, 0.202]), while the KmK_m and kcat/Kmk_\mathrm{cat}/K_m margins remain non-significant, as on the full set (Table F2).

W2 & Q2. Absolute accuracy, reproducibility, and configuration selection

Absolute accuracy on the external benchmark remains limited (Table 4), so individual zero-shot predictions still need experimental validation. Only kcat/Kmk_\mathrm{cat}/K_m exceeds chance in mutation-direction prediction (Table 7). Nevertheless, the significant co-substrate benefit described in Item 2 of our common official comment shows that adding reactant identities omitted by conventional inputs improves kinetic prediction under identity-clustered folds (Table 3).

As described in Item 4 of our common official comment, the adapted encoder weights and the baseline inference code are now released with the training data, training code, and trained models in our Hugging Face repository; only the training-data assembly code is withheld under a third-party agreement, and the assembled data is provided in full.

The configuration search ranks configurations on one partition and one seed, as Appendix C now states; the selected configuration is then trained over the five identity-clustered folds, and Tables 3 to 7 report that model. Within the search, the two choices the configuration rests on hold without exception. Every tree-ensemble configuration outranked every neural-head configuration (R2R^2 of 0.665 to 0.712 against 0.463 to 0.619), and every ProtT5 pairing ranked above every ESM-2 pairing among the tree ensembles (Figure C1).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Official Comment to All Reviewers: Summary of Revisions

We thank the reviewers for their careful reading. We have examined the reviews closely, incorporated them, and run additional experiments, which led to revisions in parts of the paper. This common official comment summarizes the main additions and changes to the main text relative to the previous version. Every weakness point and question in each review is answered in a separate individual comment.

1. Uniform pooling control (new Section 4.4, Table 5)

With the pipeline retrained at α=0\alpha = 0 and every other setting unchanged, R2R^2 is lower for kcatk_{cat} and kcat/Kmk_{cat}/K_m in all three subsets and within 0.003 in all three KmK_m subsets, but only the kcat/Kmk_{cat}/K_m wild-type difference is significant (Table 5), so we do not claim that up-weighting substituted positions improves prediction. Wild-type records are pooled identically in both settings, so the wild-type difference comes from training on differently pooled mutant records. KiMChEE's kcatk_{cat} lead on mutant records does not depend on up-weighting; at α=0\alpha = 0, its kcatk_{cat} R2R^2 on mutant records is 0.343, still above the 0.263 of CataPro, the next-best model (Table 6). On matched wild-type and mutant pairs, sign agreement at α=0\alpha = 0 is 0.465 for kcatk_{cat}, 0.535 for KmK_m, and 0.549 for kcat/Kmk_{cat}/K_m, compared with 0.488, 0.433, and 0.634 at α=10\alpha = 10 (Table 7), so the two settings do not differ consistently in how often the predicted direction of a mutation effect is correct. On sensitivity to α\alpha, we compared the untuned α=10\alpha = 10 used in all experiments (Section 3.3) with α=0\alpha = 0, which directly tests whether the extra weight helps, rather than selecting α\alpha on evaluation data.

2. Co-substrate ablation under homology-disjoint folds (Section 4.2, Table 3, Appendix D)

The co-substrate channel improves prediction under identity-clustered folds, most on collided records, which share an enzyme and main substrate but differ in co-substrates (Table 3). Removing it lowers R2R^2 from 0.450 to 0.439 over all 25,387 out-of-fold records and from 0.469 to 0.434 on the 3,251 collided records; both intervals exclude zero, and every fold shows the same direction (Appendix D). On the external benchmark, R2R^2 falls from 0.242 to 0.240, and the interval [-0.016, +0.011] includes both zero and the cross-validation effect of −0.011-0.011, so the external set does not distinguish them; no maximum-identity bin interval excludes zero either (Table D2). KiMChEE's external lead over the baselines (Table 4) is larger than this effect, so we attribute it to the model as a whole rather than to the co-substrate channel alone. For this ablation, requested by Reviewers 26g7 and TYaY, we removed only the co-substrate block and retrained all five fold models under the same folds, encoder adaptation, and ensemble settings as the final model.

3. Description of encoder adaptation brought in line with the model as trained (Section 3.5, Appendix C)

Section 3.5 and Appendix C now describe encoder adaptation as the model was actually trained; the model and all reported numbers are unchanged. The submitted version stated that the last four ProtT5 transformer blocks with its final layer norm, and the final transformer layer of each chemical encoder, were trained. In fact, the second training stage trains only the final layer norms of ProtT5 and MolT5, and every other encoder parameter, including all of ChemBERTa, is held fixed. Appendix C accordingly describes this factor of the configuration search as whether the second training stage is run. We found the discrepancy while re-examining the training code for these responses.

4. Reproducibility

The code, data, and models requested by Reviewers 2AWU and 26g7 are now in a Hugging Face repository: the training data with its fold assignments, the training code, the trained models for every configuration reported in the paper, including the adapted encoder weights, the inference and scoring code, the predictions of every baseline we evaluated, and the code we used to run each baseline's released weights. Its README maps each table of the paper to the command that produces it. The code that assembled the training data is covered by a third-party agreement and cannot be released, but the assembled data is provided in full, and the filtering rules that produced it are described in Section 4.1 and Appendix B. We will provide the repository path and a read token in a separate comment visible only to the reviewers and the Area Chair.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.