Pith. sign in

REVIEW 4 major objections 5 minor 40 references

FGW-CLIP achieves state-of-the-art enzyme–reaction retrieval by aligning both cross-domain and within-domain structure.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 17:38 UTC pith:OKJG6JWT

load-bearing objection Useful incremental advance in enzyme-reaction retrieval, but the ReactZyme gains are confounded by extra EC supervision and dataset-wide imputation. the 4 major comments →

arxiv 2512.08508 v2 pith:OKJG6JWT submitted 2025-12-09 q-bio.BM cs.LG

Multi-Alignment Contrastive Learning for Enzyme--Reaction Retrieval

Gengmo Zhou , Feng Yu , Wenda Wang , Zhifeng Gao , Guolin Ke , Zhewei Wei , Zhen Wang This is my paper
classification q-bio.BM cs.LG
keywords enzyme–reaction retrievalcontrastive learningfused Gromov–Wasserstein distancevirtual screeningEC numberrepresentation learningbiocatalysisoptimal transport
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

FGW-CLIP aims to prove that enzyme–reaction retrieval improves when the learned representation space respects the internal organization of each domain, not just the pairwise catalytic relationship. It jointly optimizes inter-domain contrastive alignment (enzyme–reaction), intra-domain contrastive alignment (reaction–reaction and enzyme–enzyme, with positive pairs defined by shared EC numbers), an EC-classification auxiliary loss, and a Gromov–Wasserstein-inspired regularization that encourages geometric consistency between the two spaces. The paper derives this objective as optimization of a fused Gromov–Wasserstein distance and reports state-of-the-art early-recognition metrics on EnzymeMap and consistent gains across all three ReactZyme splits, including settings with unseen enzymes and unseen reactions. If correct, this gives computational enzyme discovery a retrieval model that better ranks catalytically relevant candidates early, with direct implications for biocatalyst design and functional annotation.

Core claim

The central claim is that contrastive enzyme–reaction retrieval can be materially improved by adding higher-order relational alignment. The paper introduces FGW-CLIP, which treats reactions and enzymes as two metric spaces and uses a regularization term to minimize the Gromov–Wasserstein distance between their intra-domain similarity structures, while standard inter-domain contrastive loss handles the catalytic pairs and EC-based contrastive losses handle within-domain similarity. Theoretically, the full objective is shown to be equivalent to optimizing a constrained fused Gromov–Wasserstein distance, with the GW term acting as a structural regularizer that supplements sparse enzyme–reaction

What carries the argument

The load-bearing object is the fused Gromov–Wasserstein (FGW) distance, a metric that compares two spaces by aligning their internal pairwise-distance structures rather than only matching individual points. In FGW-CLIP, soft coupling matrices are built from learned cosine similarities: Γθ between reactions and enzymes, and Γψ1/Γψ2 within reactions and within enzymes. The regularization term LGW minimizes the GW discrepancy between the two intra-domain similarity matrices, with gradients of these matrices detached to stabilize training. This term is what injects higher-order relational information; the EC-number-supervised intra-domain contrastive losses provide the functional signal that sha

Load-bearing premise

The central claim rests on the premise that EC-number overlap is a reliable, informative signal for functional similarity within enzymes and reactions, and that the imputed EC values on ReactZyme do not distort the learned geometry or leak test information; if EC annotations are noisy or the imputation is arbitrary, the within-domain and GW losses could bias the representation and inflate retrieval gains.

What would settle it

Train FGW-CLIP on ReactZyme with only entries whose EC numbers are complete and verified (excluding '-' and 'n' tokens), keeping the evaluation protocol fixed; if the advantage over baselines shrinks or disappears, the claimed gains depend on the imputed EC labels. Alternatively, shuffle EC numbers assigned to intra-domain positive pairs; if BEDROC and MRR stay the same, the EC signal is not the active ingredient.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • On EnzymeMap, FGW-CLIP raises BEDROC(α=85) from roughly 44.7% to 48.7% and BEDROC(α=20) from about 63.0% to 66.7% over the strongest baseline, with higher enrichment factors.
  • On ReactZyme, FGW-CLIP reports state-of-the-art Top-k and MRR in both enzyme-to-reaction and reaction-to-enzyme retrieval on the time split, and on the enzyme-similarity and reaction-similarity splits, including unseen enzymes and unseen reactions.
  • Ablations show each added component contributes: intra-domain reaction contrast improves BEDROC, EC prediction improves enrichment, and the GW regularization with detached intra-domain matrices and weight α=0.1 gives the best overall performance.
  • The theoretical derivation frames contrastive retrieval as a constrained FGW optimization, indicating a principled way to add structural supervision to contrastive learning models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A direct test of the paper's mechanism would be to train on ReactZyme using only high-confidence, complete EC annotations and rerun retrieval; if the gains shrink, the imputation of '-' and 'n' tokens is doing more work than the structural alignment itself.
  • The same multi-alignment recipe should transfer to other retrieval problems with a functional taxonomy, such as drug-target interaction or metabolic-reaction annotation, where within-domain structure could be defined by shared pathways or gene ontologies.
  • Because the enzyme encoder is frozen, the structural alignment operates on a fixed feature space; fine-tuning the protein encoder or using a structure-aware encoder would test whether the GW regularizer remains beneficial or is partly compensating for a fixed backbone.
  • The paper leaves reaction directionality out of the reaction representation; adding directionality or atom mapping as additional structure within the reaction domain is a natural extension that could further improve the GW alignment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes FGW-CLIP, a contrastive learning framework for enzyme–reaction retrieval and virtual screening. The method adds to a standard reaction–enzyme contrastive loss (Eq. 5) two intra-domain contrastive losses defined by shared EC numbers (Eq. 6), an EC classification loss (Eq. 7), and a Gromov–Wasserstein-inspired regularization term (Eq. 8). The total objective (Eq. 10) is claimed to optimize a constrained fused Gromov–Wasserstein distance (Proposition 1). Experiments on EnzymeMap and ReactZyme report state-of-the-art performance on BEDROC/EF and Top-k/MRR metrics, with ablations used to attribute gains to the different components.

Significance. If the empirical results hold, FGW-CLIP would be a useful contribution to enzyme screening, and the idea of coupling intra-domain structure with inter-domain contrastive learning is appealing. The paper provides extensive evaluations on two benchmarks and includes a derivation connecting the loss to FGW distance. However, the central attribution of the gains to the GW regularization is weakened by the simultaneous introduction of EC-based supervision, by EC imputation that may leak test-set information, and by ablations that are not fully controlled. The theoretical result is also presented as stronger than the proof supports.

major comments (4)
  1. [§4.3–4.4, Eq. (10); Table 3] The central empirical claim—that the GW term drives the SOTA results—is confounded by EC supervision. In every ReactZyme comparison, FGW-CLIP differs from baselines by adding both the GW term (Eq. 8) and the EC-based objectives L_reaction, L_enzyme (Eq. 6) and L_EC (Eq. 7); baselines (including CLIPZyme) are trained without EC supervision. The large gains in Table 3 (e.g., reaction-to-enzyme Top1 0.392 vs 0.176 for CLIPZyme) could therefore be due to EC labels. The EnzymeMap ablation in Table 4 does not isolate GW: the step from R_E+R_R+E_E+EC (45.83) to FGW-CLIP (48.66) adds GW, but R_E+R_R alone (48.08) already nearly matches the full model, and adding E_E+EC to R_E+R_R decreases performance (45.83), indicating strong interaction effects. A ReactZyme ablation with and without the GW term at fixed α and λ is needed to support the attribution.
  2. [§5.2.4] The EC imputation procedure is likely to leak test-set information. The Mode strategy fills missing EC components with the most frequent value at that position in 'the dataset', and the Max strategy uses the current maximum plus 1 or 2. If these statistics are computed over the entire dataset (train+test), the imputed EC labels used to construct L_reaction, L_enzyme, and L_EC encode information from test samples. The Max rule also assigns arbitrary unique values to many missing entries, creating spurious shared-EC positive pairs. The imputation should be fitted on the training split only, and the authors should report the proportion of missing EC values and the sensitivity of Table 3 to the imputation rule.
  3. [§4.4, Appendix C, Proposition 1] The theoretical connection is an approximation, not the exact equivalence stated. The proof of Lemma 2 drops the sample-dependent coefficient a_i ('we disregard ai and constant in Liot', after Eq. 16), and the symmetrization step leading to Eq. (21) modifies the loss. Consequently, Eq. (11) describes a related objective rather than the actual L_FGW of Eq. (10). The paper should state this explicitly and justify dropping a_i, which weights positives in the multi-positive InfoNCE loss. As written, the claim that FGW-CLIP 'could be derived from optimizing a specific constraint-fused Gromov-Wasserstein distance' is too strong.
  4. [§5.3, Tables 4–6] No error bars or multiple-seed results are reported, so the significance of the GW term is unclear. The gain of the full model over R_E+R_R is only 0.58 BEDROC85 (48.66 vs 48.08), whereas adding R_R to R_E gives 2.14 points. The non-monotonic ablation (R_E+R_R = 48.08 vs R_E+R_R+E_E+EC = 45.83) suggests optimization instability. For a state-of-the-art claim, at least three seeds with mean±std should be reported, and a ReactZyme ablation isolating the GW term is essential.
minor comments (5)
  1. [Tables 2 and 3] Several entries are concatenated, e.g., Table 2 '13.407.81' and '13.577.61'; Table 3 has rows like '165.30660.1789MAT-2D' and '0.38300.7090'. Please fix the formatting.
  2. [§5.2.4] The phrase 'improvements of over 20 percentage in Top-k metrics' is imprecise; the Top1 gain in reaction-to-enzyme retrieval is about 17 points (0.392 vs 0.228 for the strongest baseline).
  3. [References] Reference [31] appears to duplicate [10] (both are Tianhao Yu et al., 'Enzyme function prediction using contrastive learning'). Please correct.
  4. [General] No code availability statement is included; releasing code would substantially help reproducibility.
  5. [Figure 1] The caption says 'for clarity, only one classification head is shown', but the main text (Eq. 7) describes four level-specific heads; please clarify how the four EC-level heads are combined.

Circularity Check

1 steps flagged

Empirical SOTA claims are external and not circular; the FGW theoretical framing is a post-hoc restatement of the defined loss.

specific steps
  1. self definitional [Section 4.4 / Eq. 10; Appendix C / Eq. 20]
    "By integrating the training objectives in Section 4.3, we can derive the overall training objective for FGW-CLIP, denoted as LFGW, as follows: LFGW = (1−α)(L reaction-enzyme +L reaction +L enzyme)−2αL GW +λL EC (10). ... we establish a connection between LFGW and the fused Gromov-Wasserstein distance optimization problem under a specific constraint through the proposition 1. ... Disregarding the constant terms, we can simplify the optimization objective as follows: GW(Γψ1_d ,Γψ2_d ,Γθ) =−2tr((Γθ)⊤Γψ1_d ΓθΓψ2_d) (20)"

    Eq. 8 defines L_GW as the negative of the quadruple sum over Γψ1_d Γψ2_d Γθ Γθ; Eq. 20 identifies exactly that trace (up to constants) with GW. Thus the 'derivation' of FGW-CLIP from a fused Gromov-Wasserstein objective in Proposition 1 is a restatement of the definition of L_GW, not an independent first-principles result. The theoretical novelty claim reduces to naming the loss; it does not generate or constrain the benchmark predictions.

full rationale

The central empirical claims—SOTA on EnzymeMap and across all ReactZyme splits—are evaluated on external benchmarks against independent baselines, so they are not forced by construction. The only definitional reduction is the theoretical packaging: the loss is assembled so that L_GW is the GW trace term, and Proposition 1 then announces that the method optimizes FGW distance. That is a consistent equivalence but a post-hoc one. The self-citation of S-MolSearch [28] for the IOT link is not load-bearing because Appendix C supplies a self-contained proof and the underlying IOT connection is attributed to Shi et al. [27]. The ReactZyme EC Mode/Max imputation (§5.2.4) is a correctness/leakage concern—dataset-wide statistics could encode test information—but it is not circular: imputed EC values are supervision, not model outputs that are then scored. The appended limitation (E.1) concerns scope, not circularity. Overall, partial circularity in the FGW framing is worth a moderate score, but the benchmark results retain independent content.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central method rests on standard contrastive learning plus a GW-inspired regularizer; no new physical entities are postulated. The free parameters alpha, lambda, tau are tuned through ablations on EnzymeMap. The main load-bearing assumption is that EC-number overlap is a meaningful within-domain alignment signal, which is especially delicate on ReactZyme where EC annotations are incomplete and imputed.

free parameters (3)
  • alpha (GW loss weight) = 0.1
    Selected via ablation on EnzymeMap; balances inter-domain vs intra-domain alignment (Table 5).
  • lambda (EC loss weight) = 0.1
    Selected via ablation on EnzymeMap; smaller value prevents EC loss from dominating (Table 6).
  • tau (temperature) = 0.05
    Used in all contrastive losses; set following prior contrastive learning practice (Section D).
axioms (4)
  • domain assumption EC-number overlap defines meaningful positive pairs for within-domain alignment
    Section 4.3.1 treats any two samples sharing an EC number as positive; if EC annotations are noisy/incomplete, intra-domain losses are biased.
  • domain assumption Frozen ESM2 and Uni-Mol provide adequate base representations
    Section 4.2 keeps ESM2 frozen and uses sum readout of Uni-Mol; the method's gains assume these pretrained features are sufficient.
  • standard math The IOT equivalence from Shi et al. (2023) extends to the multi-alignment setting
    Lemma 2 and Proposition 1 rely on the inverse optimal transport formulation; the proof in Appendix C involves approximations ('disregarding constants').
  • domain assumption Benchmark catalytic labels are correct
    Ground-truth enzyme-reaction pairs from EnzymeMap/ReactZyme are assumed accurate for training and evaluation.

pith-pipeline@v1.3.0-alltime-deepseek · 22351 in / 15358 out tokens · 145955 ms · 2026-08-03T17:38:11.280361+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Multi-Alignment Contrastive Learning for Enzyme--Reaction Retrieval." pith.science (2026). https://pith.science/paper/OKJG6JWT

@misc{pith2026251208508,
  author       = {Pith},
  title        = {Pith review of: Multi-Alignment Contrastive Learning for Enzyme--Reaction Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OKJG6JWT}},
  note         = {Machine review of arXiv:2512.08508}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Identifying enzymes that catalyze target biochemical reactions is a key step in computational enzyme discovery and biocatalyst design. Recent representation-learning methods formulate this problem as enzyme--reaction matching, where paired enzymes and reactions are embedded into a shared space. However, most existing approaches primarily rely on pairwise enzyme--reaction supervision and make limited use of the relationships within reaction sets or enzyme families. This work introduces a multi-alignment contrastive learning framework for biochemical retrieval. The framework jointly models cross-domain compatibility between enzymes and reactions and within-domain relationships induced by functional annotations. In addition, a Gromov--Wasserstein-inspired regularization objective encourages geometric consistency between the learned enzyme and reaction representation spaces. By combining pairwise catalytic supervision with higher-order relational alignment, the model captures both direct enzyme--reaction associations and broader functional organization. We evaluate the approach on enzyme virtual screening and bidirectional enzyme--reaction retrieval tasks. Experiments on EnzymeMap show improved early-recognition performance under BEDROC and enrichment-factor metrics compared with strong contrastive baselines. On ReactZyme, the method achieves consistent gains across time-based, enzyme-similarity, and reaction-similarity splits, demonstrating robustness to unseen enzymes and unseen reactions. Ablation studies further indicate that within-domain alignment, functional supervision, and the geometric regularization term each contribute to the observed improvements. These results suggest that modeling multiple forms of alignment can improve contrastive retrieval models for enzyme discovery, reaction annotation, and related computational biology applications.

Figures

Figures reproduced from arXiv: 2512.08508 by Feng Yu, Gengmo Zhou, Guolin Ke, Wenda Wang, Zhen Wang, Zhewei Wei, Zhifeng Gao.

Figure 1
Figure 1. Figure 1: Overview of FGW-CLIP Framework. Reactants and products, along with their structures, [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: t-SNE visualization of enzyme representations. Left: pretrained checkpoint. Right: FGW [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 1 canonical work pages

  1. [1]

    UniProt: the Universal protein knowledgebase in 2025

    “UniProt: the Universal protein knowledgebase in 2025”. In:Nucleic Acids Research53.D1 (2025), pp. D609–D617

    2025

  2. [2]

    Enzyme function and evolution through the lens of bioinformatics

    Antonio JM Ribeiro et al. “Enzyme function and evolution through the lens of bioinformatics”. In:Biochemical Journal480.22 (2023), pp. 1845–1863

    2023

  3. [3]

    Basic local alignment search tool

    Stephen F Altschul et al. “Basic local alignment search tool”. In:Journal of molecular biology 215.3 (1990), pp. 403–410

    1990

  4. [4]

    ModEnzA: accurate identification of metabolic enzymes using func- tion specific profile HMMs with optimised discrimination threshold and modified emission probabilities

    Dhwani K Desai et al. “ModEnzA: accurate identification of metabolic enzymes using func- tion specific profile HMMs with optimised discrimination threshold and modified emission probabilities”. In:Advances in bioinformatics2011.1 (2011), p. 743782

    2011

  5. [5]

    Gapped BLAST and PSI-BLAST: a new generation of protein database search programs

    Stephen F Altschul et al. “Gapped BLAST and PSI-BLAST: a new generation of protein database search programs”. In:Nucleic acids research25.17 (1997), pp. 3389–3402

    1997

  6. [6]

    Hidden Markov models in computational biology: Applications to protein modeling

    Anders Krogh et al. “Hidden Markov models in computational biology: Applications to protein modeling”. In:Journal of molecular biology235.5 (1994), pp. 1501–1531

    1994

  7. [7]

    HH-suite3 for fast remote homology detection and deep protein annotation

    Martin Steinegger et al. “HH-suite3 for fast remote homology detection and deep protein annotation”. In:BMC bioinformatics20 (2019), pp. 1–15

    2019

  8. [8]

    COFACTOR: an accurate comparative algorithm for structure-based protein function annotation

    Ambrish Roy, Jianyi Yang, and Yang Zhang. “COFACTOR: an accurate comparative algorithm for structure-based protein function annotation”. In:Nucleic acids research40.W1 (2012), W471–W477

    2012

  9. [9]

    COFACTOR: improved protein func- tion prediction by combining structure, sequence and protein–protein interaction information

    Chengxin Zhang, Peter L Freddolino, and Yang Zhang. “COFACTOR: improved protein func- tion prediction by combining structure, sequence and protein–protein interaction information”. In:Nucleic acids research45.W1 (2017), W291–W299. 12

    2017

  10. [10]

    Enzyme function prediction using contrastive learning

    Tianhao Yu et al. “Enzyme function prediction using contrastive learning”. In:Science 379.6639 (2023), pp. 1358–1363

    2023

  11. [11]

    CLIPZyme: Reaction-Conditioned Virtual Screening of Enzymes

    Peter G Mikhael, Itamar Chinn, and Regina Barzilay. “CLIPZyme: Reaction-Conditioned Virtual Screening of Enzymes”. In:arXiv preprint arXiv:2402.06748(2024)

    Pith/arXiv arXiv 2024

  12. [12]

    ReactZyme: A Benchmark for Enzyme-Reaction Prediction

    Chenqing Hua et al. “ReactZyme: A Benchmark for Enzyme-Reaction Prediction”. In:The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track. 2024.URL:https://openreview.net/forum?id=xepxnDQoGq

    2024

  13. [13]

    Care: a benchmark suite for the classification and retrieval of enzymes

    Jason Yang et al. “Care: a benchmark suite for the classification and retrieval of enzymes”. In: Advances in Neural Information Processing Systems37 (2024), pp. 3094–3121

    2024

  14. [14]

    Learning transferable visual models from natural language supervision

    Alec Radford et al. “Learning transferable visual models from natural language supervision”. In:International conference on machine learning. PmLR. 2021, pp. 8748–8763

    2021

  15. [15]

    MLIP: Efficient Multi-Perspective Language-Image Pretraining with Exhaus- tive Data Utilization

    Yu Zhang et al. “MLIP: Efficient Multi-Perspective Language-Image Pretraining with Exhaus- tive Data Utilization”. In:arXiv preprint arXiv:2406.01460(2024)

    Pith/arXiv arXiv 2024

  16. [16]

    iclip: Bridging image classification and contrastive language-image pre- training for visual recognition

    Yixuan Wei et al. “iclip: Bridging image classification and contrastive language-image pre- training for visual recognition”. In:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023, pp. 2776–2786

    2023

  17. [17]

    Cross-Modal Retrieval Meets Inference: Improving Zero-Shot Classifica- tion with Cross-Modal Retrieval

    Seongha Eom et al. “Cross-Modal Retrieval Meets Inference: Improving Zero-Shot Classifica- tion with Cross-Modal Retrieval”. In:arXiv preprint arXiv:2308.15273(2023)

    Pith/arXiv arXiv 2023

  18. [18]

    Gromov–Wasserstein distances and the metric approach to object matching

    Facundo Mémoli. “Gromov–Wasserstein distances and the metric approach to object matching”. In:Foundations of computational mathematics11 (2011), pp. 417–487

    2011

  19. [19]

    Linear-time gromov wasserstein distances using low rank couplings and costs

    Meyer Scetbon, Gabriel Peyré, and Marco Cuturi. “Linear-time gromov wasserstein distances using low rank couplings and costs”. In:International Conference on Machine Learning. PMLR. 2022, pp. 19347–19365

    2022

  20. [20]

    Optimal transport for structured data with application on graphs

    Vayer Titouan et al. “Optimal transport for structured data with application on graphs”. In: International Conference on Machine Learning. PMLR. 2019, pp. 6275–6284

    2019

  21. [21]

    Fused gromov-wasserstein graph mixup for graph-level classifications

    Xinyu Ma et al. “Fused gromov-wasserstein graph mixup for graph-level classifications”. In: Advances in Neural Information Processing Systems36 (2024)

    2024

  22. [22]

    DeepEC: An error correction framework for dose prediction and organ segmentation using deep neural networks

    Han Wang et al. “DeepEC: An error correction framework for dose prediction and organ segmentation using deep neural networks”. In:International Journal of Intelligent Systems 35.12 (2020), pp. 1987–2008

    2020

  23. [23]

    ProteInfer, deep neural networks for protein functional inference

    Theo Sanderson et al. “ProteInfer, deep neural networks for protein functional inference”. In: Elife12 (2023), e80942

    2023

  24. [24]

    Uni-Mol: A Universal 3D Molecular Representation Learning Frame- work

    Gengmo Zhou et al. “Uni-Mol: A Universal 3D Molecular Representation Learning Frame- work”. In:The Eleventh International Conference on Learning Representations. 2023.URL: https://openreview.net/forum?id=6K2RM6wVqKu

    2023

  25. [25]

    Evolutionary-scale prediction of atomic-level protein structure with a language model

    Zeming Lin et al. “Evolutionary-scale prediction of atomic-level protein structure with a language model”. In:Science379.6637 (2023), pp. 1123–1130

    2023

  26. [26]

    Representation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. “Representation learning with contrastive predictive coding”. In:arXiv preprint arXiv:1807.03748(2018)

    Pith/arXiv arXiv 2018

  27. [27]

    Understanding and generalizing contrastive learning from the inverse optimal transport perspective

    Liangliang Shi et al. “Understanding and generalizing contrastive learning from the inverse optimal transport perspective”. In:International conference on machine learning. PMLR. 2023, pp. 31408–31421

    2023

  28. [28]

    S-MolSearch: 3D Semi-supervised Contrastive Learning for Bioactive Molecule Search

    Gengmo Zhou et al. “S-MolSearch: 3D Semi-supervised Contrastive Learning for Bioactive Molecule Search”. In:The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2024.URL:https://openreview.net/forum?id=wJAF8TGVUG

    2024

  29. [29]

    EnzymeMap: curation, validation and data-driven prediction of enzymatic reactions

    Esther Heid et al. “EnzymeMap: curation, validation and data-driven prediction of enzymatic reactions”. In:Chem. Sci.14 (48 2023), pp. 14229–14242.DOI:10.1039/D3SC02048G

    work page doi:10.1039/d3sc02048g 2023

  30. [30]

    BRENDA, the ELIXIR core data resource in 2021: new developments and updates

    Antje Chang et al. “BRENDA, the ELIXIR core data resource in 2021: new developments and updates”. In:Nucleic Acids Research(2020)

    2021

  31. [31]

    Enzyme function prediction using contrastive learning

    Tianhao Yu et al. “Enzyme function prediction using contrastive learning”. In:Science(2023)

    2023

  32. [32]

    Evaluating virtual screening methods: good and bad metrics for the “early recognition

    Jean-François Truchon and Christopher I Bayly. “Evaluating virtual screening methods: good and bad metrics for the “early recognition” problem”. In:Journal of chemical information and modeling47.2 (2007), pp. 488–508

    2007

  33. [33]

    Condensed graph of reaction: considering a chemical reaction as one single pseudo molecule

    Frank Hoonakker et al. “Condensed graph of reaction: considering a chemical reaction as one single pseudo molecule”. In:Int. J. Artif. Intell. Tools20.2 (2011), pp. 253–270. 13

    2011

  34. [34]

    Frame averaging for invariant and equivariant network design

    Omri Puny et al. “Frame averaging for invariant and equivariant network design”. In:arXiv preprint arXiv:2110.03336(2021)

    Pith/arXiv arXiv 2021

  35. [35]

    The SWISS-PROT protein knowledgebase and its supplement TrEMBL in 2003

    Brigitte Boeckmann et al. “The SWISS-PROT protein knowledgebase and its supplement TrEMBL in 2003”. In:Nucleic acids research31.1 (2003), pp. 365–370

    2003

  36. [36]

    Rhea, the reaction knowledgebase in 2022

    Parit Bansal et al. “Rhea, the reaction knowledgebase in 2022”. In:Nucleic acids research 50.D1 (2022), pp. D693–D700

    2022

  37. [37]

    A general model to predict small molecule substrates of enzymes based on machine and deep learning

    Alexander Kroll et al. “A general model to predict small molecule substrates of enzymes based on machine and deep learning”. In:Nature communications14.1 (2023), p. 2787

    2023

  38. [38]

    Molecule attention transformer

    Łukasz Maziarka et al. “Molecule attention transformer”. In:arXiv preprint arXiv:2002.08264 (2020)

    Pith/arXiv arXiv 2002

  39. [39]

    SaProt: protein language modeling with structure-aware vocabulary

    Jin Su et al. “SaProt: protein language modeling with structure-aware vocabulary”. In:bioRxiv (2023), pp. 2023–10. 14 A Dataset Details and Baselines A.1 Enzyme Screening EnzymeMapBased on the original EnzymeMap dataset [29], this dataset consists of biochemical reactions linked to UniProt IDs and EC numbers. It contains 46,356 enzyme-driven reactions, in...

    2023

  40. [40]

    For enzyme representations, ESM [25] and the structure- aware protein language model SaProt [39] are utilized

    and UniMol-2D [24] for 2D molecular graphs, and MAT-3D and UniMol-3D for 3D molecular conformations as reaction representations. For enzyme representations, ESM [25] and the structure- aware protein language model SaProt [39] are utilized. Additionally, an equivariant graph neural network FANN [34]) is employed to enhance residue-level representations. Fo...

    arXiv 2001