UniZyme: A Unified Protein Cleavage Site Predictor Enhanced with Enzyme Active-Site Knowledge

Thanks for your valuable feedback. We address your questions in the following:

Response to clarity of framework diagram (Weakness 1)

Thank you very much for your valuable and practical suggestion. We will revise this diagram in the camera‑ready version.

Response to limitation of enzyme encoder (Weakness 2)

We use a single, consistent AWSEM parameter set for all enzyme families, as detailed in Appendix C and our GitHub repository. Specifically, we disable electrostatics (kelectrostatics = 0), enforce a 12‑residue sequence separation, generate 500 decoys by preserving backbone coordinates and randomizing side‑chains, and include the standard bond, angle, dihedral, contact, hydrogen‑bond, water‑mediated, and burial terms. We chose this uniform configuration because Freiberger et al. demonstrated that the same AWSEM coarse‑grained potential successfully identifies frustration patterns around active sites across diverse protease families [1].

Response to Data presentation (Weakness 3)

Thank you for the suggestion. In the camera‑ready version, we will update the figures and include more informative visualizations.

Response to pooling weight function f(.) (Weakness 4)

Here, f(.) is a learnable mapping that transforms each predicted active‑site probability into its final pooling weight, in direct analogy to how we use Gaussian kernels to map energy and distance into attention biases.

Concretely, we first pass the scalar probability $\hat a_i$ through a Gaussian basis expansion:

$\phi_{\mathrm{act}}(\hat a_i) = \bigl[\phi_{\mathrm{act},1}(\hat a_i), \dots, \phi_{\mathrm{act},K}(\hat a_i)\bigr] \in \mathbb{R}^K,$

which produces a richer, multi‑dimensional embedding. We then apply an MLP to collapse this embedding to a single weight:

$w_i = f(\hat a_i) = \mathrm{MLP}\bigl(\phi_{\mathrm{act}}(\hat a_i)\bigr).$

Response to interpretable analysis of the relationship between the enzyme's active site and its cleavage site (Weakness 5 and Question 1)

To address this, we provide an interpretability analysis that explores the relationship between predicted active-site probabilities and the quality of cleavage site prediction. Specifically, we aim to test whether higher-confidence active site signals are associated with more accurate cleavage predictions, thus demonstrating that the model indeed uses active-site information in a meaningful and interpretable way.

Here we present the PR-AUC of cleavage site prediction across different ranges of predicted active site probabilities, aiming to illustrate the interpretability of how active site signals influence substrate cleavage site prediction. We categorize all test samples into five groups based on the average predicted active-site probability (of actual active sites) $\hat{a}_i$ of the enzyme. From the table below, we can find that cleavage prediction improves with higher predicted active site probabilities in both zero-shot and supervised settings, indicating that the model effectively leverages active site signals.

Response to impact of the length of enzymes (Question 2)

To assess whether enzyme length influences cleavage prediction performance, we grouped test enzymes by sequence length and reported the PR-AUC under both zero-shot and supervised settings.

As shown in the table below, UniZyme maintains strong and stable performance across a wide range of lengths. While shorter enzymes ([0, 300)) achieve slightly higher accuracy, the model continues to perform robustly even for longer sequences, demonstrating its generalization ability across enzyme sizes.

Reference

[1] M.I. Freiberger, A.B. Guzovsky, P.G. Wolynes, R.G. Parra, & D.U. Ferreiro, Local frustration around enzyme active sites, Proc. Natl. Acad. Sci. U.S.A. 116 (10) 4037-4043

2. Interpretable Analysis Based on UniZyme

2.1 Sensitivity of Cleavage-site Predictions to Active-site Information

We decompose the influence of active-site information by computing gradient-based sensitivities to the upstream active-site prediction probabilities $\hat{a}_i$ via input attribution $\partial y/\partial \hat{a}_i$, as well as sensitivities to active-site and background residue embeddings via $\partial y/\partial h_i$.

From Table 1, we observe that predicted active-site residues consistently exhibit higher attribution scores across both supervised and zero-shot settings. This indicates that the UniZyme model significantly and consistently relies on active-site predictions to determine cleavage-site specificity. This biologically plausible dependency aligns closely with established experimental observations and biological knowledge.

Table 1. Attribution magnitudes (higher means more influence) for active-site vs. background residues in supervised and zero-shot settings.

2.2 Perturbation on Predicted Active-site Pooling Weights

To further demonstrate that UniZyme correctly relates cleavage-site predictions to active-site information, we perform an intervention analysis on the pooling step. Specifically, we selectively disrupt the contributions from residues predicted as active sites (the top 5 residues ranked by predicted probabilities $\hat{a}_i$) and compare this intervention to perturbations on randomly selected non-active-site residues.

Table 2 shows that disrupting predicted active-site contributions causes a large drop in PR-AUC in both supervised and zero-shot settings, whereas perturbing arbitrary residues has minimal effect. These results indicate that cleavage-site prediction performance critically depends on the amplified signal provided by the predicted active sites, further confirming the biological validity of the UniZyme model’s internal logic.

Table 2. Effect of perturbing pooling weights on cleavage PR-AUC in supervised and zero-shot settings.

2.3 Ablation Study (Active-site Supervision and Component Removal)

We build upon the original ablation experiments (UniZyme\A) that isolate the contributions of active-site-aware pooling. Additionally, we introduce a variant trained without active-site prediction supervision (cleavage-only) to specifically highlight the role of mechanistic guidance during training.

Results in Table 3 show that active-site prediction supervision and active-site-aware pooling provide significant benefits. This finding reinforces that active-site information is effectively injected during training, amplified through pooling, and consistently relied upon by UniZyme for accurate cleavage prediction.

Table 3. Ablation results on supervised and zero-shot enzyme families (PR-AUC).

Thanks for your comprehensive feedback and suggestions. We address your questions in the following:

Response to the comparative analysis of active site prediction module (Weaknesses 1)

We clarify that the primary objective of UniZyme is to leverage active-site prediction as an auxiliary task to enhance enzyme modeling, rather than optimizing active-site prediction performance alone. Our ablation studies in Section 4.4 clearly demonstrate the effectiveness of this auxiliary task for improving enzyme-catalyzed substrate cleavage-site prediction.

We will include a comparative analysis of our active‑site prediction module against other established active‑site prediction methods in the camera‑ready version.

Response to the limitation of 3D structures (Weaknesses 2)

We clarify that our method is primarily built upon predicted structures:

Training phase: around 90% of enzyme and substrate structures in our training set were predicted using OmegaFold or from AlphaFoldDB.

Test phase: as shown in the table below, over 80% of enzyme-substrate pairs in the test set rely entirely on predicted structures.

To assess structural sensitivity, we partitioned the test set into four categories according to the structural source of the enzyme and substrate. The results below indicate only minor performance differences between predicted and experimental structures, demonstrating that our model maintains strong predictive capability even with predicted or lower-quality structural data.

Response to the Reference of Table 2 and Table 7 (Question 1)

Thank you for catching this point. We will correct it in the camera‑ready version.

Response to the reason of using ESM-2 (Question 2)

We chose the ESM‑2 esm2_t12_35M_UR50D variant primarily for its lightweight architecture and fast inference speed, which made it highly suitable for large-scale training under limited computational resources.

At the time of model development, ESM-3 was only released at sizes ≥1.6B parameters, and the smaller ESM-C variants (300M and 600M) were not yet available. To validate this design choice, here we present ESM‑2 with ESM‑C 300M and 600M and observed only marginal performance improvements, while the inference cost increased significantly. This confirms that the use of a lightweight model like ESM‑2 35M strikes a practical and effective balance between performance and efficiency.

Response to the related protein function prediction tasks (Question 3)

To assess the general benefits of our UniZyme enzyme encoder, we conducted the EC number classification task on protease (EC 3.4..), reusing the cleavage dataset split described in our work. Specifically, we compared UniZyme with existing enzyme encoders of ReactZyme and ClipZyme in EC functional prediction. From the table below, we observe that UniZyme achieves the highest AUROC. These results suggest that our encoder captures generalizable functional signals beyond cleavage prediction. We plan to extend this analysis to GO (Gene Ontology) function prediction and incorporate more baseline comparisons in future work.

Results of predicting EC numbers with different enzyme encoders

Response to the impact of flexible regions (Question 4)

We thank the reviewer for the suggestion. While proteins are flexible, prior studies show that catalytic residues at enzyme active sites are significantly more rigid than other regions, with lower B-factors on average. For example, a study of 69 apo-enzyme structures found that active site residues lie in regions of reduced atomic displacement compared to non-catalytic residues [1] This structural rigidity ensures that a static representation captures the essential catalytic geometry.

Response to statistical hypothesis testing (Limitation 1)

To further support the reported performance gaps, we conducted statistical hypothesis testing using two-sample t-tests. The table below presents the t-values and p-values comparing UniZyme against ClipZyme and ReactZyme across both Supervised and Zero-shot settings on the 69-family and 23-family test sets.

Overall, UniZyme shows statistically significant improvements over both ReactZyme and ClipZyme in both Supervised and Zero‑shot evaluations, with the strongest effects observed in the Supervised setting.

Reference

[1] Yuan Z, Zhao J, Wang ZX. Flexibility analysis of enzyme active sites by crystallographic temperature factors. Protein Eng. 2003 Feb;16(2):109-14.

We thank the reviewer for highlighting the importance of benchmarking the active-site prediction module. In response, we conducted additional experiments comparing UniZyme to two representative structure-based baselines:

1. GraphEC [1]: a recent graph neural network developed for enzyme function prediction and active site prediction
2. NodeCoder [2]: a graph-based model to predict active sites of modeled protein structures

These models utilize the same information as UniZyme, which include enzyme structures and sequences. For implementation, we employed their official GitHub repositories. To ensure a fair comparison, we retrained both GraphEC and NodeCoder on the same large-scale active-site prediction dataset used by UniZyme. Statistics for the training and test sets of the active-site prediction task are presented below.

Table: Data Statistics for the active site prediction

We report the AUROC, AUPR, precision, recall, and F1 scores in the table below. These results clearly demonstrate that UniZyme significantly outperforms the baseline models in active site prediction. UniZyme achieves superior performance in active-site prediction because its biochemically-informed enzyme encoder effectively leverages local energetic frustration.

Specifically, previous studies indicate that local energetic frustration, referring to regions in a protein structure that are not optimized for minimal energy, commonly occurs around enzyme active sites. Notably, the energetic frustration scores are computed directly from protein structures without requiring additional experimental measures. By structurally embedding this biochemical insight into its enzyme encoder, UniZyme more accurately identifies active sites, thereby significantly outperforming baseline models.

Table: Active-Site Prediction Performance

References

[1] Song, Y., Yuan, Q., Chen, S. et al. Accurately predicting enzyme functions through geometric graph learning on ESMFold-predicted structures. Nat Commun 15, 8180 (2024).

[2] Abdollahi, N., Tonekaboni, S. A. M., Huang, J., Wang, B., & MacKinnon, S. NodeCoder: a graph-based machine learning platform to predict active sites of modeled protein structures. NeurIPS MLSB 2021.

Thanks for agreeing with our contributions. We address your questions in the following:

Response to Weakness 1

Thank you for your suggestions in presentation. We will revise according to your suggestions to improve the presentation quality in the camera-ready version.

Response to methodological contributions (Weakness 2)

We’d like to highlight how UniZyme goes well beyond simply combining existing elements:

First of all, our work addresses a novel problem by developing a unified protein cleavage site predictor for diverse proteolytic enzymes. Unlike prior approaches that train separate models per enzyme family, our UniZyme leverages a biochemically informed enzyme encoder that integrates active-site knowledge. The following technology is specifically designed for the challenge of enzyme catalyzed cleavage site prediction:

Biochemically-Informed Enzyme Encoder: We clarify that previous findings indicating that energetic frustration scores can highlight functional regions within enzymes were purely observational and had not been incorporated into enzyme modeling. Inspired by this, we propose a novel enzyme encoder in UniZyme, which integrates energetic frustration scores into the enzyme encoding process. This innovation significantly enhances the modeling of enzyme functions in substrate cleavage tasks.

Enzyme Encoder Augmented by Active Site Prediction: Active sites play a critical role in protein cleavage, thus providing essential insights into enzyme functions. To leverage this information, we introduce an auxiliary active-site prediction task to strengthen the enzyme encoder. To our knowledge, UniZyme is among the earliest methods to integrate an active-site prediction task to capture functionally relevant enzyme representations.

Active Site-Aware Pooling: We design an active site-aware pooling mechanism, whose pooling weights are based on the predicted active site probabilities. This encourages the model to focus on catalytically relevant segments of the enzyme

Extensive ablation studies in Sec. 4.4 further empirically demonstrate the effectiveness of our novel designs.

Response to the similarity threshold in splitting training and test set (Question 1)

A wealth of studies on enzyme function has shown that when sequence similarity falls below 60%, enzymes exhibit markedly greater functional divergence [1]. Moreover, previous deep-learning efforts in this area have likewise adopted a 60% similarity cutoff [2].

To further verify the impact of sequence similarity, we grouped the test set by sequence identity to the training set. From the table below, we observe that even at very low similarity levels, UniZyme’s performance only slightly decreases, and it still consistently outperforms baseline methods. This result demonstrates the strong generalization capability of UniZyme under challenging conditions.

Response to the encoder consistency (Question 2)

DeepDigest, DeepCleave, ProsperousPlus, CAT3, and ScreenCap3 are cleavage-site predictors trained on individual enzyme family. These models do not encode enzymes explicitly and thus lack a generalizable enzyme encoder. For ReactZyme and ClipZyme, they are originally designed for enzyme-substrate reaction prediction. Hence, we reused only their enzyme encoder modules in our framework for a fair comparison:

ReactZyme uses ESM-2 + MLP as the encoder.

ClipZyme uses a pretrained EGNN to encode enzyme structure.

Neither ReactZyme nor ClipZyme utilizes active-site information in their enzyme encoder training. In contrast, UniZyme incorporates a biochemically-informed encoder that is augmented with active-site knowledge. The comparisons among these models presented below demonstrate the effectiveness of using active-site information in UniZyme.

Response to the dataset curation (Question 3)

We curated a high-quality dataset specifically designed for the input requirements of our unified model, following a standardized and transparent pipeline described in Appendix A. Specifically, we:

1. Downloading enzyme-substrate pairs from the MEROPS database.

2. Collecting substrate sequences from UniProt and verified enzyme sequences for consistency between MEROPS and UniProt, discarding mismatches.

3. Filtering sequences exceeding 1,500 residues to ensure manageable input lengths.

4. Expanding the dataset by propagating cleavage annotations to all enzymes within the same MEROPS family, following a standard strategy previously used in cleavage site prediction literature (e.g., ProsperousPlus).

5. Collecting structural data from PDB and AlphaFoldDB; for missing entries, we generated 3D structures using OmegaFold.

This resulted in a cleavage site dataset containing ~220,000 enzyme–substrate pairs across 866 enzymes. Then we applied rigorous train/validation/test splits for evaluation:

Supervised setting: Selected enzyme families with ≥5 substrates. Within each family, we ensured test substrates shared <50% sequence identity with training data (Needleman–Wunsch alignment). The split ratio was 70/10/20. The final supervised test set include 69 enzyme families with 20,360 enzyme-substrate pairs.

Zero-shot setting: Selected enzyme families with ≥5 enzymes. We held out 20% of enzymes per family that shared <60% identity with any training or pretraining enzyme, to simulate novel enzyme scenarios. The final zero-shot test set includes 23 enzyme families with 5,345 enzyme-substrate pairs.

All baseline models were retrained on this dataset using the same train/validation/test splits to ensure fair and consistent comparisons. The dataset and code are publicly available via the GitHub link provided in our paper.

Given the large and diverse test sets covering 69 families in supervised settings and 23 in zero-shot settings, our benchmarks comprehensively evaluate model generalization across diverse enzymes and substrates.

Reference

[1] Röttig M, Rausch C, Kohlbacher O. Combining Structure and Sequence Information Allows Automated Prediction of Substrate Specificities within Enzyme Families. PLOS Comput Biol. 2010;6(1):1–8.

[2] Hua C, Zhong B, Luan S, Hong L, Wolf G, Precup D, Zheng S. ReactZyme: a benchmark for enzyme–reaction prediction. In: Proceedings of the 38th International Conference on Neural Information Processing Systems (NIPS ’24); 2025:Article 832. Vancouver, BC, Canada: Curran Associates Inc.

Response to Weakness 1: “The general benefits of the encoder or representation design is not discussed”

To assess the general benefits of our UniZyme enzyme encoder, we conducted the EC number classification task on protease (EC 3.4.*.*), reusing the cleavage dataset split described in our work. Specifically, we compared UniZyme with existing enzyme encoders of ReactZyme and ClipZyme in EC functional prediction. From the table below, we observe that UniZyme achieves the highest AUROC. These results suggest that our encoder captures generalizable functional signals beyond cleavage prediction. We plan to extend this analysis to GO (Gene Ontology) function prediction and incorporate more baseline comparisons in future work.

Results of predicting EC numbers with different enzyme encoders

Response to Question 1: Theoretical and experimental support of dataset expansion

We expanded the dataset at the enzyme family level based on empirical evidence, biological reasoning, and prior practice:

Active Sites Are Conserved Within Families. Proteases from the same family often share similar catalytic mechanisms and substrate preferences, even if their overall sequences differ. For example, trypsin and chymotrypsin (S01 family), as well as MMP-2 and MMP-9 (M10 family), have nearly identical active sites and exhibit overlapping cleavage patterns [1]. This consistency supports the use of shared annotations within a family.

Experimental Protocols Commonly Use Family Representatives. In biochemical and proteomic studies, it’s standard to select one enzyme as a stand-in for the entire family. Caspase-3, for instance, is frequently used to represent all C14 family members [2], while MMP-9 often stands for the M10 family. These choices reflect a practical assumption that members of the same family behave similarly in experiments [3].

Previous Models Follow the Same Approach. Earlier predictive models such as DeepDigest[4], DeepCleave[5], and ProsperousPlus[6] were trained using family-level data without relying on individual enzyme sequences. This modeling strategy reflects a broadly accepted assumption that enzymes from the same family tend to share cleavage patterns.

MEROPS Classification Supports Annotation Sharing. The MEROPS [7] database organizes proteases into families based on both sequence similarity and functional traits. Members of a family typically have conserved catalytic residues, similar active-site geometry, and shared substrate specificity. These features make it reasonable to extend known cleavage annotations across enzymes in the same family.

Referrence

[1] Cieplak P, Strongin AY. Matrix metalloproteinases - From the cleavage data to the prediction tools and beyond. Biochim Biophys Acta Mol Cell Res. 2017 Nov;1864(11 Pt A):1952-1963.

[2] Wejda M, Impens F, Takahashi N, Van Damme P, Gevaert K, Vandenabeele P. Degradomics reveals that cleavage specificity profiles of caspase-2 and effector caspases are alike. J Biol Chem. 2012 Oct 5;287(41):33983-95.

[3] Alessandro Bonadio, Solomon Oguche, Tali Lavy, Oded Kleifeld, Julia Shifman bioRxiv 2023.04.11.536383

[4] Jinghan Yang, Zhiqiang Gao, Xiuhan Ren, Jie Sheng, Ping Xu, Cheng Chang, and Yan Fu. Deepdigest: prediction of protein proteolytic digestion with deep learning. Analytical Chemistry, 93(15):6094–6103, 2021

[5] Fuyi Li, Jinxiang Chen, André Leier, Tatiana Marquez-Lago, Quanzhong Liu, Yanze Wang, Jerico Revote, A Ian Smith, Tatsuya Akutsu, Geoffrey I Webb, Lukasz Kurgan, and Jiangning Song. Deepcleave: a deep learning predictor for caspase and matrix metalloprotease substrates and cleavage sites. Bioinformatics, 36(4):1057–1065, 09 2019.

[6] Fuyi Li, Xudong Guo, Cong Wang, Tatsuya Akutsu, Geoffrey Webb, Lachlan Coin, Lukasz Kurgan, and Jiangning Song. Prosperousplus: a one-stop and comprehensive platform for accurate protease-specific substrate cleavage prediction and machine-learning model construction. Briefings in Bioinformatics, 24, 09 2023.

[7] Neil D Rawlings, Alan J Barrett, and Alex Bateman. Merops: the database of proteolytic enzymes, their substrates and inhibitors. Nucleic Acids Research, 40(Database issue):D343–D350, 2012.