Multimodal Protein Language Models for Enzyme Kinetic Parameters: From Substrate Recognition to Conformational Adaptation

Fei Wang; Xinye Zheng; Kun Li; Yanyan Wei; Yuxin Liu; Ganpeng Hu; Tong Bao; Jingwen Yang

cs.CVMar 13, 2026

Multimodal Protein Language Models for Enzyme Kinetic Parameters: From Substrate Recognition to Conformational Adaptation

Authors: Fei Wang, Xinye Zheng, Kun Li, Yanyan Wei, Yuxin Liu, Ganpeng Hu, Tong Bao, Jingwen Yang

Organizations: School of Computer Science and Information Engineering, Hefei University of Technology · Institute of Artificial Intelligence, Hefei Comprehensive National Science Center · CVLab, College of Information Technology, United Arab Emirates University · Intelligent Interconnected Systems Laboratory of Anhui Province ·

Abstract

Predicting enzyme kinetic parameters quantifies how efficiently an enzyme catalyzes a specific substrate under defined biochemical conditions. Canonical parameters such as the turnover number ( $k_\text{cat}$ ), Michaelis constant ( $K_\text{m}$ ), and inhibition constant ( $K_\text{i}$ ) depend jointly on the enzyme sequence, the substrate chemistry, and the conformational adaptation of the active site during binding. Many learning pipelines simplify this process to a static compatibility problem between the enzyme and substrate, fusing their representations through shallow operations and regressing a single value. Such formulations overlook the staged nature of catalysis, which involves both substrate recognition and conformational adaptation. In this regard, we reformulate kinetic prediction as a staged multimodal conditional modeling problem and introduce the Enzyme-Reaction Bridging Adapter (ERBA), which injects cross-modal information via fine-tuning into Protein Language Models (PLMs) while preserving their biochemical priors. ERBA performs conditioning in two stages: Molecular Recognition Cross-Attention (MRCA) first injects substrate information into the enzyme representation to capture specificity; Geometry-aware Mixture-of-Experts (G-MoE) then integrates active-site structure and routes samples to pocket-specialized experts to reflect induced fit. To maintain semantic fidelity, Enzyme-Substrate Distribution Alignment (ESDA) enforces distributional consistency within the PLM manifold in a reproducing kernel Hilbert space. Experiments across three kinetic endpoints and multiple PLM backbones, ERBA delivers consistent gains and stronger out-of-distribution performance compared with sequence-only and shallow-fusion baselines, offering a biologically grounded route to scalable kinetic prediction and a foundation for adding cofactors, mutations, and time-resolved structural cues.

Figures & tables

Figure 2: Architecture of the proposed ERBA. It augments a sequence-only PLM with multimodal conditioning on substrate chemistry and pocket geometry. MRCA injects substrate fingerprints into enzyme embeddings to capture recognition specificity, and G-MoE integrates local 3D pocket structure to model conformational adaptation. Update and query paths couple both modules to the backbone, while ESDA aligns representations with the prior PLM, enabling accurate prediction of enzyme kinetic parameters.

Figure 3: G-Gating & Router pools sequence-substrate and structural cues to produce gating logits, activate the top- k geometry-relevant experts, and suppress others. Selected experts transform structure-conditioned features, and an MLP fuses them into a geometry-adaptive representation for kinetic regression.

Method	Venue	Modality			Kinetic Endpoint: $k_{\text{cat}}$				Kinetic Endpoint: $K_{\text{m}}$				Kinetic Endpoint: $K_{\text{i}}$
Method	Venue	$\mathbf{S}_{e}$	$\mathbf{S}_{m}$	$\mathbf{S}_{g}$	R 2 ↑	PCC↑	RMSE↓	MAE↓	R 2 ↑	PCC↑	RMSE↓	MAE↓	R 2 ↑	PCC↑	RMSE↓	MAE↓
Exp I: Existing SOTA Models
DLKcat [ 78 ]	Nat.Catal 2022	✓	✓	-	0.01	0.37	1.78	1.30	-	-	-	-	-	-	-	-
TurNup [ 76 ]	Nat.C 2023	✓	✓	-	0.39	0.65	1.60	1.19	-	-	-	-	-	-	-	-
UniKP [ 61 ]	Nat.C 2023	✓	✓	-	0.13	0.42	1.56	1.14	0.36	0.63	1.02	0.78	-	-	-	-
MPEK [ 48 ]	Brief.Bio 2024	✓	✓	-	0.33	0.59	1.34	1.05	0.30	0.56	1.40	1.09	-	-	-	-

Table 1: Performance comparison of existing SOTA methods and fine-tuned PLM backbones on enzyme kinetic prediction tasks. Results are presented for three kinetic parameters: turnover number (

k_{\text{cat}}

), Michaelis constant (

K_{\text{m}}

), and inhibition constant (

K_{\text{i}}

). The performance metrics include R 2 ↑, PCC↑, RMSE↓, and MAE↓. ERBA consistently outperforms the baselines across all metrics. Notably, the best value in each column is highlighted in red , and the second best in blue . ‡ indicates that the backbone of the PLMs is fine-tuned.

Index	Class Name	Kinetic Endpoint: $k_{\text{cat}}$					Kinetic Endpoint: $K_{\text{m}}$					Kinetic Endpoint: $K_{\text{i}}$
Index	Class Name	Count	R 2 $\uparrow$	PCC $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$	Count	R 2 $\uparrow$	PCC $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$	Count	R 2 $\uparrow$	PCC $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$
EC-1	Oxidoreductases	679	0.37	0.62	1.32	0.95	1,382	0.55	0.74	0.90	0.69	282	0.58	0.77	1.19	0.92
EC-2	Transferases	334	0.46	0.69	1.22	0.87	1,000	0.53	0.73	0.81	0.61	270	0.38	0.65	1.37	0.99
EC-3	Hydrolases	682	0.43	0.68	1.18	0.88	1,145	0.51	0.72	0.90	0.65	408	0.48	0.71	1.50	1.10
EC-4	Lyases	131	0.50	0.71	1.27	0.92	293	0.56	0.75	0.81	0.61	157	0.65	0.81	1.19	0.84
EC-5	Isomerases	72	0.46	0.73	1.39	0.94	123	0.61	0.78	0.84	0.64	42	0.67	0.83	1.20	0.95

Table 2: Enzyme category analysis across kinetic prediction tasks. The table provides a breakdown of enzyme categories in the datasets for

k_{\text{cat}}

K_{\text{m}}

, and

K_{\text{i}}

prediction tasks. It includes the distribution of each enzyme class, the number of samples per category, and performance metrics for each class. Notably, the best value in each column is highlighted in red , and the second best in blue .

Method	OOD Test @EITLEM- $k_{\text{cat}}$				OOD Test @EITLEM- $K_{\text{m}}$
Method	R 2 ↑	PCC↑	RMSE↓	MAE↓	R 2 ↑	PCC↑	RMSE↓	MAE↓
DLKcat [ 78 ]	0.01	0.15	3.87	3.05	-	-	-	-
TurNup [ 76 ]	0.21	0.38	2.46	2.11	-	-	-	-
UniKP [ 61 ]	0.08	0.20	1.76	1.24	0.14	0.32	1.54	1.08
MPEK [ 48 ]	0.17	0.41	1.65	1.18	0.12	0.35	1.71	1.32
EITLEM [ 82 ]	0.27	0.50	1.43	0.95	0.23	0.43	1.31	1.03

Table 3: Comparison with existing SOTA models on OOD tests.

k_{\text{cat}}

and

K_{\text{m}}

using the EITLEM [ 82 ] test sets.

Figure 4: Log-scaled experimental versus predicted values for the kinetic parameters $k_{\text{cat}}$ , $K_{\text{m}}$ , and $K_{\text{i}}$ . Each plot shows the absolute error less than or equal to 1 as a percentage, denoted as 1- $\text{Radio}_{\text{AE}}$ . The dashed red line represents perfect predictions.

Figure 5: Ablation studies on fusion order and manner. Comparison of different fusion strategies: $\mathbf{S}_{e}$ → $\mathbf{S}_{g}$ → $\mathbf{S}_{m}$ , Concat & MLP, and the proposed $\mathbf{S}_{e}$ → $\mathbf{S}_{m}$ → $\mathbf{S}_{g}$ /ERBA. Percentage improvements across metrics are highlighted in red .

Method	Modality			Kinetic Endpoint: $K_{\text{i}}$
Method	$\mathbf{S}_{e}$	$\mathbf{S}_{m}$	$\mathbf{S}_{g}$	R 2 ↑	PCC↑	RMSE↓	MAE↓
PLM Baseline	✓	-	-	0.49	0.71	1.38	1.05
w / MRCA	✓	✓	-	0.51	0.73	1.32	0.98
w / G-MoE	✓	✓	✓	0.54	0.74	1.27	0.94
w / ESDA (Ours)	✓	✓	✓	0.61	0.78	1.26	0.91

Table 4: Ablation study on the effect of different modules. Performance comparison for

K_{i}

prediction with various configurations: PLM baseline, MRCA, G-MoE, and our ESDA framework.

Class	Kinetic Dataset			Total	Details
Class	$\mathit{k}_{\text{cat}}$	$K_{\text{m}}$	$K_{\text{i}}$	Total
Dataset I (Main Test): $k_{\text{cat}},K_{\text{m}}$ and $K_{\text{i}}$ from [ 71 ]
Entries	23,151	41,174	11,929	76,254
EC 1 Oxidoreductases	7,756	13,931	2,744	24,431	Oxidoreductases catalyze diverse redox reactions by transferring electrons between molecules and substrates, and represent the largest EC class with broad kinetic and functional diversity.
EC 2 Transferases	4,155	10,182	2,706	17,043	Transferases catalyze the transfer of functional groups between diverse substrates, enabling the formation of new chemical bonds and the rapid and efficient interconversion of molecules.
EC 3 Hydrolases	8,065	11,025	4,083	23,173	Hydrolases catalyze the hydrolysis of diverse chemical bonds in biological molecules. They efficiently cleave covalent linkages in various substrates with variable enzyme kinetic parameters.

Table 5: Detailed breakdown of the enzyme kinetic datasets from [ 71 , 82 ] used for training and out-of-distribution (OOD) testing. We provide the number of entries, unique sequences, EC classes, organisms, and additional information for each enzyme class in the main and OOD test datasets.

Index	Class Name	Kinetic Endpoint: $k_{\text{cat}}$				Kinetic Endpoint: $K_{\text{m}}$				Kinetic Endpoint: $K_{\text{i}}$
Index	Class Name	R 2 $\uparrow$	PCC $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$	PCC $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$	PCC $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$
(a) Ankh3-5.7B [ 70 ] w/ ERBA
EC-1	Oxidoreductases	0.39	0.68	1.25	0.89	0.51	0.78	0.83	0.61	0.53	0.73	1.29	0.96
EC-2	Transferases	0.44	0.65	1.23	0.87	0.55	0.74	0.78	0.58	0.39	0.65	1.27	0.97
EC-3	Hydrolases	0.45	0.70	1.13	0.85	0.53	0.73	0.85	0.60	0.29	0.60	1.72	1.27
EC-4	Lyases	0.52	0.73	1.20	0.91	0.60	0.79	0.73	0.55	0.54	0.74	1.36	0.99

Table 6: Performance Comparison Across Enzyme Commission Classes. The table compares the performance of different backbone models and the impact of ERBA on kinetic parameter prediction for six EC classes: EC 1 (Oxidoreductases), EC 2 (Transferases), EC 3 (Hydrolases), EC 4 (Lyases), EC 5 (Isomerases), and EC 6 (Ligases). Metrics include R 2 , PCC, RMSE, and MAE for

k_{\text{cat}}

K_{\text{m}}

, and

K_{\text{i}}

, showcasing the effect of model size and architecture on the predictive accuracy across enzyme classes.

Figure 6: Error Distribution Comparison Across Different Backbone Models and ESM Sizes. It shows the error distribution of predicted versus experimental values for three kinetic parameters ( $k_{\text{cat}}$ , $K_{\text{m}}$ , and $K_{\text{i}}$ ) across four backbone models: (a) Ankh3-5.7B [ 70 ] , (b) ProtT5-3B [ 72 ] , (c) ESM2-650M [ 79 ] , and (d) ESM2-3B [ 79 ] , each augmented with ERBA. The plots show the proportion of predictions with absolute error less than or equal to 1, denoted as 1- $\text{Radio}_{\text{AE}}$ .

	Method	Kinetic Endpoint: $K_{\text{i}}$
	Method	R 2 ↑	PCC↑	RMSE↓	MAE↓
Stage 1	w / Concat & Self-Atention [ 83 ]	0.47	0.69	1.44	1.10
Stage 1	w / MRCA (Ours)	0.61	0.78	1.26	0.91

Table 7: Stage 1 (Recognition) ablation on

K_{\text{i}}

: Replacing Concat & Self-Attention with our MRCA markedly improves

R^{2}

and PCC while reducing RMSE/MAE under identical settings.

	Method	Kinetic Endpoint: $K_{\text{i}}$
	Method	R 2 ↑	PCC↑	RMSE↓	MAE↓
Stage 2	w / MoE [ 81 ]	0.50	0.71	1.39	1.05
Stage 2	w / G-MoE (Ours)	0.61	0.78	1.26	0.91

Table 8: Stage 2 (Adaptation) ablation on

K_{\text{i}}

: standard MoE vs. our G-MoE that routes by pocket geometry; geometry-aware experts yield consistent gains across all metrics.

	Method	Kinetic Endpoint: $K_{\text{i}}$
	Method	R 2 ↑	PCC↑	RMSE↓	MAE↓
Optimization	w / $\mathcal{L}_{2}$ Loss [ 85 ]	0.48	0.69	1.42	1.08
Optimization	w / ESDA (Ours)	0.61	0.78	1.26	0.91

Table 9: Objective ablation on

K_{\text{i}}

: plain

\mathcal{L}_{2}

vs. ESDA -regularized objective. Distribution alignment in RKHS stabilizes fine-tuning and improves R 2 /PCC while lowering RMSE/MAE.

Molecular Property Prediction

Molecular property prediction estimates chemical, physical, and biological properties from molecular structure and context.

See all 202 papers

Retrieval-Augmented Multimodal Learning for Enzyme-Substrate Interaction Prediction Under Low-Homology Shift
Jun 22, 2026Chen Liu, Bingxin Zhou, Xinyuan Wang +3Cross-Modal Representation Learning Domain Generalization
School of Information Science and Engineering, East China University of Science and Technology, Shanghai 200237, China · Institute of Natural Sciences and Zhangjiang Institute for Advanced Study, Shanghai Jiao Tong University, Shanghai 200240, China · School of Life Sciences and Biotechnology, Shanghai Jiao Tong University. Shanghai 200240, China +1
RXNRECer Enables Fine-grained Enzymatic Function Annotation through Active Learning and Protein Language Models
Mar 13, 2026Zhenkun Shi, Jun Zhu, Dehang Wang +7
Biodesign Center, Key Laboratory of Engineering Biology for Low-carbon Manufacturing, Tianjin Institute of Industrial Biotechnology, Chinese Academy of Sciences, 300308, Tianjin, China. · College of Computer Science and Technology, Harbin Engineering University, Street, Harbin, 10587, State, China.
Elucidating the Space of Enzymatic Reaction: A Unified Benchmark and Pretrained Model
Oct 8, 2026Yutong Hu, Tianming Huang, Yanbo Zhao +6Sequence-to-Sequence Models Chemical Reaction Prediction
Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory · The University of Sydney

Jun 22, 2026J/K move · Enter open · S save

Retrieval-Augmented Multimodal Learning for Enzyme-Substrate Interaction Prediction Under Low-Homology Shift

Chen Liu, Bingxin Zhou, Xinyuan Wang, Ming Li+2

School of Information Science and Engineering, East China University of Science and Technology, Shanghai 200237, China · Institute of Natural Sciences and Zhangjiang Institute for Advanced Study, Shanghai Jiao Tong University, Shanghai 200240, China · School of Life Sciences and Biotechnology, Shanghai Jiao Tong University. Shanghai 200240, China+1

Cross-Modal Representation Learning Domain Generalization

Open paper AskPDF arXiv

Enzyme substrate interaction (ESI) prediction is a fundamental computational task for biocatalyst discovery and reaction screening in large biochemical spaces. In practical settings, ESI prediction is challenged by sparse positive supervision and low-homology distribution shift, where test enzymes share limited sequence identity with those observed during training. To address these challenges, we propose RAMMESI, a retrieval-augmented multimodal framework for robust ESI prediction. RAMMESI learns explicit pairwise enzyme-substrate representations through directional cross-modal interaction modeling and adaptive fusion. To enhance robustness, RAMMESI retrieves neighboring enzymes at inference time, recombines them with the query substrate, and aggregates the resulting pairwise predictions as contextual evidence. To improve learning under sparse positive supervision, we further adopt an imbalance-aware weighted-BCE objective. Experiments on two ESI benchmarks under sequence-identity-aware splits demonstrate that RAMMESI achieves consistently strong performance, with particular advantages in more challenging low-identity regimes. In addition, the retrieval module improves multiple ESI backbones in a plug-and-play manner, suggesting that retrieval provides a general mechanism for improving robustness under homology shift.

Abstract

Molecular Property Prediction

Explore similar work

Retrieval-Augmented Multimodal Learning for Enzyme-Substrate Interaction Prediction Under Low-Homology Shift

RXNRECer Enables Fine-grained Enzymatic Function Annotation through Active Learning and Protein Language Models

Elucidating the Space of Enzymatic Reaction: A Unified Benchmark and Pretrained Model