Multimodal Protein Language Models for Enzyme Kinetic Parameters

Introduction

Enzyme kinetics describes the rates at which enzymes catalyze biochemical reactions. Predicting kinetic parameters from molecular information can help researchers identify promising enzymes, understand substrate specificity, and accelerate enzyme engineering and synthetic biology research.

Three important parameters are commonly used to characterize enzyme behavior:

Turnover number (\(k_{\text{cat}}\)): Measures the number of substrate molecules an enzyme can convert into products per unit time under specified conditions when the enzyme is saturated with substrate. Michaelis constant (\(K_m\)): Represents the substrate concentration at which the reaction rate reaches half of its maximum under the standard Michaelis-Menten model. Inhibition constant (\(K_i\)): Characterizes the strength of an inhibitor's interaction with an enzyme in a specified inhibition model.

Accurately predicting these parameters is challenging because enzyme activity depends on more than the compatibility between an enzyme sequence and a substrate molecule. The substrate must first be recognized by the enzyme, after which the active site may undergo conformational changes that influence binding and catalysis. Traditional machine learning pipelines often represent the enzyme using a pretrained protein language model and encode the substrate using molecular features. These representations are then combined using operations such as concatenation or shallow cross-attention before a regression model predicts the kinetic parameter.

Although these approaches can learn useful molecular relationships, they may fail to capture the distinct stages of substrate recognition and conformational adaptation. They can also struggle to incorporate structural information without disrupting the biochemical knowledge encoded in pretrained protein representations. The researchers address these limitations by introducing ERBA, a staged multimodal framework that conditions protein language models on substrate chemistry and active-site geometry. Its architecture is designed to reflect the biological stages of enzyme-substrate interaction rather than treating catalysis as a single static compatibility score.

Methodology

Enzyme-Reaction Bridging Adapter (ERBA)

ERBA is designed to integrate enzyme sequence, substrate chemistry, and active-site structural information into a unified prediction framework. Rather than combining all molecular inputs through a single fusion operation, the model introduces information in stages. This design follows the principle that substrate recognition and conformational adaptation represent related but distinct processes in enzyme catalysis. The framework incorporates three principal components: Molecular Recognition Cross-Attention, Geometry-aware Mixture-of-Experts, and Enzyme-Substrate Distribution Alignment.

Molecular Recognition Cross-Attention (MRCA)

The first stage focuses on substrate recognition.

MRCA injects substrate information into the enzyme representation using cross-attention. This mechanism allows the model to learn relationships between the molecular characteristics of a substrate and the sequence-level representation of an enzyme. By conditioning the enzyme representation on the substrate, the model aims to capture substrate specificity: the ability of an enzyme to recognize and interact with particular molecules. This stage provides a substrate-conditioned representation that serves as the foundation for subsequent structural adaptation modeling.

Geometry-aware Mixture-of-Experts (G-MoE)

The second stage focuses on conformational adaptation.

Enzyme active sites can change their conformations during substrate binding. These structural adjustments, commonly associated with induced fit, can influence the stability of the enzyme-substrate complex and the resulting catalytic behavior. G-MoE integrates active-site geometry into the substrate-conditioned enzyme representation. It uses a mixture-of-experts architecture to route samples to specialized experts associated with different active-site or binding-pocket characteristics. This design allows the model to represent diverse structural configurations rather than relying on a single shared transformation for every enzyme-substrate pair. By combining substrate-conditioned sequence information with geometric features, G-MoE aims to capture structural factors that influence enzyme kinetics.

Enzyme-Substrate Distribution Alignment (ESDA)

Integrating information from multiple modalities creates an additional challenge: the resulting representations must remain compatible with the biochemical information learned during protein language model pretraining. ERBA addresses this issue through Enzyme-Substrate Distribution Alignment (ESDA). ESDA encourages distributional consistency between the learned representations and the pretrained protein language model's representation space. The method uses a reproducing kernel Hilbert space (RKHS) formulation to align the relevant distributions.

This component is intended to preserve the semantic and biochemical information encoded in pretrained representations while allowing the model to incorporate substrate and structural information. Together, MRCA, G-MoE, and ESDA establish a staged framework for predicting enzyme kinetic parameters from multiple sources of molecular information.

Experimental Evaluation

The researchers evaluate ERBA across three enzyme kinetic prediction targets and multiple protein language model backbones. The study compares the proposed framework with sequence-only and shallow-fusion baselines.

The evaluation examines both predictive performance and generalization to out-of-distribution examples, where the test data differ from the distributions represented in training.

According to the reported results, ERBA achieves consistent performance improvements across the evaluated kinetic endpoints and protein language model backbones. The framework also demonstrates stronger out-of-distribution performance than the comparison baselines.

These findings support the use of staged multimodal conditioning as an alternative to treating enzyme kinetics prediction as a simple fusion of static enzyme and substrate representations.

Open Access CVF+1

Discussion

The ERBA framework emphasizes that enzyme kinetics prediction involves multiple interacting biological factors. Enzyme sequence provides information about the protein, substrate chemistry describes the molecule being processed, and active-site geometry contributes information about structural adaptation during binding.

By separating substrate recognition from conformational adaptation, the framework introduces a more biologically structured approach to multimodal learning.

MRCA models the relationship between substrate chemistry and enzyme sequence representations. G-MoE adds structural information through specialized experts, while ESDA helps preserve the distributional properties of pretrained protein representations.

This design is particularly relevant to enzyme engineering, where researchers may need to estimate kinetic behavior across many enzyme-substrate combinations before conducting expensive experimental validation.

However, computational predictions do not replace biochemical experiments. Predicted kinetic parameters should be interpreted in the context of the model's training data, evaluation settings, and the biochemical conditions under which the parameters are measured.

Conclusion

Multimodal Protein Language Models for Enzyme Kinetic Parameters introduces ERBA, a framework for predicting enzyme kinetics by integrating protein sequences, substrate chemistry, and active-site geometry through staged multimodal conditioning.

The approach combines Molecular Recognition Cross-Attention for substrate specificity, Geometry-aware Mixture-of-Experts for conformational adaptation, and Enzyme-Substrate Distribution Alignment to maintain consistency with pretrained protein representations.

The reported experiments demonstrate consistent gains across three kinetic prediction targets and improved out-of-distribution performance compared with sequence-only and shallow-fusion baselines.

By aligning multimodal learning with key stages of enzyme-substrate interaction, the research offers a framework for more biologically informed enzyme kinetics prediction and provides a foundation for future extensions involving cofactors, mutations, and time-resolved structural information.