Decompose and Recompose: Reference-Centered Multimodal Adaptation of Protein Language Models for Enzyme Turnover Prediction

Predicting enzyme turnover numbers ($k_{\mathrm{cat}}$) from protein sequences and reaction chemistry supports virtual screening, which requires distinguishing activity differences between enzyme--reaction pairs. Recent interaction models face two representation challenges. First, they merge reactants and products before enzyme interaction, leaving separate enzyme responses implicit despite the two sides' distinct roles in recognition, transformation, and release. Second, cross-attention readouts mix reaction-dependent changes with an enzyme-specific residue background, which can obscure activity-relevant differences across samples. To address these challenges, we introduce EnzRef, a lightweight framework for multimodal adaptation of frozen protein language models. EnzRef decomposes reaction chemistry into reactant and product conditions, models their interactions with the enzyme separately, and recomposes the resulting responses for activity prediction. During these interactions, Reference-Centered Cross-Attention expresses each response as a deviation from the enzyme's own residue background to expose chemically conditioned differences for activity discrimination. With 1.72M trainable parameters, EnzRef achieves state-of-the-art source prediction on CatPred-DB and zero-shot transfer to TurNuP-DB. Across ten independent runs, EnzRef achieves 19.8\% higher mean $R^2$ than the strongest baseline MCKcat in zero-shot transfer to unseen enzymes. It also predicts activity differences more accurately between pairs sharing an enzyme or a reaction.

Authors

Institutions

Publication Details

Journal
bioRxiv (Cold Spring Harbor Laboratory)
Published
2026-11-17
DOI
https://doi.org/10.64898/2026.11.15.766718
Primary Topic
Machine Learning in Bioinformatics
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Decompose and Recompose: Reference-Centered Multimodal Adaptation of Protein Language Models for Enzyme Turnover Prediction

Tengfei Ma, Shan Lin, Aocheng Ding, Leyi Wei et al.
3 links
bioRxiv (Cold Spring Harbor Laboratory)
Machine Learning in Bioinformatics
preprint

Decompose and Recompose: Reference-Centered Multimodal Adaptation of Protein Language Models for Enzyme Turnover Prediction

Tengfei Ma, Shan Lin, Aocheng Ding, Leyi Wei, Xuan Lin, Xiangrong Liu
preprint en

Abstract

Predicting enzyme turnover numbers ($k_{\mathrm{cat}}$) from protein sequences and reaction chemistry supports virtual screening, which requires distinguishing activity differences between enzyme--reaction pairs. Recent interaction models face two representation challenges. First, they merge reactants and products before enzyme interaction, leaving separate enzyme responses implicit despite the two sides' distinct roles in recognition, transformation, and release. Second, cross-attention readouts mix reaction-dependent changes with an enzyme-specific residue background, which can obscure activity-relevant differences across samples. To address these challenges, we introduce EnzRef, a lightweight framework for multimodal adaptation of frozen protein language models. EnzRef decomposes reaction chemistry into reactant and product conditions, models their interactions with the enzyme separately, and recomposes the resulting responses for activity prediction. During these interactions, Reference-Centered Cross-Attention expresses each response as a deviation from the enzyme's own residue background to expose chemically conditioned differences for activity discrimination. With 1.72M trainable parameters, EnzRef achieves state-of-the-art source prediction on CatPred-DB and zero-shot transfer to TurNuP-DB. Across ten independent runs, EnzRef achieves 19.8\% higher mean $R^2$ than the strongest baseline MCKcat in zero-shot transfer to unseen enzymes. It also predicts activity differences more accurately between pairs sharing an enzyme or a reaction.

bioRxiv (Cold Spring Harbor Laboratory)
Hunan University (CN), Xiamen University (CN), State Key Laboratory of Chemobiosensing and Chemometrics (CN), Xiangtan University (CN), Macao Polytechnic University (MO), Xiamen University of Technology (CN)
National Natural Science Foundation of China
Machine Learning in Bioinformatics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon