LM2KM: feature extraction from pretrained protein and molecular language models combined with quantum mechanics (QM) features for KM predictions
- 1Department of Pharmaceutical Sciences and Computational Chemical Genomics Screening Center, School of Pharmacy, University of Pittsburgh Pittsburgh Pennsylvania 15261 USA [email protected].
- 2Center for Drug Discovery & School of Pharmacy and Pharmaceutical Sciences, Northeastern University Boston MA USA.
- 3Department of Computer Science, Hunter College, The City University of New York New York NY USA [email protected].
Abstract
The Michaelis constant (K M) is a key parameter for characterizing enzyme-substrate interactions and plays an essential role in enzyme kinetics modeling and physiologically based pharmacokinetic (PBPK) applications. However, experimental determination of K M is labor-intensive, and existing computational prediction approaches remain limited in accuracy and scalability. In this study, we systematically evaluated pretrained protein and molecule language models for K M prediction by integrating four protein language models (PLMs) and two molecule language models (MLMs) to encode enzyme sequences and substrate structures, respectively. Principal component analysis (PCA) was applied to reduce feature dimensionality, and a wide range of representation combinations were assessed. Our results indicated that protein representations derived from the last four layers of ProtT5 using approximately 70 features, combined with 20 CLS-based molecular features from MolFormer, provided a well-balanced and efficient representation for K M prediction. In addition, the incorporation of quantum mechanics (QM) descriptors consistently improved prediction performance across all tested configurations, with the model R ² of 0.537. In summary, this work provides a flexible and extensible framework for large-scale K M prediction, with strong potential to facilitate enzyme pathway identification, drug and lead screening, and sequence-based protein design, supporting broader applications in enzyme kinetics-driven modeling and drug development.
Related
- Toward Autonomous Extraction and Building of Machine-Readable Molecular Models from Publications Using Large Language Models
- Accuracy and Efficiency Benchmarks of Pretrained Machine Learning Potentials for Molecular Simulations
- BigMixSolDB: Extraction of a solubility database in solvent mixtures with an uncertainty-quantified large language model-based pipeline
- Embedded Morgan Fingerprints for more efficient molecular property predictions with machine learning
- Crystal structure, Hirshfeld surface analysis, computational study, molecular docking predictions and biological evaluation of a new Schiff base (E)-2-((2,4-dichlorobenzylidene)amino)aniline
BibTeX
@article{zhai2026lm2km,
title = {{LM2KM: feature extraction from pretrained protein and molecular language models combined with quantum mechanics (QM) features for KM predictions}},
author = {Jingchen Zhai and Lianjin Cai and Taoyu Niu and Lei Xie and Junmei Wang},
journal = {Digital discovery},
year = {2026},
month = oct,
doi = {10.1039/d6dd00139d},
url = {https://europepmc.org/article/PMC/PMC13725812},
note = {CC BY},
}
RIS
TY - JOUR
TI - LM2KM: feature extraction from pretrained protein and molecular language models combined with quantum mechanics (QM) features for KM predictions
AU - Jingchen Zhai
AU - Lianjin Cai
AU - Taoyu Niu
AU - Lei Xie
AU - Junmei Wang
PY - 2026
DA - 2026/10/20
JO - Digital discovery
DO - 10.1039/d6dd00139d
UR - https://europepmc.org/article/PMC/PMC13725812
AB - The Michaelis constant (K M) is a key parameter for characterizing enzyme-substrate interactions and plays an essential role in enzyme kinetics modeling and physiologically based pharmacokinetic (PBPK) applications. However, experimental determination of K M is labor-intensive, and existing computational prediction approaches remain limited in accuracy and scalability. In this study, we systematically evaluated pretrained protein and molecule language models for K M prediction by integrating four protein language models (PLMs) and two molecule language models (MLMs) to encode enzyme sequences and substrate structures, respectively. Principal component analysis (PCA) was applied to reduce feature dimensionality, and a wide range of representation combinations were assessed. Our results indicated that protein representations derived from the last four layers of ProtT5 using approximately 70 features, combined with 20 CLS-based molecular features from MolFormer, provided a well-balanced and efficient representation for K M prediction. In addition, the incorporation of quantum mechanics (QM) descriptors consistently improved prediction performance across all tested configurations, with the model R ² of 0.537. In summary, this work provides a flexible and extensible framework for large-scale K M prediction, with strong potential to facilitate enzyme pathway identification, drug and lead screening, and sequence-based protein design, supporting broader applications in enzyme kinetics-driven modeling and drug development.
SN - 2635-098X
ER -