KmPred: prediction of Michaelis constants (Km) using an integrative machine learning framework

Area of Science:

• Biochemistry
• Computational Biology
• Machine Learning

Background:

• The Michaelis constant (Km) quantifies enzyme-substrate affinity, crucial for understanding enzyme kinetics.
• Traditional in vitro assays for determining Km are time-consuming and labor-intensive.
• Advancements in protein and chemical databases enable computational prediction of kinetic parameters.

Purpose of the Study:

• To develop and validate KmPred, a machine learning framework for accurate Km prediction.
• To integrate protein sequence embeddings and substrate molecular descriptors for enhanced predictive power.
• To establish a computational approach for accelerating enzyme characterization.

Main Methods:

• Developed KmPred, a machine learning framework combining protein sequence embeddings (from language models) and substrate SMILES-derived molecular descriptors.
• Utilized LSTM and Transformer models for feature extraction from enzyme sequences.
• Employed XGBoost for final Km regression predictions.
• Benchmarked performance on the MPEK and Kroll et al. datasets.

Main Results:

• KmPred achieved competitive performance on both MPEK and Kroll datasets, outperforming baseline models.
• On the MPEK dataset, the best model yielded an R² of 0.7049 and PCC of 0.8398.
• On the Kroll dataset, KmPred achieved an R² of 0.5519 and PCC of 0.7440.
• Demonstrated robust and generalizable Km prediction by combining multi-modal features and advanced ML architectures.

Conclusions:

• The integration of multi-modal features (protein sequence and ligand properties) with advanced machine learning enables robust Km prediction.
• KmPred offers a scalable computational approach for predictive enzymology, accelerating enzyme characterization.
• This AI-driven methodology has significant implications for biotechnology, metabolic engineering, and drug discovery pipelines.