Pseudodata-Guided Invariant Representation Learning Boosts the Out-of-Distribution Generalization in Enzymatic...

Area of Science:

• Biochemistry and Bioinformatics
• Computational Biology
• Enzyme Kinetics

Background:

• Accurate enzyme kinetic parameter prediction is crucial for understanding enzyme mechanisms and guiding enzyme engineering.
• Current deep learning enzyme-substrate interaction (ESI) predictors struggle with sequence-divergent, out-of-distribution (OOD) data, limiting their real-world applicability.

Purpose of the Study:

• To develop a robust module, O²DENet, that enhances the generalization ability of ESI predictors on OOD datasets.
• To improve the accuracy and stability of enzyme kinetic parameter predictions, specifically kcat and Km.

Main Methods:

• O²DENet employs biologically and chemically informed perturbation augmentation for enzyme-substrate pairs.
• It enforces invariant representation learning by ensuring consistency between original and augmented data representations.
• The module is designed as a lightweight, plug-and-play component for existing ESI models.

Main Results:

• O²DENet consistently improved predictive performance for kcat and Km across stringent OOD benchmarks.
• The module demonstrated state-of-the-art accuracy and robustness compared to other methods.
• Significant enhancements in predictive performance were observed when O²DENet was integrated with various ESI models.

Conclusions:

• O²DENet offers a general and effective strategy for enhancing the stability and deployability of data-driven enzyme kinetics predictors.
• The proposed method addresses the limitations of current ESI predictors in handling biologically relevant perturbations.
• O²DENet facilitates more reliable enzyme engineering applications through improved prediction robustness.