Enzyme Substrate Prediction from Three-Dimensional Feature Representations Using Space-Filling Curves.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biochemistry

Background:

• Accurate protein property prediction requires compact and interpretable 3D structural representations.
• Existing methods often rely on sequence embeddings, but geometric approaches offer complementary insights.

Purpose of the Study:

• To construct and evaluate 3D feature representations of protein structures using space-filling curves (SFCs).
• To assess the performance of SFC-based representations for enzyme substrate and cofactor prediction.
• To compare SFC-based methods with existing protein feature representations.

Main Methods:

• Utilized space-filling curves (Hilbert, Morton) to create 1D representations from 3D protein structures.
• Generated 3D protein structures using AlphaFold2 for short-chain dehydrogenase/reductases (SDRs) and S-adenosylmethionine-dependent methyltransferases (SAM-MTases).
• Employed gradient-boosted tree classifiers for enzyme classification tasks, including substrate and cofactor selectivity.

Main Results:

• Achieved binary prediction accuracy of 0.77-0.91 and AUC of 0.83-0.92 for enzyme classification tasks.
• Investigated the impact of amino acid encoding, spatial orientation, and SFC parameters on prediction accuracy.
• Demonstrated the effectiveness of SFC-based representations for enzyme function prediction.

Conclusions:

• Space-filling curves provide a promising geometry-based approach for generating effective protein structural representations.
• SFC-based representations are complementary to sequence-based embeddings like evolutionary scale modeling (ESM).
• This work advances the development of interpretable and accurate protein feature representations for predicting enzyme function.