Toward Automatic Derivation of Geometry-Based Descriptors as Surrogates for Complex Computational Approaches in...

Area of Science:

• Biocatalysis and enzyme engineering
• Computational chemistry and structural biology
• Drug discovery and molecular recognition

Background:

• Predicting enzyme-substrate interactions is crucial for biocatalysis and drug discovery.
• Current machine learning methods often require extensive data and lack mechanistic interpretability.
• Developing interpretable, data-efficient methods for substrate specificity prediction is essential.

Purpose of the Study:

• To develop a novel methodology for predicting substrate specificity using geometry-based descriptors derived from enzyme-substrate complexes.
• To simplify complex catalytic mechanisms into interpretable geometric filters.
• To provide mechanistic insights into substrate recognition and enable enzyme engineering.

Main Methods:

• Automatic derivation of geometry-based descriptors from enzyme-substrate complex structures.
• Simplification of catalytic mechanisms into interpretable geometric filters (inter-atomic distances, atomic pair accessibility).
• Validation using haloalkane dehalogenases (HLDs) and aldehyde reductases (AldRs) with minimal training data.

Main Results:

• Geometric filters demonstrated robust performance across diverse substrates for HLDs and AldRs.
• Achieved 77% accuracy and 94% sensitivity for HLDs, and 57% recall for AldRs on testing datasets.
• Outperformed state-of-the-art machine learning methods in substrate prediction for these enzyme families.
• Derived descriptors provided mechanistic insights into substrate recognition.

Conclusions:

• The proposed interpretable, mechanism-based approach requires minimal training data.
• Geometric descriptors offer a powerful tool for enzyme engineering and substrate screening.
• This methodology can be readily applied to newly characterized enzymes for predicting substrate specificity.