AUKAT: Conditional VAE-Driven Augmentation and Neural Modeling of Enzyme Turnover Numbers

Area of Science:

• Biochemistry
• Computational Biology
• Bioinformatics

Background:

• Enzyme turnover number (kcat) prediction is crucial for systems biology, metabolic engineering, and drug discovery.
• Accurate kcat estimation is hindered by limited and unevenly distributed experimental data.

Purpose of the Study:

• To develop AUKAT, an integrated framework combining conditional generative modeling and deep neural prediction for improved kcat estimation.
• To address the challenge of data scarcity in biochemical modeling.

Main Methods:

• Utilized a conditional variational autoencoder to generate synthetic training data in embedding space.
• Implemented a selection pipeline to ensure data reliability based on agreement across independent evaluators.
• Employed a hybrid convolutional neural network and transformer architecture for kcat prediction using substrate, enzyme functional, and species embeddings.

Main Results:

• Incorporating synthetic data improved predictive performance for both random forest and neural network models, with greater gains for neural networks.
• Benchmarking against DLKcat showed comparable accuracy on standard datasets and improved generalization on unseen subsets, particularly for low-similarity substrates and enzymes.
• AUKAT-human demonstrated enhanced prediction accuracy for human enzyme kinetics through pre-training and fine-tuning.

Conclusions:

• AUKAT offers a scalable approach to enzyme kinetics prediction, effectively mitigating data scarcity issues in biochemical modeling.
• The framework provides a practical solution for enhancing the accuracy and reliability of kcat estimations.
• AUKAT balances the utilization of substrate, enzyme functional, and species information for more robust predictions.