A Deep Retrieval-Enhanced Meta-Learning Framework for Enzyme Optimum pH Prediction.

Area of Science:

• Biochemistry and Bioinformatics
• Computational Biology
• Enzyme Engineering

Background:

• Enzyme function is highly sensitive to pH, making optimal pH (pHopt) prediction critical for enzyme design and application.
• Experimental determination of pHopt for numerous newly discovered or designed enzymes is time-consuming and costly.
• In-silico methods, particularly machine learning, offer a cost-effective alternative for predicting enzyme pHopt.

Purpose of the Study:

• To develop and validate Venus-DREAM, a novel computational model for predicting enzyme optimal pH (pHopt).
• To leverage few-shot learning and k-nearest neighbors (kNN) with protein language models for accurate pHopt prediction.
• To enable high-throughput virtual screening of enzymes for desired pH profiles.

Main Methods:

• Venus-DREAM employs a k-nearest neighbors (kNN) regression algorithm to determine the optimal number of neighbors (k).
• Enzyme similarity is quantified using cosine similarity of embeddings derived from protein language models (PLMs).
• A few-shot learning approach, based on the Reptile algorithm, adapts a model to k-nearest enzymes for specialized pHopt prediction.

Main Results:

• Venus-DREAM achieves state-of-the-art accuracy in predicting enzyme optimal pH (pHopt).
• The model effectively treats pHopt prediction as a few-shot learning task, learning from limited labeled data.
• The method facilitates efficient, high-throughput virtual exploration of protein sequence space for enzymes with specific pHopt.

Conclusions:

• Venus-DREAM provides a highly accurate and efficient computational tool for predicting enzyme optimal pH (pHopt).
• This approach significantly reduces the need for extensive experimental validation in enzyme engineering.
• The methodology is adaptable for predicting other enzyme functions, broadening its applicability in protein science.