Accurately predicting optimal conditions for microorganism proteins through geometric graph learning and language...

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Enzyme Engineering

Background:

• Extremophilic proteins offer valuable industrial applications due to their stability under extreme conditions.
• Experimental determination of optimal protein conditions is time-consuming.
• Previous computational models lacked comprehensive data and structural information.

Purpose of the Study:

• To develop a fast and accurate computational model for predicting protein optimal temperature, pH, and salt concentration.
• To leverage protein structures and sequence embeddings for improved prediction accuracy.
• To address limitations of previous studies regarding data scarcity and structural information.

Main Methods:

• Constructed a dataset of 175,905 non-redundant proteins.
• Developed GeoPoc, a novel model based on geometric graph learning.
• Utilized protein structures and sequence embeddings from a pre-trained language model.
• Employed a geometric graph transformer network to capture sequence and spatial information.

Main Results:

• Achieved a Pearson Correlation Coefficient (PCC) of 0.78 for optimal temperature prediction in in-house validation.
• Outperformed state-of-the-art methods by 2.3% in AUC on an independent test set for temperature prediction.
• Obtained AUC scores of 0.78 for pH and 0.77 for salt concentration prediction.
• Identified critical physicochemical properties contributing to protein thermostability through interpretable analysis.

Conclusions:

• GeoPoc provides a robust and accurate method for predicting optimal conditions of extremophilic proteins.
• The model's ability to integrate structural and sequence data enhances prediction performance.
• GeoPoc facilitates enzyme engineering and industrial applications by enabling rapid identification of suitable proteins.