StoPred: Accurate Stoichiometry Prediction for Protein Complexes Using Protein Language Models and Graph Attention

Area of Science:

• Biochemistry
• Computational Biology
• Structural Biology

Background:

• Protein complexes are crucial for biological functions, but determining their subunit stoichiometry is experimentally challenging.
• Existing computational methods for stoichiometry prediction are limited, often requiring homologous structures or predefined assembly states.
• Current protein language model (pLM) approaches predict homomeric stoichiometry but fail with heteromeric complexes.

Purpose of the Study:

• To develop a novel computational method for accurate prediction of protein complex stoichiometry.
• To address the limitations of existing methods, particularly for heteromeric complexes.
• To integrate sequence and structural information with advanced deep learning techniques.

Main Methods:

• Developed StoPred, a method integrating pLM embeddings with a graph attention network.
• Modeled inter-subunit relationships to predict stoichiometry directly from sequence or structure features.
• Applied the method to both homomeric and heteromeric protein complexes.

Main Results:

• StoPred achieved higher accuracy and efficiency compared to template-based and deep learning methods on benchmark datasets.
• Demonstrated significant improvements in top-1 accuracy: up to 16% for homomeric and 41% for heteromeric complexes.
• StoPred is the first deep learning method capable of accurately predicting heteromeric complex stoichiometry.

Conclusions:

• StoPred offers a robust and accurate solution for predicting protein complex stoichiometry, including challenging heteromeric assemblies.
• The method advances computational approaches by leveraging pLM embeddings and graph attention networks.
• StoPred has broad implications for understanding protein function and biological pathways through accurate complex composition prediction.