Geometricus represents protein structures as shape-mers derived from moment invariants

Area of Science:

• Structural bioinformatics
• Computational biology
• Machine learning in proteomics

Background:

• Increasing availability of experimentally solved protein structures necessitates efficient methods for their analysis.
• Existing protein structure embedding methods often rely on time-consuming alignments or may not generalize well.
• Need for a universally applicable, fast, and accurate approach to represent protein structures for machine learning.

Purpose of the Study:

• To introduce Geometricus, a novel method for embedding protein structures into fixed-dimensional vectors.
• To provide a fast, accurate, and interpretable approach for protein structure representation.
• To demonstrate the utility of Geometricus across diverse bioinformatics tasks.

Main Methods:

• Utilizes 3D moment invariants to discretize protein structure fragments into shape-mers.
• Represents full protein structures as vectors of shape-mer counts.
• Applies the embedding approach to tasks such as structure similarity search, clustering, and classification.

Main Results:

• Geometricus provides a fast and accurate method for protein structure embedding.
• The approach is universally applicable across different protein superfamilies and families.
• Demonstrated effectiveness in unsupervised clustering, classification, and similarity search tasks.

Conclusions:

• Geometricus offers a robust and efficient solution for representing protein structures in a fixed-dimensional space.
• The method's interpretability and broad applicability make it valuable for various predictive tasks in proteomics.
• Geometricus facilitates advanced machine learning applications on protein structural data.