ProtSpace: A Tool for Visualizing Protein Space.

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Protein language models (pLMs) generate high-dimensional embeddings capturing evolutionary information from protein sequences.
• Interpreting these complex protein embeddings remains a significant challenge in bioinformatics.
• Understanding protein function from sequence-derived representations is crucial for biological research.

Purpose of the Study:

• To introduce ProtSpace, an open-source Python package for interactive visualization of protein embeddings.
• To facilitate the discovery of functional patterns by integrating embedding space with protein 3D structures.
• To demonstrate the utility of ProtSpace in analyzing biological datasets and challenging existing hypotheses.

Main Methods:

• Development of an interactive 2D and 3D visualization tool for protein embeddings.
• Integration of protein embedding visualization with 3D protein structure viewers.
• Application of ProtSpace to analyze phage and venom protein datasets.

Main Results:

• Phage dataset analysis revealed distinct functional clusters and a mixed region, potentially indicating biases in pLM training data.
• Venom protein analysis uncovered convergent evolution between scorpion and snake toxins, challenging current classifications.
• ProtSpace successfully identified functional patterns missed by traditional sequence analysis methods.

Conclusions:

• ProtSpace offers a powerful solution for interpreting protein language model embeddings.
• The tool aids in discovering novel functional insights and evolutionary relationships.
• Interactive visualization of embeddings alongside structural data enhances biological discovery.