Multiscale topology-enabled structure-to-sequence transformer for protein-ligand interaction predictions.

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Pretrained natural language processing (NLP) models excel in many fields but struggle with biological data.
• Current NLP models for biology often overlook crucial 3D structural information due to their sequential nature.

Purpose of the Study:

• To develop a novel approach that integrates 3D structural data with NLP models for biological applications.
• To overcome the limitations of sequential architectures in representing complex biological structures.

Main Methods:

• Introduction of TopoFormer, a model combining NLP with a multiscale topology technique called persistent topological hyperdigraph Laplacian (PTHL).
• PTHL converts 3D protein-ligand complexes into NLP-admissible sequences of topological invariants and shapes across various spatial scales.

Main Results:

• TopoFormer demonstrates exemplary scoring accuracy in benchmark datasets.
• The model shows excellent performance in ranking, docking, and screening tasks for protein-ligand interactions.
• The approach effectively captures essential interactions across different spatial scales.

Conclusions:

• TopoFormer successfully bridges the gap between 3D structural biology and NLP.
• This method enables the conversion of general high-dimensional structured data into NLP-compatible sequences.
• Paves the way for broader applications of NLP in biological research and drug discovery.