Protein Binding Site Representation in Latent Space.

Area of Science:

• Computational chemistry
• Structural biology
• Artificial intelligence in drug discovery

Background:

• Deep learning models are increasingly used in computer-based drug discovery.
• Ensuring the interpretability and reliability of these models is crucial for their adoption.
• Understanding how these models perceive features, such as ligand binding sites, is key to building trust.

Purpose of the Study:

• To investigate the feature perception of a graph neural network (GNN) used for protein-ligand affinity prediction.
• To analyze the latent representation of ligand binding sites within the GNN.
• To explore the geometric structure of this latent space and its relationship to protein function.

Main Methods:

• Development of an automated computational pipeline for latent space analysis.
• Application of dimensionality reduction, clustering, hypothesis testing, and visualization techniques.
• Utilizing a graph neural network for protein-ligand complex affinity prediction.

Main Results:

• The learned latent space of protein binding sites is inherently structured, not random.
• Identified clusters in the latent space correspond to known functional protein families.
• Ligand size was identified as a significant factor influencing the geometry of these clusters.

Conclusions:

• The developed computational pipeline effectively enables analysis and interpretation of latent spaces in deep learning models.
• The findings demonstrate that GNNs learn meaningful representations of protein binding sites related to function.
• The methodology is adaptable for diverse datasets and deep learning architectures in drug discovery.