EquiPNAS

EquiPNAS predicts protein-nucleic acid binding sites by combining protein language model (pLM) embeddings with an E(3) equivariant deep graph neural network to identify protein-DNA and protein-RNA interaction sites.

Key Features:

• Protein Language Model Integration: EquiPNAS uses pLMs trained on large corpora of protein sequences to generate embeddings that capture sequence-derived information.
• E(3) Equivariant Deep Graph Neural Networks: The framework employs symmetry-aware deep graph learning that is invariant to three-dimensional rotations and translations to model spatial interactions.
• Performance Superiority: EquiPNAS outperforms state-of-the-art methods for predicting both protein-DNA and protein-RNA binding sites across multiple datasets.
• Validation on Experimental and AlphaFold2 Structures: Predictive performance has been validated on experimental inputs and AlphaFold2 predicted structures.
• Reduced Dependence on Evolutionary Information: An ablation study demonstrates that pLM embeddings substantially reduce reliance on evolutionary information while maintaining accuracy.

Scientific Applications:

• Binding site identification: Identification of protein-DNA and protein-RNA binding sites to support studies of protein–nucleic acid interactions.
• Drug discovery: Prioritization of nucleic-acid-interacting residues relevant to therapeutic target characterization.
• Genetic regulation studies: Analysis of interaction sites to inform mechanisms of genetic regulation and nucleic acid recognition.

Methodology:

EquiPNAS integrates pLM-derived embeddings with an E(3) equivariant graph neural network architecture; an ablation study was used to assess the contribution of evolutionary information.