COLLAPSE

COLLAPSE generates compressed latent representations of protein structural environments (Compressed Latents Learned from Aligned Protein Structural Environments) to identify and characterize structural sites critical for protein function.

Key Features:

• Deep Representation Learning: Employs deep learning on 3D atomic positions surrounding protein sites to capture detailed structure–function relationships.
• Self-supervised Learning via Evolutionary Relationships: Leverages evolutionary relationships between homologous proteins as a self-supervision signal to train without extensive labeled datasets.
• Transfer Learning Capabilities: Produces embeddings applied in transfer learning, demonstrating state-of-the-art performance on benchmarks including protein–protein interaction prediction, mutation stability prediction, and functional-site prediction from the Prosite database.
• Comprehensive Search and Annotation: Enables searching for similar structural sites across large protein datasets and annotating proteins using a curated database of known functional sites.
• Computational Efficiency and Interpretability: Generates tunable embeddings designed for computational efficiency and interpretability to support high-throughput analyses.

Scientific Applications:

• Functional Annotation of Proteins: Identifies and characterizes structural sites to aid elucidation of biological mechanisms and disease-related pathways.
• Drug Discovery and Development: Predicts functional and potential drug-binding sites to inform targeted therapy design.
• Protein Engineering and Stability Prediction: Facilitates protein design and assessment of mutation effects, including mutation stability prediction and protein–protein interaction tasks.

Methodology:

Generates compressed latent embeddings from aligned 3D atomic positions via deep representation learning, trains using self-supervision from evolutionary relationships among homologous proteins, and applies embeddings for transfer learning and for searching/annotating against a curated database of known functional sites.