DIPS-Plus: The enhanced database of interacting protein structures for interface prediction.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Protein interface prediction (PIP) is crucial for understanding protein function and interactions.
• Existing datasets like the Database of Interacting Protein Structures (DIPS) provide structural information but lack rich features.
• Developing advanced machine learning models for PIP requires comprehensive and feature-rich datasets.

Purpose of the Study:

• To present DIPS-Plus, an expanded and feature-rich dataset for protein interface prediction.
• To provide researchers with a curated feature bank for training machine learning models.
• To demonstrate the efficacy of DIPS-Plus in improving state-of-the-art PIP methods.

Main Methods:

• Expansion of the DIPS dataset to create DIPS-Plus, comprising 42,112 protein complexes.
• Inclusion of diverse residue-level features: surface proximities, half-sphere amino acid compositions, and profile hidden Markov model (HMM)-based sequence features.
• Benchmarking of a state-of-the-art PIP model trained on DIPS-Plus against existing methods.

Main Results:

• DIPS-Plus offers a significantly richer feature set compared to the original DIPS dataset.
• Training a state-of-the-art model on DIPS-Plus achieved new state-of-the-art performance in protein interface prediction.
• The enhanced model surpassed previous leading models trained on less comprehensive feature encodings.

Conclusions:

• DIPS-Plus is a valuable, feature-rich resource for advancing machine learning-based protein interface prediction.
• The dataset facilitates the development of more accurate and robust PIP methods.
• This work highlights the importance of curated, feature-rich datasets in achieving SOTA results in bioinformatics.