Deep learning models for unbiased sequence-based PPI prediction plateau at an accuracy of 0.65.

Area of Science:

• Computational biology
• Bioinformatics
• Protein interaction analysis

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions.
• Accurate computational prediction of PPIs remains a significant challenge.
• Previous evaluation schemes and data leakage issues have obscured progress in sequence-based PPI prediction.

Purpose of the Study:

• To investigate the impact of protein embeddings on sequence-based PPI prediction.
• To evaluate the performance of different model architectures and embedding strategies.
• To determine if sequence-based models can implicitly learn contact maps for PPI prediction.

Main Methods:

• Utilized ESM-2 protein embeddings for sequence-based PPI prediction.
• Compared models with varying complexity, per-protein, and per-token embeddings.
• Assessed the influence of self-attention and cross-attention mechanisms.
• Analyzed the ability of models to learn contact maps as intermediate representations.

Main Results:

• ESM-2 embeddings significantly explain performance gains in PPI prediction, irrespective of model architecture.
• All tested sequence-based models plateaued at an accuracy of 0.65.
• Sequence-based models cannot implicitly learn a contact map.
• Performance gains are attributed to embeddings rather than inherent sequence-based learning capabilities.

Conclusions:

• Protein embeddings like ESM-2 are key drivers of recent performance improvements in sequence-based PPI prediction.
• Current sequence-based models have limitations and cannot implicitly learn contact maps.
• Structural information may be essential for achieving reliable and accurate PPI predictions.