CSI: Contrastive data Stratification for Interaction prediction and its application to compound-protein interaction...

Area of Science:

• Computational Biology
• Machine Learning
• Bioinformatics

Background:

• Accurate prediction of interactions between biological entities (e.g., compound-protein) is crucial for drug discovery and synthetic biology.
• Current deep learning models often struggle to fully leverage the relational information inherent in interaction datasets.
• Exploiting multi-view representations of interacting objects can enhance model performance through contrastive learning.

Purpose of the Study:

• To develop a novel method, Contrastive Stratification for Interaction Prediction (CSI), for partitioning interaction datasets.
• To improve the learning of object representations by utilizing congruent and non-congruent data views via contrastive learning.
• To apply CSI to the compound-protein interaction prediction problem to accelerate drug discovery and related applications.

Main Methods:

• CSI stratifies (partitions) datasets by assigning a key and multiple views to each data point.
• Data partitions under a specific key form congruent views, enabling contrastive multiview coding.
• The method learns embeddings that maximize mutual information across these congruent views.

Main Results:

• CSI significantly improved average precision in compound-protein interaction prediction, with gains ranging from 13.7% to 39% when using compounds/sequences as keys.
• Further gains of 16.9% to 63% were observed when using reaction features as keys in enzymatic datasets.
• These results demonstrate the effectiveness of data stratification and contrastive learning for interaction prediction.

Conclusions:

• CSI offers a powerful approach to enhance interaction prediction by effectively leveraging multi-view data representations.
• The method shows substantial improvements over baseline models lacking data stratification and contrastive learning.
• CSI has the potential to expedite drug discovery, metabolic engineering, and synthetic biology applications.