Boost-RS: boosted embeddings for recommender systems and its application to enzyme-substrate interaction prediction

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Enzyme promiscuity and substrate interactions are crucial for various biological and synthetic applications but remain under-explored.
• Current computational tools for predicting enzyme-substrate interactions are limited, hindering advancements in areas like drug metabolism and biomolecule synthesis.
• Recommender systems (RS) offer a promising avenue for predicting enzyme-substrate interactions, but their performance relies heavily on high-quality embedding vectors, which are challenging to enhance with auxiliary data.

Purpose of the Study:

• To develop a general recommender system (RS) framework, named Boost-RS, designed to improve the prediction of enzyme-substrate interactions.
• To enhance the performance of collaborative filtering (CF) based RS by effectively incorporating heterogeneous auxiliary data, particularly relational data.
• To validate the efficacy of Boost-RS in boosting embedding vectors for enzyme-substrate interaction prediction using contrastive learning tasks.

Main Methods:

• Proposed an innovative general RS framework, Boost-RS, that enhances embedding vectors by utilizing auxiliary data through multiple relevant auxiliary learning tasks.
• Employed contrastive learning tasks within Boost-RS to effectively exploit relational data for improving embedding quality.
• Applied the Boost-RS framework to several baseline collaborative filtering (CF) models for enzyme-substrate interaction prediction.

Main Results:

• Demonstrated that each auxiliary task within Boost-RS significantly boosts the learning of embedding vectors.
• Showcased that contrastive learning, as implemented in Boost-RS, outperforms traditional methods like attribute concatenation and multi-label learning.
• Confirmed that Boost-RS surpasses the performance of existing similarity-based models in predicting enzyme-substrate interactions.
• Ablation studies and visualization confirmed the critical role of contrastive learning on auxiliary data in enhancing embedding vectors.

Conclusions:

• Boost-RS provides an effective framework for enhancing recommender system performance in predicting enzyme-substrate interactions.
• The use of contrastive learning on auxiliary relational data is a key factor in boosting embedding vector quality and prediction accuracy.
• The developed Boost-RS framework offers a valuable computational tool for exploring the enzyme-substrate interaction space, with implications for various biological and synthetic applications.