Nearest Neighbor Gaussian Process for Quantitative Structure-Activity Relationships.

Area of Science:

• Computational Chemistry
• Machine Learning in Drug Discovery

Background:

• Gaussian process models are powerful but limited to small datasets.
• Quantitative Structure-Activity Relationship (QSAR) models are crucial in drug discovery.
• Large datasets in industrial drug discovery pose scalability challenges for traditional models.

Purpose of the Study:

• To adapt Gaussian process models for large-scale QSAR analysis in industrial drug discovery.
• To introduce a novel nearest neighbor Gaussian process (NNGP) model.
• To improve prediction efficiency and model updatability for growing datasets.

Main Methods:

• Incorporation of locality-sensitive hashing for efficient nearest neighbor searches.
• Development of a sub-linear time complexity prediction model.
• Utilizing a small number of hyperparameters for robustness against overfitting.

Main Results:

• The NNGP model demonstrates efficient prediction with sub-linear time complexity relative to sample size.
• The model is robust against overfitting and generalizes comparably to other QSAR models.
• NNGP provides principled and well-calibrated uncertainty estimates, similar to standard Gaussian processes.

Conclusions:

• The NNGP model significantly extends the applicability of Gaussian processes to large datasets in QSAR.
• This novel approach offers efficient computation, robust generalization, and reliable uncertainty quantification for drug discovery.
• The NNGP model presents a promising alternative to existing QSAR methods for industrial applications.