Linear-Scaling Kernels for Protein Sequences and Small Molecules Outperform Deep Learning While Providing Uncertainty...

Area of Science:

• Machine Learning
• Computational Chemistry
• Bioinformatics

Background:

• Gaussian processes (GPs) offer uncertainty quantification and interpretability for regression.
• Traditional GPs face computational challenges and difficulties with sequence/graph data.

Purpose of the Study:

• To develop scalable Gaussian process regression methods for sequence and graph data.
• To introduce the xGPR Python library for efficient GP regression.

Main Methods:

• Introduced random feature-approximated kernels for linear scaling with data and input size.
• Developed an efficient algorithm for fitting GPs to large datasets using the xGPR library.
• Compared xGPR performance against deep learning models on 17 benchmarks.

Main Results:

• Achieved competitive accuracy compared to state-of-the-art deep learning models.
• Demonstrated well-calibrated uncertainty quantification and improved interpretability.
• Showcased xGPR's utility in active learning for automated protein engineering.

Conclusions:

• Scalable GP regression is viable for sequence and graph analysis.
• xGPR provides an efficient and accurate tool for machine learning tasks.
• GP regression with xGPR facilitates automated scientific discovery.