AssayMatch: Learning To Select Data for Molecular Activity Models.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Machine learning model performance in drug discovery relies heavily on training data quality.
• Aggregating bioactivity data from diverse sources like ChEMBL introduces noise due to experimental variability.
• Existing methods struggle with selecting compatible training data for novel drug candidates.

Purpose of the Study:

• To introduce AssayMatch, a novel framework for selecting homogeneous training datasets in drug discovery.
• To improve the predictive accuracy and data efficiency of machine learning models by reducing experimental noise.
• To enable data selection for test sets with unknown labels, mimicking real-world drug discovery scenarios.

Main Methods:

• AssayMatch utilizes data attribution methods to quantify assay contributions to model performance.
• It fine-tunes language embeddings of assay descriptions using attribution scores for semantic and compatibility matching.
• The framework ranks training data based on fine-tuned embeddings for effective selection.

Main Results:

• Models trained on AssayMatch-selected data outperformed models trained on complete datasets.
• AssayMatch effectively filters out noisy or incompatible experimental data.
• Improved prediction capabilities were observed across different machine learning architectures and model-target pairs.

Conclusions:

• AssayMatch offers a data-driven approach to curate high-quality datasets for drug discovery.
• The framework reduces noise from incompatible experiments, enhancing model predictive power and data efficiency.
• AssayMatch represents a significant advancement in optimizing machine learning applications for pharmaceutical research.