A Small Step Toward Generalizability: Training a Machine Learning Scoring Function for Structure-Based Virtual...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Machine learning scoring functions aim to predict protein-small molecule binding affinity.
• Many current functions rely on dataset biases, limiting generalization to new targets.
• Understanding the physics of binding is crucial for accurate predictions.

Purpose of the Study:

• Develop a machine learning scoring function that avoids dataset biases.
• Validate the function's ability to learn physical binding interactions.
• Utilize input attribution to extract meaningful binding information for drug design.

Main Methods:

• Built a machine learning scoring function (PointVS) with rigorous data filtering to minimize bias.
• Evaluated performance on the Comparative Assessment of Scoring Functions (CASF-2016) benchmark.
• Applied input attribution to analyze feature importance and identify binding interactions.
• Used attribution insights for fragment elaboration in drug design.

Main Results:

• PointVS achieved performance comparable to leading methods on CASF-2016.
• Attribution analysis showed PointVS identified interactions correlating with a physics-based profiler.
• Extracted binding pharmacophores using attribution enabled improved docking scores via fragment elaboration.
• Demonstrated proof-of-concept for a deep learning method in structure-based molecule design.

Conclusions:

• The developed scoring function learns physical binding interactions beyond dataset biases.
• Input attribution is a valuable tool for interpreting machine learning models in drug discovery.
• This work presents a novel deep learning approach for structure-based molecule design and pharmacophore extraction.