Machine learning prediction of multiple distinct high-affinity chemotypes for α-synuclein fibrils

TL;DR

A machine learning model predicted new high-affinity ligands for α-synuclein aggregates, useful for PET imaging.

Contribution

A novel machine learning approach enabled ligand discovery with strong generalization from limited training data.

Findings

Five high-affinity binders were experimentally validated from a large compound library.

The model demonstrated robust generalization despite being trained on fewer than 300 binding measurements.

Scaffold-guided curation improved the efficiency of ligand selection for testing.

Abstract

To identify new ligands for positron emission tomography imaging of α-synuclein aggregates, we developed a machine learning model trained on <300 binding measurements. We used scaffold-guided curation to select a 30 compound prospective set from a 140-million-member library. Experimental validation yielded five high-affinity binders, showing robust generalization for ligand discovery.

Machine learning model predicts α-synuclein ligands.

Figures