A self-organizing map (SOM) projects high-dimensional data onto a low-dimensional grid while preserving neighborhood relationships. For chemical screening, it turns 2,048-dimensional Morgan fingerprints into a two-dimensional map that helps visualize similar compounds and potential clusters. section 7.1.2
A SOM training workflow:
- Initialize a rectangular or hexagonal grid whose nodes contain randomly initialized vectors matching the input dimensionality. section 7.1.2
- Select an input vector and compare it with every node using a distance metric such as Euclidean distance. section 7.1.2
- Choose the closest node as the best matching unit (BMU), then update the BMU and neighboring nodes toward the input. section 7.1.2
From colored beads to drug repurposing
The intuition uses colored beads encoded as RGB vectors: nearby colors map to nearby nodes. The same process maps molecular fingerprints. In drug repurposing, nodes receive the plurality activity label of their training compounds; unseen compounds assigned to active nodes become candidate hits for further study. section 7.1.2 section 7.1.3

