FoldSeek diversity clustering#

Pass --cluster_with_foldseek and structures get clustered by structural similarity using FoldSeek, which adds a foldseek_cluster_id column so you can count structurally distinct designs per task.

Clustering only looks at the designed chain. In a multimer output such as binder plus target, only the most freely designed chain gets extracted and clustered. The pipeline identifies that chain from the conditioning annotations and stores it as chain_of_interest in the metrics CSV. Clustering reads the extracted per design chain CIFs, which only get written when you generate with save_chain_of_interest=true, so set that if you plan to cluster.

Point FOLDSEEK_EXEC in .env.local at the FoldSeek executable. Clustering runs foldseek easy-cluster on the extracted single chain with a TM-score threshold of 0.5.