foldseek createtsv on a clustering result outputs a tsv file where representatives are outliers, not centroids

[shiraz-shah]

Expected Behavior

For createtsv on a cluster db to return a clusters where representatives (column 1) are more representative of the full cluster (column 2).

This could be a centroid or at least a sequence with bidirectional hits to relatively many different cluster members.

Current Behavior

The createtsv tsv file often selects outliers as representatives.

Steps to Reproduce (for bugs)

After a successful foldseek cluster result, execute:
foldseek createtsv DB DB DB_C DB_C.tsv

Context

Current behaviour is problematic for two reasons:

• When clustering is used to reduce redundancy, the non-redundant set becomes a poor representative of the full set
• When outputing cluster MSAs with result2msa the msas are bad quality because of the outlier a3m reference
• Both of the above lead to excessive information-loss as a result of clustering, which can be avoided by selecting a more appropriate cluster representative.

[martin-steinegger]

We currently do lot provide this. Do you want to have a more central representative based on alignment score?

[shiraz-shah]

Yes, please. That would be a huge improvement!