Clustering · SMARTFlexDB

Redundancy-aware clustering (RNA-only)

The RNA-only apo–holo pairs are clustered in three complementary ways to support redundancy-controlled, leakage-aware machine-learning data splits. Assign all members of each cluster to the same data split; colour = cluster. geometry-derived

Sequence — one point = one whole host RNA chain's sequence (not the pocket's). MMseqs2 at 70% identity. It is single-chain on all 266 pairs. On 219 of them the chain is longer than that pair's own 10 Å pocket (8I7N chain N: 418 nt of chain against a 27 nt pocket).
Structure — one point = that same host RNA chain, as 3D coordinates. RMalign, RMscore ≥ 0.75.
Pocket — one point = one binding pocket: the RNA within a wider 10 Å shell of the ligand (broader than the 8 Å binding site shown on the pair pages). RMalign, RMscore ≥ 0.75. Unlike the other two, a pocket can span more than one chain — on 135 of the 266 pairs the 10 Å shell reaches a second chain. Those chains are aligned together as one rigid body.

Clusters (by size) ↓ assignments (CSV)

Each cluster is a redundancy group: its members are similar enough that putting them in different folds would leak information between those folds. It is a redundancy map for splitting, not a phylogeny.

Methods & parameters

Sequence clusters come from MMseqs2 cluster (createdb + cluster + createtsv) at 70 % identity (--min-seq-id 0.7 -c 0.8 --cov-mode 1 -s 7.5, nucleotide).

Structure and pocket clusters come from RMalign average-normalized RMscore with single-linkage clustering cut at 0.75.

A cluster is a connected component of the RMscore ≥ 0.75 graph, so no two similar chains/pockets (even transitively linked) can be split across folds.

Caveats & how to read the map

The distribution of sequence similarity is bimodal (near-identical families vs distinct singletons), so the exact identity threshold barely changes the clustering.

Many of the host RNA chains are short (median host-chain length ≈ 29 nt per pair; range 9–418 nt), so RMscore/identity are low-information below ~15–20 nt.

Singleton clusters should not necessarily be interpreted as biologically unique.

An & in a host-chain sequence (84 of 266 pairs) is DSSR's strand-break marker inside that one chain. The break itself is an unmodeled or renumbered gap, a backbone nick, or a HET residue of the same chain that DSSR reads as a nucleotide. It is stripped before clustering, and it never joins two chains.

The sequence map is laid out by k-mer composition, colored by MMseqs2 cluster.