Clusters of similar RNAs for machine-learning data splits: put every member of a
cluster in the same split, to keep similar RNAs from landing in both training and test sets. Each
holo structure enters once, through its largest binding pocket, and is clustered three ways.
tool-derived
Sequence — the ligand’s host RNA chain, whole (not only the pocket).
MMseqs2, ≥ 70% identity.
Structure — the same chain in 3D. RMalign, RMscore ≥ 0.75.
Pocket — the RNA within 8 Å of the ligand, from any
chain.the binding site the pair page shows: RNA within
8 Å of the ligand. RMalign, RMscore ≥ 0.75.
Colour = cluster; grey = a one-point cluster. Click a point to open its holo
structure on the Database page (a Sequence point shared by several holo
structures selects its cluster instead), or a cluster in the list to highlight it and link each of its
holo structures. Cluster numbers are the CSV’s.
What is clustered, and the CSV
One binding pocket per holo structure: the one with the most binding-site residues (ties:
lower RMSD, then ID). Its host chain is the one on the Sequence and Structure maps.
No apo structure is used: every entry of a holo structure shares its clusters.
CSV, one row per holo structure: id (the binding pocket clustered),
holo_pdb, apo_pdbs (every apo structure the holo structure forms an
entry with, by PDB ID, separated by ;),
ligand (that pocket’s),
rna_type (Functional type), seq_cluster / struct_cluster /
pocket_cluster (empty when not on that map),
seq_apo_cluster / struct_apo_cluster /
pocket_apo_cluster (see Caveats).
Methods & parameters
Sequence: MMseqs2 cluster (Steinegger & Söding, 2017), nucleotide,
--min-seq-id 0.7 -c 0.8 --cov-mode 1 -s 7.5 --alignment-mode 3 --max-seqs 10000 --cluster-reassign.
Criterion: ≥ 70% identity to the cluster’s representative over ≥ 80% of the
member’s length (MMseqs2’s cascaded mode, used here, does not
strictly enforce it), met by every member. Input:
the modelled chain’s sequence from DSSR (Lu et al., 2015), chain breaks
(&) removed, modified nucleotides as their parent base or N.
Structure, Pocket: all-against-all RMalign (Zheng et al., 2019; RMscore normalised
by the mean length), single linkage cut at 0.75, so two items with RMscore ≥ 0.75 always share
a cluster. Sequence clusters (MMseqs2’s greedy set cover) carry no such guarantee.
Pocket: RNA nucleotides with a heavy atom within 8 Å of a
ligand heavy atom; a pocket over two or more chains is aligned as one rigid body.the pair page’s binding-site residues (RNA within 8 Å of the
ligand); a pocket over two or more chains is aligned as one rigid body.
Maps: t-SNE (van der Maaten & Hinton, 2008) of 1 − RMscore, or of 3-mer
composition for Sequence; perplexity 30, fixed seed. Only nearby points are meaningful.
Caveats
9I9W and 9IF0 are on the Sequence map
only: their chain IDs (AAA, BBB) do not fit
RMalign’s PDB input.
One apo structure can pair with holo structures from different clusters, so a split by holo cluster alone can put it on both sides. The CSV’s seq_apo_cluster / struct_apo_cluster /
pocket_apo_cluster merge the clusters an apo structure connects (numbered by the
smallest cluster merged), so a split on them keeps every apo structure on one side.
Short chains: the median host chain is 33 nt and 45 of 266 are under 20 nt; scores between
very short chains are less reliable.
One-point clusters are not necessarily biologically unique.