Schematic overview of the clustering procedures We start with a single linkage tree constructed from pairwise distances. Each leaf in the tree corresponds to a protein sequence. Superfamilies are determined based on the internal structure of the tree. For each superfamily a distinct superfamily distance graph is built. This weighted graph is cut at weak connections into subclusters.

Excerpt from the single linkage tree The superfamily of sequence O93431 is determined as follows (traversing the tree along the branches depicted as bold lines). The first internal node connects this sequence with the four sequences P52794, P20827, P52793, and P97553 at an E-value of 1e-52. Thus, the ratio of the size of the merging subtree and the size of the current subtree at this point is 4/1. Stepping up the hierarchy, the next node (E-value 4e-38) connects these five sequences with a subtree consisting of 13 sequences, resulting in a ratio of 13/5 (= 2.6). Stepping further up the hierarchy, the following ratios are 1/18 (= 0.056 at E-value 6e-38), 2/19 (= 0.105 at E-value 2e-37), 15/21 (= 0.714 at E-value 2e-13), 1/36 (= 0.028 at E-value 5e-10), 211 975/37 (= 5729.054 at E-value 0.022), 259/212 012 (= 0.001 at E-value 0.023), etc. Taking the maximum of the ratios we find the superfamily root at E-value 5e-10 as the last node before the largest relative increase (depicted as a bullet in the tree). The superfamily of sequence O93431 hence consists of the 37 sequences belonging to the ephrin type A and type B families plus a few predicted proteins.

Multi domain proteins Sequences with different domain compositions belong to the same family of Adenylate cyclases but form different "true" clusters (Pfam domains: RA: Ras association (RalGDS/AF-6) domain; LRR: Leucine Rich Repeats; PP2C: Protein phosphatase 2C; guanylate_cyc: Adenylate and Guanylate cyclase catalytic domain)

The SYSTERS algorithms

The superfamily distance graph of the ephrin superfamily The graph contains only those edges which represent E-values of at least the superfamily cutoff 5e-10. The width of an edge is according to its E-value, here ranging from 5e-10 (thinnest edge) to 3e-149 (thickest edge). The subclustering procedure first splits off nodes from the bottom right of the graph as single sequence clusters. These sequences are predicted proteins which are not yet confirmed as functioning by any experiment. The last accepted split in the graph results in the partitioning into the two major groups of ephrin type A (left) and type B (right) sequences as shown by the dashed line. Single sequence clusters are added to the ephrin type B family in the subsequent singleton adoption step.