Figure 2.

SCIMM pipeline. To initialize the IMMs, we initially partition a subset of the sequences into k clusters with a previously published method such as CompostBin [39] or LikelyBin [40]. We train an IMM on each cluster, and then compute the likelihood that each sequence was generated by each IMM for all sequences and all IMMs. Next, we reassign each sequence to the cluster corresponding to the IMM which generated it with greatest likelihood. If > 0.1% of the sequences changed clusters, we repeat the process. Otherwise we consider the clusters to be stable and halt.