BATCH-SCAMPP: Scaling Phylogenetic Placement Methods to Place Many Sequences (Abstract)

Phylogenetic placement is the problem of placing one or more query sequences into a phylogenetic "backbone" tree, which may be a maximum likelihood tree on a multiple sequence alignment for a single gene, a taxonomy with leaves labeled by sequences for a single gene [Nidhi Shah et al., 2021], or a species tree [Jiang et al., 2023]. When the backbone tree is a tree estimated on a single gene, the most accurate techniques for phylogenetic placement are likelihood-based, and can be computationally intensive when the backbone trees are large [Chu and Warnow, 2023]. Phylogenetic placement into gene trees occurs when updating existing gene trees with newly observed sequences, but can also be applied in the "bulk" context, where many sequences are placed at the same time into the backbone tree. For example, phylogenetic placement can be used to taxonomically characterize shotgun sequencing reads generated for an environmental sample in metagenomic analysis [Nidhi Shah et al., 2021; Barbera et al., 2019]. The two most well known maximum likelihood phylogenetic placement methods are pplacer [Chu and Warnow, 2023] and EPA-ng [Barbera et al., 2019]. Of these two, EPA-ng is optimized for scaling the number of query sequences and is capable of placing millions of sequences into phylogenetic trees of up to a few thousand sequences [Barbera et al., 2019], and achieves sublinear runtime in the number of query sequences (see Figure 2 from [Balaban et al., 2022]). Previously we introduced the SCAMPP framework [Wedell et al., 2022] to enable both pplacer and EPA-ng to perform phylogenetic placement into ultra-large backbone trees, and we demonstrated its utility for placing into backbone trees with up to 200,000 sequences. By using maximum likelihood methods pplacer or EPA-ng within the SCAMPP framework, the resulting placements are more accurate than with APPLES-2 [Balaban et al., 2022], with the most notable accuracy improvement for fragmentary sequences, and are computationally similar for single query sequence placement [Wedell et al., 2022]. However, SCAMPP was designed to incrementally update a large tree, one query sequence at a time, and was not optimized for the other uses of phylogenetic placement, where batch placement of many sequencing reads is required. Here we introduce BATCH-SCAMPP, a technique that improves scalability in both dimensions: the number of query sequences being placed into the backbone tree and the size of the backbone tree. Furthermore, BATCH-SCAMPP is specifically designed to improve EPA-ng’s scalability to large backbone trees. Although BATCH-SCAMPP is based on SCAMPP, it uses a substantially modified design in order to be able to take advantage of EPA-ng’s ability to place many query sequences efficiently. The BATCH-SCAMPP method operates by allowing the input set of query sequences to suggest and then vote on placement subtrees, thus enabling many query sequences to select the same placement subtree. We pair BATCH-SCAMPP with EPA-ng to explore the capability of this approach for scaling to many query sequences. We show that this combination of techniques (which we call BSCAMPP+EPA-ng, or BSCAMPP(e)) not only provides high accuracy and scalability to large backbone trees, matching that of SCAMPP used with EPA-ng (i.e., SCAMPP(e)), but also achieves the goal of scaling sublinearly in the number of query sequences. Moreover, it is much more scalable than EPA-ng and faster than SCAMPP+EPA-ng: when placing 10,000 sequences into a backbone tree of 50,000 leaves, EPA-ng is unable to run due to memory issues, SCAMPP+EPA-ng requires 1421 minutes, and BSCAMPP(e) places all sequences in 7 minutes (all given the same computational resources. Figure 1 gives an example of this performance advantage on the nt78 [Chu and Warnow, 2023] simulated dataset.