Phylogenetic placement, the problem of placing a sequence into a precomputed phylogenetic “backbone” tree, is useful for constructing large trees, performing taxon identification of newly obtained sequences, and other applications. The most accurate current method, pplacer, performs the placement using maximum likelihood but fails frequently on backbone trees with 5000 sequences. We show a simple technique, pplacer-XR (pplacer-eXtra Range), that extends pplacer to large datasets. We show, using challenging large datasets, that pplacer-XR provides the accuracy of pplacer and the scalability to ultra-large datasets of a leading fast phylogenetic placmement method, APPLES. pplacer-XR is available in open source form on github.
Protein fossils, i.e. noncoding DNA descended from coding DNA, arise frequently from transposable elements (TEs), decayed genes, and viral integrations. They can reveal, and mislead about, evolutionary history and relationships. They have been detected by comparing DNA to protein sequences, but current methods are not optimized for this task. We describe a powerful DNA-protein homology search method. We use a 64×21 substitution matrix, which is fitted to sequence data, automatically learning the genetic code. We detect subtly homologous regions by considering alternative possible alignments between them, and calculate significance (probability of occurring by chance between random sequences). Our method detects TE protein fossils much more sensitively than blastx, and > 10× faster. Of the ~7 major categories of eukaryotic TE, three have not been found in mammals: we find two of them in the human genome, polinton and DIRS/Ngaro. This method increases our power to find ancient fossils, and perhaps to detect non-standard genetic codes. The alternative-alignments and significance paradigm is not specific to DNA-protein comparison, and could benefit homology search generally.
The Maximum Weight Trace (MWT) is an optimization problem for multiple sequence alignment that takes a set of sequences and weights on pairs of letters from different sequences and seeks a multiple sequence alignment that maximizes the sum of the weights for the pairs of letters that appear in the same column. MWT was introduced by Kececioglu in 1993, then proven to be NP-hard, and heuristics and exact solutions for MWT developed. Unfortunately none of the MWT methods are scalable to even moderate-sized datasets. Here we propose the MWT-AM problem (MWT for Alignment Merging), an extension of the MWT problem to be used in a divide-and-conquer setting, where we seek a merged alignment of a set of disjoint alignments that optimizes the MWT score. We present variations of GCM (the Graph Clustering Merger, originally developed for the MAGUS multiple sequence alignment method) that are specifically designed for MWT-AM. We show that the best of these variants, which we refer to as GCM-MWT, perform well for the MWT-AM criterion. We explore GCM-MWT in comparison to other methods for merging alignments, T-coffee and MAFFT–merge, and find that GCM-MWT produces more accurate merged alignments. GCM-MWT is available in open source form at https://github.com/vlasmirnov/MAGUS .
We consider the following problem: A researcher identified a small number of molecules with a certain property of interest and now wants to find further molecules sharing this property in a database. This can be described as learning molecular classes from small numbers of positive examples. In this work, we propose a method that is based on learning a graph grammar for the molecular class. We consider the type of graph grammars proposed by Althaus et al. [2], as it can be easily interpreted and allows relatively efficient queries. We identify rules that are frequently encountered in the positive examples and use these to construct a graph grammar. We then classify a molecule as being contained in the class if it matches the computed graph grammar. We analyzed our method on different known groups of molecules defined by structural properties and show that our method achieves low false-negative and low false-positive rates.
Group testing saves time and resources by testing each preassigned group instead of each individual, and one-stage group testing emerged as essential for cost-effectively controlling the current COVID-19 pandemic. Yet, the practical challenge of adjusting pooling designs based on infection rate has not been systematically addressed. In particular, there are both theoretical interests and practical motivation to analyze one-stage group testing at finite, practical problem sizes, rather than asymptotic ones, under noisy, rather than perfect tests, and when the number of positives is randomly distributed, rather than fixed. Here, we study noisy group testing under the probabilistic framework by modelling the infection vector as a random vector with Bernoulli entries. Our main contributions include a practical one-stage group testing protocol guided by maximizing pool entropy and a maximum-likelihood recovery algorithm under the probabilistic framework. Our findings highlight the implications of introducing randomness to the infection vectors – we find that the combinatorial structure of the pooling designs plays a less important role than the parameters such as pool size and redundancy.
