OptiClust, an Improved Method for Assigning Amplicon-Based Sequence Data to Operational Taxonomic Units.
Westcott SL, Schloss PD. · 2017
Read OptiClust, an Improved Method for Assigning Amplicon-Based Sequence Data to Operational Taxonomic Units. by Westcott SL, Schloss PD. (2017) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc5343174.txt.
Introduction
Sarah L. Westcott and Patrick D. Schloss presented a new grouping method for microbe data in mSphere in 2017. Gene tag studies sort millions of reads into groups of kin, called operational taxonomic units, without leaning on a reference library. The sorting step shapes all later claims about which microbes live where. Weak sorting blurs the map of a community.
Their method, OptiClust, shuffles reads between groups and keeps the moves that raise a score of grouping quality. That score counts both right calls and misses, so the method does not trade one error for another. In tests, the groups held up better than those from older neighbor rules. The runs were fast and light on memory. The authors say the gain will let field studies rest on firmer groups as data sets keep growing.
Catalog record
| CREATOR | Westcott SL, Schloss PD. |
|---|---|
| YEAR | 2017 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
OptiClust, an Improved Method for Assigning Amplicon-Based Sequence Data to Operational Taxonomic Units By Sarah L Westcott, Patrick D Schloss, Katherine McMahon, Mick Watson, Katherine Pollard Published in: mSphere DOI: 10.1128/mSphereDirect.00073-17 ABSTRACT The analysis of microbial communities from diverse environments using 16S rRNA gene sequencing has expanded our knowledge of the biogeography of microorganisms. An important step in this analysis is the assignment of sequences into taxonomic groups based on their similarity to sequences in a database or based on their similarity to each other, irrespective of a database. In this study, we present a new algorithm for the latter approach. The algorithm, OptiClust, seeks to optimize a metric of assignment quality by shuffling sequences between taxonomic groups. We found that OptiClust produces more robust assignments and does so in a rapid and memory-efficient manner. This advance will allow for a more robust analysis of microbial communities and the factors that shape them. KEYWORDS: 16S rRNA gene, bioinformatics, microbial ecology, microbiome INTRODUCTION Amplicon-based sequencing has provided incredible insights into Earth’s microbial biodiversity (1, 2). It has become common for studies to include sequencing millions of 16S rRNA gene sequences across hundreds of samples (3, 4). This is a sequencing depth 3 to 4 orders of magnitude greater than was previously achieved using Sanger sequencing (5, 6). The increased sequencing depth has revealed novel taxonomic diversity that is not adequately represented in reference databases (1, 3). However, the advance has forced reengineering of methods to overcome the rate- and memory-limiting steps in computational pipelines that process raw sequences through the generation of tables containing the number of sequences in different taxa for each sample (7–10). A critical component of these pipelines has been the assignment of amplicon sequences to taxonomic units that are defined either based on similarity to a reference or operationally based on the similarity of the sequences to each other within the data set (11, 12). A growing number of algorithms have been developed to cluster sequences into operational taxonomic units (OTUs). These algorithms can be classified into three general categories. The first category of algorithms has been termed closed reference or phylotyping (13, 14). Sequences are compared to a reference collection and clustered based on the reference sequences to which they are similar. This approach is fast; however, the method struggles when a sequence is similar to multiple reference sequences that may have different taxonomies and when it is not similar to sequences in the reference (15). The second category of algorithms has been called de novo because they assign sequences to OTUs without the use of a reference (14). These include hierarchical algorithms such as nearest, furthest, and average neighbor (16) and algorithms that employ heuristics such as abundance- or distance-based greedy clustering (AGC or DGC, respectively) as implemented in USEARCH (17) or VSEARCH (18), Sumaclust and OTUCLUST (19), and Swarm (20). De novo methods are agglomerative and tend to be more computationally intense. It has proven difficult to know which method generates the best assignments. A third category of algorithm is open-reference clustering, which is a hybrid approach (3, 14). Here, sequences are assigned to OTUs using closed-reference clustering, and sequences that are not within a threshold of a reference sequence are then clustered using a de novo approach. This category blends the strengths and weaknesses of the other method and adds the complication that closed-reference and de novo clustering use different OTU definitions. These three categories of algorithms take different approaches to handling large data sets to minimize the time and memory requirements while attempting to assign sequences to meaningful OTUs. Several metrics have emerged for assessing the quality of OTU assignment algorithms. These have included the time and memory required to run the algorithm (3, 20–22), agreement between OTU assignments and the sequences’ taxonomy (20, 22–32), sensitivity of an algorithm to stochastic processes (33), the number of OTUs generated by the algorithm (23, 34), and the ability to regenerate the assignments made by other algorithms (3, 35). Unfortunately, these methods fail to directly quantify the quality of the OTU assignments. An algorithm may complete with minimal time and memory requirements or generate an idealized number of OTUs, but the composition of the OTUs could be incorrect. These metrics also tend to be subjective. For instance, a method may appear to recapitulate the taxonomy of a synthetic community with known taxonomic structure but do a poor job when applied to real communities with poorly defined taxonomic structure or for sequences that are prone to misclassification. As an alternative, we developed an approach to objectively benchmark the clustering quality of OTU assignments (13, 15, 36). This approach counts the number of true positives (TPs), true negatives (TNs), false positives (FPs), and false negatives (FNs) based on the pairwise distances. Sequence pairs that are within the user-specified threshold and are clustered together represent TPs, and those in different OTUs are FNs. Those sequence pairs that have a distance larger than the threshold and are not clustered in the same OTU are TNs, and those in the same OTU are FPs. These values can be synthesized into a single correlation coefficient, the Matthews correlation coefficient (MCC), which measures the correlation between observed and predicted classifications and is robust to cases where there is an uneven distribution across the confusion matrix (37). Consistently, the average neighbor algorithm was identified as among the best or as the best algorithm. Other hierarchical algorithms, such as furthest and nearest neighbor, which do not permit the formation of FPs or FNs, respectively, fared significantly worse. The distance-based greedy clustering as implemented in VSEARCH has also performed well. The computational resources required to complete the average neighbor algorithm can be significant for large data sets, and so there is a need for an algorithm that efficiently produces consistently high-quality OTU assignments. These benchmarking efforts have assessed the quality of the clusters after the completion of the algorithm. In the current study, we developed and benchmarked a new de novo clustering algorithm that uses real-time calculation of the MCC to direct the progress of the clustering. The result is the OptiClust algorithm, which produces significantly better sequence assignments while making efficient use of computational resources. RESULTS OptiClust algorithm. The OptiClust algorithm uses the pairs of sequences that are within a desired threshold of each other (e.g., 0.03), a list of all sequence names in the data set, and the metric that should be used to assess clustering quality. A detailed description of the algorithm is provided for a toy data set in the supplemental material. Briefly, the algorithm starts by placing each sequence either within its own OTU or into a single OTU. The algorithm proceeds by interrogating each sequence and recalculating the metric for the cases where the sequence stays in its current OTU, is moved to each of the other OTUs, or is moved into a new OTU. The location that results in the best clustering quality indicates whether the sequence should remain in its current OTU or be moved to a different or new OTU. Each iteration consists of interrogating every sequence in the data set. Although numerous options are available for optimizing the clusters and for assessing the quality of