MetaTax: Meta-classification of taxonomy classification

MetaTax is a meta-classification algorithm of taxonomy classification.

The input to MetaTax is a file with reads names (metagenomics) and several taxonomic classification for those reads that come from different classification tools (such as Kraken, Clark, USEARCH, ...).

The output of MetaTax is a file where each read is given a classification according to MetaTax voting algorithm. This algorithm will try to classify reads at the lowest level possible, always using the majority voting criteria.

Details about MetaTax algorithm will be released soon. Tests with several datasets has shown that at the species level MetaTax achieves a higher sensitivity and F1 score compared to individual classification tools and to WEVOTE, an existing voting algorithm based on the lowest common ancestor (LCA).

If you want to test MetaTax, create a file with the following structure:

reads1_file_path (could be relative or absolute)
N               (number of classifiers)
classifier1_name
classifier1_output_file_path
...
classifierN_name
classifierN_output_file_path

... and so on, for any number of reads files

Currently MetaTax supports output from the following taxonomic classifiers:

• Clark
• Centrifuge
• Kraken
• OneCodex
• USEARCH/Blastn*

* Note for USEARCH/Blastn

It is important to note that for USEARCH/Blastn, the supported format is ``blast6out''. Furthermore, before calling MetaTax with an output from USEARCH/Blastn, the output must be formatted with the script process_usearch.py. It uses the best-match criteria to select the taxonID for each read from the USEARCH/Blastn output. You can configure the values used to detect the best hit by editing this script. In the very beginning it defines these values:

• maxE (defaul 1e-4): maximum E-value to consider a hit (i.e. hits with E-value above this limit will not be considered)
• minIdentity (default 80): min percent of identity to consider a hit
• minSizeFactor (default 0.5): min percent of aligment length relative to the mean reads size
• readSize: put here the mean length of your reads

When 2 hits for the same read are considered, we choose the one with the highest identity. If both have the same identity, then we choose the one with the lowest E-value. Please feel free to implement your own ``best-hit'' approach by modifying the script process_usearch.py.

Output format

MetaTax outputs several files for each input. The output files are:

• _file_nameclassified.tsv: tab-separated file with reads that were classified according to MetaTax. For each read it provides:
  • Read: sequence name
  • ToolsN: number of tools specified in the input file
  • TotalClassif: number of tools that classified the sequence
  • Votes: number of votes that supported the sequence classification
  • PercentVotes: percent of votes relative to ToolsN
  • FullLineage: complete lineage of the classification. Each node in the lineage has the format <rank|name|taxID>.
• _file_namedisagreement.tsv: tab-separated file with the reads for which an agreement was not possible, e.g. some tools consider it as a bacteria while other tools consider as virus, without a majority consensus.
• _file_namelow_weight.tsv: tab-separated file with reads that do not achieved a minimum weight to consider the classification as valid. This weight is automatically computed as follows: given a taxon i and a taxon j such that i is an ancestral of j, in the classification tree, then the weight of i supporting the classification of j is the value of the probability density function of a normal distribution with mean 0 and std 0.5, with x value equal to level(i) - level(j). The function level() returns the level of the node in the taxonomy tree.
• _file_nameNAs.tsv: tab-separated file with reads that were not classified.
• _file_namelineage.tsv: optional tab-separated file (activate with -lineage when calling MetaTax) with reads that were classified and their full lineage, without any statistics.
• _file_namelog.txt: optional text file (activate with -log) with plenty of details for each read analysed. This file can be really large and should be used for debug only.

Doubts or comments: diaztula@ime.usp.br