Fork me on GitHub

Training feature classifiers with q2-feature-classifier

Note

This guide assumes you have installed QIIME 2 using one of the procedures in the install documents.

This tutorial will demonstrate how to train q2-feature-classifier for a particular dataset. We will train the Naive Bayes classifier using Greengenes reference sequences and classify the representative sequences from the Moving Pictures dataset.

Note that several pre-trained classifiers are provided in the QIIME 2 data resources. These can be used for some common marker-gene targets (e.g., 16S rRNA genes). Pre-trained classifiers for other marker genes are also available on the QIIME 2 forum.

We will download and create several files, so first create a working directory.

mkdir training-feature-classifiers
cd training-feature-classifiers

Obtaining and importing reference data sets

Two elements are required for training the classifier: the reference sequences and the corresponding taxonomic classifications. To reduce computation time for this tutorial we will use the relatively small Greengenes 13_8 85% OTU data set. Do not use the 85% OTU data set used in this tutorial for classification of real experimental data. We recommend using more information-rich sequences, e.g., reference sequences clustered at 99% sequence similarity, for classification of real data. See the QIIME 2 data resources page for links to complete QIIME-compatible reference datasets.

Note

All sequence IDs present in the reference sequences must also be present in the reference taxonomy. If using reference sequence databases that have been clustered into OTUs, make sure that the corresponding reference taxonomy is used. For example, if using the Greengenes 99% OTU sequences, the 99% OTU taxonomy must also be used.

We will also download the representative sequences from the Moving Pictures tutorial to test our classifier.

Please select a download option that is most appropriate for your environment: 
wget \
  -O "85_otus.fasta" \
  "https://data.qiime2.org/2024.2/tutorials/training-feature-classifiers/85_otus.fasta"
curl -sL \
  "https://data.qiime2.org/2024.2/tutorials/training-feature-classifiers/85_otus.fasta" > \
  "85_otus.fasta"
Please select a download option that is most appropriate for your environment: 
wget \
  -O "85_otu_taxonomy.txt" \
  "https://data.qiime2.org/2024.2/tutorials/training-feature-classifiers/85_otu_taxonomy.txt"
curl -sL \
  "https://data.qiime2.org/2024.2/tutorials/training-feature-classifiers/85_otu_taxonomy.txt" > \
  "85_otu_taxonomy.txt"
Please select a download option that is most appropriate for your environment: 
wget \
  -O "rep-seqs.qza" \
  "https://data.qiime2.org/2024.2/tutorials/training-feature-classifiers/rep-seqs.qza"
curl -sL \
  "https://data.qiime2.org/2024.2/tutorials/training-feature-classifiers/rep-seqs.qza" > \
  "rep-seqs.qza"

Next we import these data into QIIME 2 Artifacts. Since the Greengenes reference taxonomy file (85_otu_taxonomy.txt) is a tab-separated (TSV) file without a header, we must specify HeaderlessTSVTaxonomyFormat as the source format since the default source format requires a header.

qiime tools import \
  --type 'FeatureData[Sequence]' \
  --input-path 85_otus.fasta \
  --output-path 85_otus.qza

qiime tools import \
  --type 'FeatureData[Taxonomy]' \
  --input-format HeaderlessTSVTaxonomyFormat \
  --input-path 85_otu_taxonomy.txt \
  --output-path ref-taxonomy.qza

Output artifacts:

Extract reference reads

It has been shown that taxonomic classification accuracy of 16S rRNA gene sequences improves when a Naive Bayes classifier is trained on only the region of the target sequences that was sequenced (Werner et al., 2012). This may not necessarily generalize to other marker genes (see note on fungal ITS classification below). We know from the Moving Pictures tutorial that the sequence reads that we’re trying to classify are 120-base single-end reads that were amplified with the 515F/806R primer pair for 16S rRNA gene sequences. We optimize for that here by extracting reads from the reference database based on matches to this primer pair, and then slicing the result to 120 bases.

qiime feature-classifier extract-reads \
  --i-sequences 85_otus.qza \
  --p-f-primer GTGCCAGCMGCCGCGGTAA \
  --p-r-primer GGACTACHVGGGTWTCTAAT \
  --p-trunc-len 120 \
  --p-min-length 100 \
  --p-max-length 400 \
  --o-reads ref-seqs.qza

Output artifacts:

Note

The --p-trunc-len parameter should only be used to trim reference sequences if query sequences are trimmed to this same length or shorter. Paired-end sequences that successfully join will typically be variable in length. Single-end reads that are not truncated at a specific length may also be variable in length. For classification of paired-end reads and untrimmed single-end reads, we recommend training a classifier on sequences that have been extracted at the appropriate primer sites, but are not trimmed.

Note

The primer sequences used for extracting reads should be the actual DNA-binding (i.e., biological) sequence contained within a primer construct. It should NOT contain any non-biological, non-binding sequence, e.g., adapter, linker, or barcode sequences. If you are not sure what section of your primer sequences are actual DNA-binding, you should consult whoever constructed your sequencing library, your sequencing center, or the original source literature on these primers. If your primer sequences are > 30 nt long, they most likely contain some non-biological sequence.

Note

The example command above uses the min-length and max-length parameters to exclude simulated amplicons that are far outside of the anticipated length distribution using those primers. Such amplicons are likely non-target hits and should be excluded. If you adapt this command for your own use, be sure to select settings that are appropriate for the marker gene, not the settings used here. The min-length parameter is applied _after_ the trim-left and trunc-len parameters, and max-length _before_, so be sure to set appropriate settings to prevent valid sequences from being filtered out.

Train the classifier

We can now train a Naive Bayes classifier as follows, using the reference reads and taxonomy that we just created.

qiime feature-classifier fit-classifier-naive-bayes \
  --i-reference-reads ref-seqs.qza \
  --i-reference-taxonomy ref-taxonomy.qza \
  --o-classifier classifier.qza

Output artifacts:

Test the classifier

Finally, we verify that the classifier works by classifying the representative sequences from the Moving Pictures tutorial and visualizing the resulting taxonomic assignments.

qiime feature-classifier classify-sklearn \
  --i-classifier classifier.qza \
  --i-reads rep-seqs.qza \
  --o-classification taxonomy.qza

qiime metadata tabulate \
  --m-input-file taxonomy.qza \
  --o-visualization taxonomy.qzv

Output artifacts:

Output visualizations:

Classification of fungal ITS sequences

In our experience, fungal ITS classifiers trained on the UNITE reference database do NOT benefit from extracting/trimming reads to primer sites. We recommend training UNITE classifiers on the full reference sequences. Furthermore, we recommend the “developer” sequences (located within the QIIME-compatible release download) because the standard versions of the sequences have already been trimmed to the ITS region (excluding portions of flanking rRNA genes that may be present in amplicons generated with standard ITS primers).