Skip to content

Latest commit

 

History

History
52 lines (36 loc) · 2.36 KB

prediction.md

File metadata and controls

52 lines (36 loc) · 2.36 KB

How to make predictions on a metagenome dataset

This tutorial assumes that fastq/fasta files have been converted to TFRecord.

To get a full list of options for DeepMicrobes.py:

DeepMicrobes.py --helpfull

The shell scripts used to make predictions with all tested DNNs in the paper can be found in pipelines. In these scripts we use the --model_name option to tell DeepMicrobes.py which DNN architecture we would like to use.
The scripts of models called by DeepMicrobes.py are indicated in square brackets below.

The final best DNN:

  • attention: Embed + LSTM + Attention (DeepMicrobes) [./models/embed_lstm_attention.py]

Other tested DNNs:

  • deep_cnn: ResNet-like CNN [./models/resnet_cnn.py]
  • cnn_lstm: CNN + LSTM [./models/cnn_lstm.py]
  • seq2species: Seq2species [./models/seq2species.py]
  • embed_pool: Embed + Pool [./models/embed_pool.py]
  • embed_cnn: Embed + CNN [./models/embed_cnn.py]
  • embed_lstm: Embed + LSTM [./models/embed_lstm.py]

Classifying reads using DeepMicrobes

To make prediction on a metagenome dataset (referred to as sample.tfrec) using DeepMicrobes :

predict_DeepMicrobes.sh -i sample.tfrec -b 8192 -l species -p 8 -m model_dir -o prefix

Arguments:

  • -i TFRecord input containing interleaved paired-end reads
  • -m Dictionary containing model weights (should match the taxonomic level)
  • -o Output prefix
  • -b (Optional) Batch size (a multiple of 4) (default: 8192)
  • -l (Optional) Taxonomic level, species/genus (should match the weights) (default: species)
  • -p (Optional) Number of parallel calls for input preparation (default: 8)

Note: The model classifies sequences faster using a larger batch size. We recommend users to try different values and select the largest batch size that fits into memory.

The script takes as input a TFRecord dataset and generates a tab-delimited output file containing predictions made on each pair of reads.

  • 1st column: category labels (integer)
  • 2nd column: confidence score (decimal)

The tab-delimited file can then be used to generate a species/genus profile (see next tutorial).