alfie

alignment free identification of eDNA

Alfie is an alignment-free, kingdom level taxonomic classifier for DNA barcode data. Alfie classifies sequences using a neural network which takes k-mer frequencies (default k = 4) as inputs and makes kingdom level classification predictions. At present, the program contains trained models for classification of cytochrome c oxidase I (COI) barcode sequences to the taxonomic level: kingdom. The program is effective at classifying sequences >200 base pairs in length, and no alignment information is needed.

Alfie can be deployed from the command line for rapid file-to-file classification of sequences. This is an effective means of separating contaminant sequences in a DNA metabarcoding or environmental DNA data set from sequences of interest.

For increased control, alfie can also be deployed as a module from within Python. The alfie package also contains functions that can aid a user in the training and application of a custom alignment-free classifier, which allows the program to be applied to different DNA barcodes (or genes), as a binary classifier, or on different taxonomic levels.

Installation

Alfie is a python3 program that depends on the python packages: numpy, tensorflow, scikit-learn, and pandas. If you do not have these installed, it is recommended that you install python and the required packages via anaconda.

To install alfie, download and unzip this repository. From the terminal, enter the repository and then run the following command:

python3 setup.py install

To check that the installation was successful, open a new terminal and run the following command. It should pop up the alfie help menu.

alfie -h

Usage

Command line interface

Alfie can be run as a stand alone command line interface, just specify an input .fasta or .fastq file using the -f flag and alfie will conduct classification. The output will be a folder named alfie_out. The folder will contain five files (names same as the input, with a prefix indicating the kingdom) that respectively contain the sequences classified as belonging to the kingdom indicated in the file prefix.

You can test this out using the example files shipped with alfie.

#from within the alfie folder
alfie -f alfie/data/example_data.fasta

# This will create a folder: alfie_out
# with the files: 
# animalia_example_data.fasta bacteria_example_data.fasta fungi_example_data.fasta plantae_example_data.fasta protista_example_data.fasta

For very large files (order of millions), the input sequence file may need to be processed in a batch fashion. This will run more slowly, but less sequences will be held in memory at once. The batch size (number of sequences) is specified with the -b flag. This flag isn't required, and should be used only if the program is crashing (finding the optimal value for your own machine will require some trial and error).

alfie -f alfie/data/example_data.fastq -b 100

By default, alignment free classification is performed using the default feature set (4mer frequencies) and the corresponding pre-trained neural network (trained on COI-5P sequence fragments of varying lengths). A user can pass an alternative neural network to make predictions with by using the -m flag. If this option is exercised and the model has not been trained on 4mers, then the -k flag must be used to ensure the proper set of kmer features are generated to match the neural network input structure (see the example notebook for more info on making and using custom neural networks with alfie).

#example using the 6mer model that ships with alfie, note the -k 6 option is required
alfie -f alfie/data/example_data.fastq -m alfie/data/dnn_model_6mers -k 6

The alfie package

For more control, the alfie package can be deployed from within Python. The package contains modules for: sequence classifion, fasta and fastq input/output, and helper functions to aid a user in training and deploying a customized alignment-free sequence classifier.

Deploying alife as a kingdom-level classifier from within Python is fairly simple.

The following example data is available through alfie:

from alfie import example_fasta

#peek at the data structure, it is a list of dictonaries
example_fasta[0]
{'name': 'seq1_plantae',
 'sequence': 'TTCTAGGAGCATGTATATCTATGCTAATCCGAATGGAATTAGCTCAACCAGGTAACCATTTGCTTTTAGGTAATCACCAAGTATACAATGTTTTAATTACAGCACATGCTTTTTTAATGATTTTTTTTATGGTAATGCCTGTAATGATTGGTGGTTTTGGTAATTGGTTAGTTCCTATTATGATAGGAAGTCCAGATATGGCTTTTCCTAGACTAAATAACATATCTTTTTGACTTCTTCCACCTTCTTTATGTTTACTTTTAGCTTCTTCAATGGTTGAAGTAGGTGTTGGAACAGGATGAACTGTTTATCCTCCCCTTAGTTCGATACAAAGTCATTCAGGCGGAGCTGTTGATTTAGCAATTTTTAGCTTACATTTATCTGGAGCTTCATCGATTTTAGGAGCTGTCAATTTTATTTCTACGATTCTAAATATGCGTAATCCTGGGCAAAGCATGTATCGAATGCCATTATTTGTTTGATCTATTTTTGTAACGGCA'}
# a fastq sequence will have 'strand' and 'quality' key,value pairs as well

The classify_records function can be used to classify a list of sequences.

#import classifier
from alfie.classify import classify_records

#classify the sequences in example_fasta
seq_records, predictions = classify_records(example_fasta)

predictions[:5]
#array([3, 1, 4, 0, 0])

The function returns two outputs: a list of sequences (seq_records) and an array of classifications (predictions) assigning the records to kingdoms. The returned classifications are numeric (0 == "animalia", 1 == "bacteria", 2 == "fungi", 3 == "plantae", 4 == "protista"). Kingdom classifications can be decoded to kingdom names using the decode_predictions function.

from alfie.classify import decode_predictions

kingdom_labels = decode_predictions(predictions)

kingdom_labels[:5]
#['plantae', 'bacteria', 'protista', 'animalia', 'animalia']

Advanced application and custom neural network construction

For a more detailed demonstration of the alfie package's functionality please consult the jupyter notebook included with this repository. The notebook covers sequence input/output and kingdom-level classification in more detail, and also provides examples of how to train and deploy a custom, alignment-free classifier with alfie. Custom classifiers can be implemented for any taxonomic level or DNA barcode - you can bring your own training data or subset a taxonomic group of interest from the dataset used to train alfie. All the functions demonstrated above can also be applied in a generic fashion to efficiently conduct custom classification.

Acknowledgements

This program is dedicated to my Dad's dog, Alfie (pictured in the README). He is a good boy.