bacpp 0.1.2

BacterialPloidyPredictor (BacPP): A Machine Learning Tool for Bacterial Polyploidy Prediction from Genome Assemblies

BPP: BacterialPloidyPredictor

BacterialPloidyPredictor is the first package that aims to predict the bacterial polyploidy utilizing the complete genome assemblies. It allows users to apply one of the three machine learning models we trained for binary prediction.

Installation

To maximize the use of this tool, installation of CheckM2 is recommended. Although this BPP is fully functional without CheckM2, the incorporation of CheckM2 gives another metrics for the prediction confidence. According to our result, genome assemblies below 88% completeness would result in significant shift in feature values extracted for subsequent prediction. To enable the incorporation of CheckM2, add --checkm2. (See below for details)

Prerequisites

Ensure you have following packages:

• Python 3.9+
• Numpy
• Pandas
• Matplotlib
• sklearn
• xgboost
• statsmodels

Option 1: Through GitHub repository

git clone https://github.com/Dias-Lab/BacterialPloidyPredictor-v1.git
conda env create -n bpp -f bpp.yml
conda activate bpp
./bpp -h

Option 2: through pip install

pip install bpp
bpp -h

---

Overview

• Global genome features extraction - 1.1 to 2.10

• Genome completeness estimation through CheckM2 - 1.3

• Binary monoploidy/polyploidy prediction - 2.11

• Confidence estimation through multiple visualizations and quantitative metrics - 2.12 to 3.6

Usage

If installed via pip, use bpp ...; if installed via git clone, use bpp.py ...

To pull out the instruction page for full list of arguments,

bpp -h

usage: bpp [-h] [--predict] [--out OUT] [--cpus CPUS] [--images] [--checkm2] [--model {knn,lg,xgb}] [--model-path MODEL_PATH] [--num-windows NUM_WINDOWS]
           [--no-interactions] [--id-col ID_COL] [--pred-input PRED_INPUT] [--pred-output PRED_OUTPUT]
           folder

BPP: Bactererial Ploidy Predictor to identify bacteiral polyploidy based on global genomic architecture.

positional arguments:
  folder                Input folder containing FASTA/FA/FNA files

options:
  -h, --help            show this help message and exit
  --predict             After feature extraction, run polyploidy prediction using a trained model.
  --out OUT             Output directory (default: <folder>/outputs)
  --cpus CPUS           Number of CPU cores (1=serial).
  --images              Generate GC/AT skew images into <output folder>/image
  --checkm2             Run CheckM2 and append completeness/contamination to predictions.csv.
  --model {knn,lg,xgb}  Model to use for prediction if --predict is set. Default: knn
  --model-path MODEL_PATH
                        Path to model file (defaults to ./models/kNNPC.json / ./models/MLG.json / ./models/XGBoost.json).
  --num-windows NUM_WINDOWS
                        Number of windows for extracting global genomic architecture (default: 4096)
  --no-interactions     Do not add interaction terms
  --id-col ID_COL       ID column name in the features CSV for prediction. Default: file
  --pred-input PRED_INPUT
                        Optional: features CSV to use for prediction (overrides --out).
  --pred-output PRED_OUTPUT
                        Optional: predictions CSV path (2 columns: ID, polyploidy_pred). Default: <features_csv_dir>/predictions.csv

Quick start

To serve as the test run of this tool, the following command can be copied directly and executed at package root directory to perform polyploidy prediction, visualizing global genomic architectures and prediction confidence-related statistics (PCA as an overview and quantile of pairwise Euclidean distance among in-group and between-group pairwise Euclidean distance distribution), as well as prediction result of CheckM2:

bpp ./example --images --predict --cpus 4 --checkm2

or:

./bpp.py ./example --images --predict --cpus 4 --checkm2

If CheckM2 is not installed, please run:

bpp ./example --images --predict --cpus 4

or:

./bpp.py ./example --images --predict --cpus 4

Note

If no output directory is given, bpp will automatically generate a new subdirectory under the input directory named outputs, where prediction.csv as the main prediction result as well as the extracted feasures extracted_features.csv are included. One or two subdirectories checkm2-result (optional) and images will be generated to store raw output of CheckM2, and the exported png and interactive 3-dimensional PCA plot in html format.

CheckM2 only accept one fasta suffix among fasta, fa, and fna. However, you don't have to worry about renaming the input fasta files. All fasta files with either fasta, fa, or fna suffix can be dumped into one single input directory. BPP will read all fasta files and make prediction and plot using all recognized fasta files.

By default, the whole chromosome is split into 4096 windows for calculattion of oligonucleotide skew values. The models are all trained based on this setting. We did not observe significant predictive power change through fine-tuning number of windows. There are three models available for polyploidy prediction: logistic regression, XGBoost, and kNN. By default, kNN was deployed due to the best performance and no signs of overfitting.

As mentioned above, installation of CheckM2 is recommended nevertheless BPP is fully functional without CheckM2. We simulatd partial genome assemblies ranging from 10% to 98% to assess the predictive power of BPP (See below) and see the Euclidean distance change compared to the coordinate generated uding complete genome assemblies on 3D PCA space. According to our result, genome assemblies/scaffold below 88% completeness would result in significant shift in feature values extracted for subsequent prediction, which could cause a false prediction using BPP. Usually the verified complete genome assemblies are recommended for most accurate prediction.

Outputs

Above is the illustration from a selection of outputs from examples using genomes from *E.coli*, *Synechocystis sp. PCC 6803*, and *Citribacter freundii* Chromosome of *C.freundii* is visualized in forms of global GC skew and AT skew in circular representation (a) and linear representation (b). In panel (a), the outer ring represents global GC skew with positive and negative skew values in orange and purple colors, respecively; the inner ring represents global AT skew with positive and negative skew values in olive and grey colors, respectively. In panel (b), cumulative GC skew (blue) and AT skew (orange) are illustrated in different color. Additionally, two more plots are provided to help users assess the confidence of the prediction. (c) an interactive 3-dimentional PCA plot is generated and saved as the .html file. Below is a screenshot of the illustration, where orange (group 1: monoploidy), green (group 2: polyploidy with similar genomic architecture with group 1), and blue (polyploidy with divergent architecture in contrast to group 1) represents three groups of reference bacterial species with known ploidy. Samples for prediction is included in the scatter plot as black points. The name of the sample will be appear upon hovering above the point. (d) Pairwise Euclidean distance for reference samples within the same group (black) and between two different groups (blue) are enumerated which forms two distributions. The Euclidean distance between every sample for prediction and its nearest reference sample is calculated and included on the histogram to help assess the confidence of the prediction.

file	polyploidy_pred	PED.confidence	completeness	contamination
Citrobacter_freundii_ATCC_8090.fasta	1	0.9767195925928703	100.0	0.08
Escherichia_coli.fna	0.725	0	1	100.0
Synechocystis_sp_PCC_6803.fasta	1	1	99.99	0.1

The main prediction and prediction confidence are saved in the main output file prediction.csv. Above is the example output when kNN (default) is selected as the model for prediction.