palazzilorenzo/synthetic-data-for-antimicrobial-resistance

Conditional Variational Auto-Encoder for generation of synthetic data for Antimicrobial Resistance.

Author	Project
L. Palazzi	Synthetic data for antimicrobial resistance

synthetic-data-for-antimicrobial-resistance

This package allows to generate synthetic tabular data for AMR using Variational AutoEncoders and Conditional Variational AutoEncoders. The modules allow to build and train the models, generate new data starting from real ones, and test their quality and validity through the SDMetrics package.

1. Overview
2. Contents
3. Prerequisites
4. Installation
5. Usage
6. Graphics
7. Author
8. Citation

Overview

Antimicrobial Resistance (AMR) occurs when microbes like bacteria, viruses, and fungi develop the ability to withstand antimicrobial treatments (drugs used to treat infections), increasing the risk of disease spread, severe illness and death.
Matrix-assisted laser desorption/ionization–time of flight (MALDI-TOF) mass spectrometry can be used to rapidly identify microbial species [2]. A recent work [4] has shown that machine learning-based method may be applied to MALDI-TOF spectra of bacteria to test antimicrobial susceptibility (or resistance), thus allowing more suitable treatment decisions in a clinical context. Since real data are often not enough to properly train a machine learning model (and thus to make correct predictions), the use of synthetic data is becoming increasingly important.

This work will provide a model based on Variational AutoEncoders (VAEs) and Conditional AutoEncoders (CVAEs) to generate new synthetic sets of data, and will also test their goodness w.r.t. the original dataset using consolidated metrics.

Keywords: synthetic data, mass spectrometry, MALDI-TOF, bacteria, bacteriology, antimicrobial resistance, AMR.

References

[1] Bielow C, Aiche S, Andreotti S, Reinert K. MSSimulator: Simulation of Mass Spectrometry Data. Journal of Proteome Research 2011 10 (7), 2922-2929. doi: 10.1021/pr200155f.

[2] Seng P, Drancourt M, Gouriet F, La Scola B, Fournier PE, Rolain JM, Raoult D. Ongoing revolution in bacteriology: routine identification of bacteria by matrix-assisted laser desorption ionization time-of-flight mass spectrometry. Clin Infect Dis. 2009 Aug 15;49(4):543-51. doi: 10.1086/600885. PMID: 19583519.

[3] Sohn K, Lee H, Yan X. Learning Structured Output Representation using Deep Conditional Generative Models. Advances in Neural Information Processing Systems 28 (NIPS 2015). https://proceedings.neurips.cc/paper_files/paper/2015/file/8d55a249e6baa5c06772297520da2051-Paper.pdf.

[4] Weis C, Cuénod A, Rieck B et al. Direct antimicrobial resistance prediction from clinical MALDI-TOF mass spectra using machine learning. Nat Med 28, 164–174 (2022). https://doi.org/10.1038/s41591-021-01619-9.

Contents

synthetic-data-for-antimicrobial-resistance is composed of three main modules that allow to generate new synthetic data:

• train_model.py allow to build and train one or more models based on Conditional Variational AutoEncoders model.
• generate_data.py allow to generate new set of synthetic data using an existing model.
• test_metrics.py allow to test the quality and validity of new data w.r.t. the real ones.

Some modules are also contained in docs folder, and contain classes and methods necessary to run the scripts.

N.B. With VAE model you can generate Escherichia coli spectra susceptible to Meropenem antibiotic. With Conditional VAE model you can generate Escherichia coli spectra, both resistant and susceptible to Ampicillin antibiotic.

An example is also provided in the extras folder.

For a better description of each module:

Module	Description
class_cvae	define class for CVAE model
class_vae	define class for VAE model
create_cvae	contains the architecture of the CVAE model, i.e. encoder and decoder architecture
create_vae	contains the architecture of the VAE model, i.e. encoder and decoder architecture
sampling	contains method to sample z, the vector encoding a digit.
utils	this module contains useful methods to access data, load models, train them and run predictions
train_model	allow to build and train the model (VAE or CVAE), saving wheights and loss history
generate_data	allow to generate new set of data, using trained models
test_metrics	allow to test the quality and validity of synthetic data w.r.t. real data

Prerequisites

Supported python:

First of all ensure to have the right python version installed.

This project uses Tensorflow, Keras, Numpy: see the requirements for more information.

Installation

First, clone the git repository and change directory:

git clone https://github.com/palazzilorenzo/synthetic-data-for-antimicrobial-resistance.git
cd synthetic-data-for-antimicrobial-resistance

Then, pip-install the requirements:

pip install -r requirements.txt

⚠️ Apple Silicon: ensure to have installed a proper Conda env, then install tensorflow as explained here. Finally, install one by one the remaining dependencies using conda (if available) or pip.

Usage

Once you have installed all the requirements, you can start by training your model/s.

Step 1:

Create your model and start the training by running:

python train_model.py model_dim

where

• model is the name of the model you want to use, i.e. 'cvae' or 'vae' (required) ;
• dim is the dimension of the latent space, i.e. the space in which data are represented after encoding process (required).

⚠️ model and dim must be separated by the '_' character.

Example:

python train_model.py cvae_64

or

python train_model.py vae_64

You can train more than one model in the same execution by separate them by one space.

python train_model.py cvae_32 cvae_64

or

python train_model.py vae_64 cvae_64

Step 2:

Now that you have built and trained your model/s, you can generate new set of synthetic data.

python generate_data.py model_dim

Same rules as before apply to model_dim.

Synthetic data are saved as prediction_model_dim format, with the addition of susc or res labels at the end for respectively only susceptible and only resistant spectra.

Step 3:

You can now test the quality and validity of your new data w.r.t. the real ones by running:

python test_metrics.py prediction_model_dim

where prediction_model_dim is the name of the file containing the synthetic data you want to test (required).

⚠️ prediction_model_dim must not contain the file extension.

Examle:

python test_metrics.py prediction_cvae_64

or

python test_metrics.py prediction_cvae_64_susc

test_metrics.py uses SDMetrics package to generate two types of report:

• Quality Report, that captures the Column Shapes and Column Pair Trends properties.
• Diagnostic Report, that captures the Coverage and Boundary properties.

Graphics

With graphics.py module it is possible to visualize different properties, such as loss history, synthetic vs real data and others.

Method	Description
Show_history_loss	Shows the history loss saved during training for a certain model
Show_synth_vs_real	Shows synthetic vs real spectrum (the first of each set)
Column_similarity	Shows the overall distribution of real and synthetic data
Column_shapes	Shows the shape score
Column_pair_trends	Shows the correlation map between real and sinthetic data
Coverage	Shows the coverage score

How to use:

• Show_history_loss

python graphics.py Show_history_loss model_dim

Ex.

python graphics.py Show_history_loss cvae_64

---

• Show_synth_vs_real

python graphics.py Show_synth_vs_real prediction_model_dim

Ex.

python graphics.py Show_synth_vs_real prediction_cvae_64

---

• Column_similarity

python graphics.py Column_similarity prediction_model_dim column

where column is optional argument and represents the column for which to show the similarity (default column=4601). Ex.

python graphics.py Column_similarity prediction_cvae_64 4631

⚠️ column ranges from 4400 to 5600 with step of 3 (4400, 4403, 4406 ...). Be sure to chose an existing column.

---

• Column_shapes

python graphics.py Column_shapes model_dim

Ex.

python graphics.py Column_shapes cvae_64

---

• Column_pair_trends

python graphics.py Column_pair_trends model_dim

Ex.

python graphics.py Column_pair_trends cvae_64

---

• Coverage

python graphics.py Coverage model_dim

Ex.

python graphics.py Coverage cvae_64

For a full example on how to use this package please refer to example_cvae_64.

Author

• Lorenzo Palazzi git

Citation

If you have found synthetic-data-for-antimicrobial-resistance helpful in your research, please consider citing this project

@misc{Synthetic data for Antimicrobial Resistance,
  author = {Palazzi, Lorenzo},
  title = {Synthetic data for Antimicrobial Resistance},
  year = {2024},
  publisher = {GitHub},
  howpublished = {\url{https://github.com/palazzilorenzo/synthetic-data-for-antimicrobial-resistance}},
}