This repository contains code for emulating Quantitative Systems Pharmacology (QSP) model using machine learning methods.
This code was written and teste on python 3.10. To reproduce environment please install python 3.10 and install requirements using pip:
$ pip install -r requirements.txtThe list of direct project dependencies is stored in requirements.in file.
Generation of fata point require R interpreter and deSolve package.
We have prepared environment.yml file with both python and R dependencies.
Base on your preference for conda or mamba you can install all dependencies using one of the following commands.
To install all dependencies using conda please run:
$ conda env create --file=environment.yml -n qsp_emulationTo install all dependencies using mamba please run:
$ mamba env create --file=environment.yml -n qsp_emulationBecause of performance we suggest to use mamba instead of conda.
The environment_freeze.yml file contains mamba env export output to recreate environment in
a exactly the same version, but it also contains Linux specific packages.
The code is split to six parts. All notebooks are stored in notebooks folder.
01. Generating training dataset (1M patients) using QSP.ipynb- generates a training dataset of 1000000 patients using QSP model implemented in R language.02. Training ML models.ipynb- trains and evaluates Machine Learning (ML) model emulators on generated dataset.03. Hand-adjusting params and selecting best PCA value.ipynb- trains and evaluates multiple more variants ofMLPRegressoremulator, which was selected as best performing in notebook 2.04. ML Models learning-curves.ipynb- presents comparison of learning curves of best performing models trained in notebook 2.05. Best model evaluation.ipynb- visualizes the absolute and relative error of the best performing emulator06. Fitting population params.ipynb- shows the use case of the emulator, by fitting population params to match waterfall plots of virtual and clinical trials.