cheMLearning library provides of high-level, user-friendly tools to perform common chemometrics analysis in untargeted metabolomics. Modelling is done using scikit-learn framework, for machine learning; while visualizations are mainly done with Plotly.

Main modules are:

• unsupervised_modelling
• supervised_modelling
• utils

Installation

pip install cheMLearning

within a Jupyter Notebook cell

!pip install cheMLearning

Tutorial

Unsupervised modelling

Unsupervised class works around Principal Component Analysis, facilitating the visualization of score and loading plots. Moreover, Hierarchical Clustering Analysis can be called on the principal components and visualize a score plot where the color code is given by the clusters of the HCA.

Visualizations can be obtained without having to manipulate scores or loadings. However, if one would like to access to these outputs, they can be accessed via class modules.

Let's say we have a data frame (df) containing 3 groups from a toxicology assay and 2 groups of QCs. The number of features is 3255.

Import

from cheMLearning.unsupervised_modelling import unsupervised_class

df = pd.read_csv("path_to_your_df")

# create an object which will contain all your PCA information
output = unsupervised_class()

# now call PCA_ready method to compute scores and loadings. 
output.PCA_ready(df)

Now we are ready to visualize a PCA-score plot

output.score_viz(pcx=0, pcy=1)

Now we have our score plot, however, all data points are the same color which makes it hard to interpret. We can try adding some information to the plot, such as what to use to color the data, the color sequence we would like to use and even symbols, to help people who might have hard problems differentiating colors.

color = ["[C] 0.0", "[C] 0.0","[C] 0.0","[C] 0.0", "TNF-a","TNF-a","TNF-a","TNF-a","[C] 10.0", "[C] 10.0",
"[C] 10.0","[C] 10.0"]
color_seq=["mediumblue","darkorange", "red"]
symbol_seq=["circle","diamond", "x"]

output.score_viz(pcx=0, pcy=1, color=color, color_discrete_sequence=color_seq, symbol=color, symbol_sequence=symbol_seq)

Furthermore, we can visualize our loadings in a scatter plot, a.k.a loading plot, which can bring more information to our analysis.

output.loading_contourplot(pcx=0, pcy=1)

To save your plots, just assign the image to a variable and then use the to_image() method from plotly.

loading_plot=output.loading_contourplot(pcx=0, pcy=1)
loading_plot.to_image("path_to_your_folder/loading_plot.png")

One might suspect that the data misses normalization. cheMLearining has a function implementing median-based PQN normalization. This function returns your df normalized and also writes it for you.

Make sure your data contains both QCs and dQCs, otherwise it will give you problems.

output.PQN("path_to_your_df") #it must be in separated by commas ",".

#you can save the output by assign it to a variable
df_pqn = output.PQN("path_to_your_df")

Hierarchical Clustering Analysis can be now performed and the dendrogram visualized. Let's say we got rid of the QCs.

output.HCA_dendrogram(label=df_pqn.iloc[0:12,].index);

Supervised modelling

Supervised class includes Partial Least Squares and Random Forest which can be used for either regression or classification tasks. PLS is computed with the NIPALS algorithm.

In the PLS modelling, tools to compute goodness metrics, i.e. R² and Q², VIPs, are provided. In addition to a static functions to ease feature extraction VIP for a given VIP threshold.

For RF modelling, it can be computed the distance matrix for a RF-classifier model. It allows tracking where the observations land in the forest.

Analytical tools for classifiers are also provided by facilitating a classification report.

Import

from cheMLearning.unsupervised_modelling import supervised_class

Utils

Here, tools to create confusion matrices and ROC curves are provided. In case one is working with MetaboRank (MR) tool, a function to read MR's output is provided.

Import

from cheMLearning import utils

Change Log

0.0.1 (09/11/2022)

• First Release