Transfer Learning for Versatile Training Free High Content Screening Analyses

Transfer learning for versatile and training free high content screening analyses

Requirements

Python version : 3.10

Python libraries: pip install -r requirements.txt

This repository contains scripts used to transform and analyze HCS data in a fully automated manner. The two main scripts are featurization.py and normalisation.py

From image to feature using featurization.py

Usage : python featurization.py images.xlsx features.xlsx

This script consumes rows from an input file images.xlsx to get image paths and pass them through a ResNet as described in Transfer learning for versatile and training free high content screening analyses. It will save the output in a file named feature.xlsx.

Input.xlsx specification

An example can be found at example/1_input_deepmodule (toy).xlsx. It needs to be an Excel file with a sheet named Image_path which contains four columns named Barcode, Well, Content and Path. If optional columns Fields and Wave Length are present, features will be concatenated for each value of Wave Length and aggregated per values of Fields. Additional columns are ignored.

Note : One can modify the main part of the script to accept other file format (underlying class expect a Pandas DataFrame)

Feature normalization and analysis using analysis.py

Usage : python analysis.py features.xlsx parameters_file.json selected_hit.xlsx

This script takes as input an Excel file (example/2_data.xlsx) as first argument, it can be the output of the featurization.py script or of an handmade image analysis, a parameters file (example/2_parameters.json) as second argument and the name of the output file (such as "selected_hit.xlsx") as third argument.

features.xlsx specification

An Excel file with a sheet named All_Data.

Mandatory columns :

• Barcode = unique id per plate
• Plate = unique name per layout (compounds disposition)
• Well = plate localization, a letter for row (A-P) and a number for columns (1-24).
• Content = Name of the content
• Line = Name of the cell line
• Replicate = Replicate identifier
• At least one column of feature (will take every additional column where all values can be detected as numbers)

Other columns will be ignored.

file_parameters.json

A json file with some key info for the normalization process. Dictionary values are :

• ctrl_neg = a list of sample names to be considered as negative control.
• ctrl_pos = a list of sample names to be considered as positive control.
• features = a list of column names to apply the normalization process on. In case of features obtained by featurization.py, the special value '["deep learning feature"]' should be used.
• other_parms = a dict of parameters which can contain (among other things) the key reduced_feature_space with boolean value (True of False). If True, a PCA is launched on the feature space in order to keep 99% (default value) of the data variability.
• spatial_correction = name of the method in scripts.methods.CorrectionMethod to use as spatial correction method. (underscore can be replaced by space)
• sc_parms = dict to be passed as parameters to the spatial correction method (see list of parameters for the chosen function)
• normalization = name of the normalization method in scripts.methods.NormalisationMethod to be used. (underscore can be replaced by space)
• n_parms = dict to be passed as parameters to the normalization method (see list of parameters for the chosen function)
• selection = name of the method in scripts.methods.HitSelection to be used for hit selection. (underscore can be replaced by space)
• s_parms = dict to be passed as parameters to the hit selection method. It can contain two values :
  • "str_parms" = a list of rules to select hits. Each rule comprises 4 values :
    • 'include' = None, 'and' or 'or' indicate how to pile rules
    • 'feature' = name of the feature to be used for selection (some hit selection method can output new feature name, ex: "linear discriminant analysis" will output a feature named "LDA". See method documentation for details)
    • 'relative' = '<', '>' or '><' indicate the direction of the threshold ('><' means outside of [-|x|, |x|])
    • 'value' = threshold of the rule
  • "other_parms" = specific parameters for the chosen function in selection.