With the advent of high-throughput assays, a large number of biological experiments can be carried out. Image-based profiling is among the most accessible and inexpensive technologies for this purpose. High-throughput image-based assays have earlier shown to be effective in characterizing unknown functions of genes and small molecules. CellProfiler is a popular and commonly used tool to process and quantify the data that is produced by these assays by making various measurements, or features, for each cell. However, there may be several sources of error in the quantification pipeline that is developed in CellProfiler. In this work, we examined different steps to improve profiles that are taken from CellProfiler to identify mechanisms of action of different drugs. These steps mainly consist of data cleaning, cell level outlier detection, toxic drug detection, and regressing out the cell area from all other features, as many of them are widely affected by the cell area. We also examined deep learning based methods to improve the CellProfiler output, and finally we suggest possible avenues for future research.
Prerequisites can be found in requirements.txt.
Dataset was from GIGADB. Each plate was downloaded separately and tables of cell, nuclei and cytoplasm merged for each plate and all features saved in a CSV file in this format: plate\download.py .
The normalization by DMSO features is in normalize.py .
After downloading all plates, each plate was normalized individually.
Metadata can be found in moa.txtand final_features.txt.
In outlierdetection.py outlier cells in each plate was found by histrogm-based technique and the metadata for distinguishing outliers from others saved in a CSV file for each plate.
Output of remove_toxic_drugs.py is a CSV file that sorts drugs by median of cell counts of the wells treated by them. In the profile creation pipeline, 5% of drugs with lowest cell counts were removed.
In simple_find_representation.py profiles were created without any data cleaning method. After the profile creation, aggregation on well, aggregation on broad sample and calculating similarity and odds ratio and plotting the results added.
In final_find_representation.py profiles were created with different data cleaning methods, such as Histrogram-based outlier detection, removing outlier and toxic drugs. Regressing out was another step in the profile creation too. Other steps for evaltion was exactly similar to simple_find_representation.py.
In train_test_split.py five mechanisms were chosen and fifty random drugs with these mechnisms were picked for train/test data.
After splitting and saving train, validation and test data, different models such as simple fully connected network with mixup technique (mixup.py), simple fully connected network with HSIC loss (hsic.py) and various versions of AutoEncoder(autoencoder.py) like denoising autoencoder trained on data.
When trained model saved, by using deep_representation.py, the profile for final evalution is the output of the model (The input of the model is the cleaned profile created in previous step).