The purpose of this analysis and data processing is to prepare the food interaction data provided by drugBank for automated graph generation using FCA. This would enable to show the food interaction data in two ways:
- text - in form of a table. Can show more detailed descriptions but connections are hard to find if there are too many drugs
- graph - with FCA. This should be an additional view option to the table format and would allow the user to see the links between selected drugs and their food interactions
This project focuses on preparing the data for the automated graph visualization. The code was written in
python and the jupyter notebook files containing the code, results, graphs and explanations are located in the
analysis folder. The following is a summary of the process and the results.
The data was taken from drugbank. I used an sql script to convert the following sql tables to csv: categories, food interactions, drugs and analysed the first two. They have the following attributes:
The table has 68860 entries with 6142 unique drug ids and 3606 unique categories. The following plot shows the most common categories from this dataset.
This dataset is the most significant one and is the main interest of this project. It contains 2388 food interactions with 498 unique food interaction descriptions.
After analysing the data (see the Food interactions - data analysis & cleaning file) I gathered the
following findings:
- multiple sentence descriptions, which can be split into separate descriptions
- redundant descriptions which can be removed (ex. 'No food effects.', 'No known food interactions.')
In this section, I split each description into separate sentences in the dataset and added them as new rows with the corresponding drug id. In the previous steps I also identified some useless data which I also eliminated (ex. 'No food effects.', 'No known food interactions.'). Finally, I exported the cleaned data to a new csv file.
For the sentence splitting I had to make sure to ignore abbreviations and other not relevant punctuation marks. I found a suitable python library which I could use with some adjustments. I added a new column to my dataset with the newly generated tokens (every token symbolizes one sentence), deleted the old description column and expanded the token list according to the linked drug ids.
This section evaluates the new, cleaned data retrieved from the previous step.
This plot shows the result of the sentence splitting. The initial data (blue) has a lower rate of short
sentences (measured in words), which is improved by the processed data (orange).
The above diagram shows the most frequent food related nouns which appear in food interaction descriptions.