eriveradelgado/cyclodextrin-qsarr

A QSAR built in R for predicting binding affinity with cyclodextrin

Cyclodextrin QSAR Comparison

This repository will detail the process of creating and comparing QSARs that predict binding affinity of small molecules to α-, β-, and γ-cyclodextrin.

Workflow

Overall, the workflow can be followed by running the R Markdown files in order. However, the "zeroth" file, 00-dwnld.Rmd, can be skipped. This is the least important file for reproducibility and the data generated is already loaded in the repository. Additionally, loading the file is likely to be difficult without an academic VPN or academic credentials to download the data.

Steps

00. Downloading experimental observations

File: 00-dwnld.Rmd

• Downloads data from Rekharsky and Inoue (1997), Suzuki (2001), and Singh et al (2015)
• Converted Singh data from constant of association to Gibbs free energy change
• Abbreviated as ri, suzuki, and singh
• Basic wrangling of data into data frames
• Cyclodextrin is categorized as "alpha", "beta", "gamma" rather than by Greek letter
• Raw data is stored in affinity/raw, derived data is stored in affinity/derived

01. Cleaning experimental observations

File: 01-clean.Rmd

• Removed special or unconventional characters
• Cleaned by solvent conditions
• Rekharsky and Inoue data required most cleaning as it comes from multiple compiled sources
  • Additional cleaning is necessary in order to pass the data to 02-sdf.Rmd
  • Typos in chemical names corrected

02. Downloading ligand structures as SDFs

File: 02-sdf.Rmd

• Structure data files (SDFs) downloaded into sdf/
  • Subdirectories for each data source
  • Additional subdirectory for compiled data from each source
• Queried Chemical Identifier Resolver from NCI https://cactus.nci.nih.gov/chemical/structure
• All observations successfully downloaded
• Observations compiled into single SDFs
• The directories of individual SDFs is not uploaded onto GitHub for convenience
  • Combined SDF file is backed up in GitHub

03. Calculating chemical descriptors using CDK for R

File: 03-cdk.Rmd

• Using package rcdk to obtain chemical descriptors
  • https://cran.r-project.org/web/packages/rcdk/vignettes/using-rcdk.html
• Calculates 281 descriptors for each molecule
• Does not calculate 3D descriptors
  • 3D descriptors may not be reliable due to lack of optimization

04. Other sources for chemical descriptors

File: 04-desc_external.Rmd

• Incorporation of descriptors from other sources
• PaDEL-Descriptor
  • http://yapcwsoft.com/dd/padeldescriptor/
• Online chemical database with modeling environment
  • https://ochem.eu
• Mordred
  • https://github.com/mordred-descriptor/mordred

05. Joining affinity data to descriptors

File: 05-join.Rmd

• Combining affinity data and the descriptors
• Requires addition of guest names for OCHEM and Mordred descriptors
• Requires renaming of the column Name in results from PaDEL

06. Creating external validation sets

File: 06-extval.Rmd

• 15% of the data is set aside for external validation
• Data for external validation in directory extval/
• Data for model building and training in directory trn/

07. Preprocessing chemical descriptors

File: 07-preprocess.Rmd

• Common preprocessing functions applied to data
  • Centering and scaling
  • Removal of variables with near zero variance
  • Removal of molecules with large amounts of missing data
• If more than 5% of the chemical descriptors of a molecule are NA or NaN, the molecule is removed from the data. Otherwise, the descriptors are filled in using a simple mean.
• These steps are performed on the data saved in trn
• Removal of X-outliers
  • Use of algorithm discussed here: https://doi.org/10.1016/j.chemolab.2015.04.013
  • Assumptions of standard deviation of descriptors

08. Random forest proof of concept

File: 08-rf_poc.Rmd

• Building of a random forest
• Model built on all variables as well as variables selected by caret::rfe
• Tuning with new functions and existing caret::train functions
• Analyzing model performance over repetitions
• Decision that building models every time with RFE is too time intensive, leading to step 09
• Models saved in model/rf_pov/

09. Feature selection

File: 09-rfe.Rmd

• Use of caret::rfe to select features
• Running a single iteration with random forest functions only
  • 10 repeats of 10-fold cross-validation
• Requires multiple core processing to cut down on time
• Top 16 variables saved
• "rfe" objects saved in rfe/
• Variables saved in rfe_var/

10. Random forest

File: 10-rf.Rmd

• Random forest model built
• Tuning for two sets of variables: full variable set and selected variables
• Results from 10 repeats of 10-fold cross-validation

Notes on workflow

Because the data sources already completed screening of molecule names, there was no cleaning step to verify uniqueness of all ligands. In the future, this would be an appropriate step to take in order to curate the data.

Acknowledgements

This research was performed at Horst von Recum's lab at Case Western Reserve University under the mentorship of Edgardo Rivera-Delgado.