kudkudak / compound-activity-prediction

Code for TFML 2015 conference paper

Code for TFML 2015 paper

Supporting code for "Analysis of compounds activity concept learned by SVM using robust Jaccard based low-dimensional embedding" paper.

The paper deals with analysis of compound activity predictions. It argues that concepts learned by vastly popular SVM are degenerate given used sampling technique.

Setup

Change your misc/config.py file to match base directory and upload to data directory all .libsvm files

Code organization

Scripts fit_knn.py, fit_svm.py, fit_lr.py, fit_melc.py are used to fit models. See python fit_<model>.py -h for usage. Files in scripts folder were used to schedule many fittings, precalculate kernels for Jaccard etc. Each fitting script writes experiment file that can be printed out.

Fit new SVM

To fit svms run python fit_svms.py.

Usage: fit_svms.py [options]

Options:
  -h, --help            show this help message and exit
  -e EXPERIMENT_NAME, --e_name=EXPERIMENT_NAME
  --kernel=KERNEL       
  --experiment_name=EXPERIMENT_NAME
  --seed=SEED           
  --use_embedding=USE_EMBEDDING
  --fingerprint=FINGERPRINT
  --n_folds=N_FOLDS     
  --protein=PROTEIN     
  --max_hashes=MAX_HASHES
  --grid_w=GRID_W       
  --K=K        

Example

python fit_svms.py --kernel=linear --use_embedding=1 --protein=0 --fingerprint=4 --e_name=my_favourite_experiment

Print and plot results

python scripts/fit_svms_print_results.py my_favourite_experiment