unixjunkie / rankers

Reference implementation of the Vanishing Ranking Kernels (VRK) method

RanKers

Reference implementation of the Vanishing Ranking Kernels (VRK) method

How to install the software

For beginners/non opam users: download and execute the latest self-installer shell script from (https://github.com/UnixJunkie/rankers/releases).

Then execute:

./rankers-1.0.0.sh ~/usr/rankers-1.0.0

This will create ~/usr/rankers-1.0.0/bin/rankers_bwmine, among other things in the same directory.

For opam users:

opam install rankers

Do not hesitate to contact the author in case you have problems installing or using the software or if you have any question.

Example

Example ROC curve on a hold-out test set. The test set had 38 active molecules and 664 inactives. ROC AUC: 0.861; BEDROC AUC: 0.766; PR AUC: 0.678. The ROC curve is in purple; the precision-recall (PR) curve in cyan. The probability of activity given a raw score is the red curve. The green curve is the number of actives divided by the number of decoys as a function of the scores filtering threshold.

Train and test a model:

rankers_bwmine -i data/tox21_nrar_ligands_std_rand_01.txt

Same, but using 16 cores :

rankers_bwmine -np 16 -i data/tox21_nrar_ligands_std_rand_01.txt

Usage

rankers_bwmine -i <train.txt>
  [-p <float>]: proportion of the (randomized) dataset
  used to train (default=0.80)
  [-k {uni|tri|epa|biw}]: kernel function choice (default=biw)
  [-np <int>]: max number of processes (default=1)
  [-o <filename>]: write raw test scores to file
  [--train <train.txt>]: training set (overrides -p)
  [--valid <valid.txt>]: validation set (overrides -p)
  [--test <test.txt>]: test set (overrides -p)
  [-n <int>]: max number of optimization steps; default=150
  [--capf <float>]: keep only fraction of decoys
  [--capx <int>]: keep only X decoys per active
  [--capi <int>]: limit total number of molecules
  (but keep all actives)
  [--seed <int>: fix random seed]
  [--pr]: use PR AUC instead of ROC AUC during optimization
  [-kb <float>]: user-chosen kernel bandwidth
  [--mcc-scan]: scan classif. threshold to maximize MCC
  [--tap]: tap the train-valid-test partitions to disk
  [-q|--quick]: exit early; just after model training
  [--noplot]: turn off gnuplot
  [-v]: verbose/debug mode
  [-h|--help]: show this help message

rankers's Issues

rankers_bwexploit: add --bin option

to read the encoded database from a binary file

bench it then

experiment with counted-Bloom

cf. UnixJunkie/molenc#72

release minor version in opam

after parany.10.0.0 is accepted

bwmine: implement --brute

To disable the optimizer.
Syntax: [--brute ]: nsteps; default = 100

regression

Either Inverse Distance Weighting / Shepard Interpolation (NRBF).
Later, the more complex RBF Interpolation with LU decomposition.

remove dependency to parmap

release v2.0.0

bwmine: implement --NxCV