agitter / lightchem

Light-weight Machine Learning Framework for Drug Discovery

LightChem

LightChem provides high level machine-learning interface that served to used by researcher without deep machine-learning background. It aims to offer convenient exploration for researcher interested in machine-learning for drug discovery. LightChem is a package created by Haozhen Wu from Small Molecule Screening Facility
at University of Wisconsin-Madison.

Key features:

• XGBoost backend
• Parallel computing
• Supports tree based models and regression models
• Interface to use ensemble( stacking ) models
• Support multiple evaluation metrics
• Common featurization methods to transform molecules into fingerprint
• Competitive benchmark results for well-known public datasets

Table of contents:

• Example
  • MUV
  • TOX21
  • PCBA128
• Datasets
• Dependencies
• Installation
• Benchmark Results
• FAQ
• RoadMap
• Reference

Dependencies:

• sklearn version = 0.18.1
• XGBoost version = 0.6
• numpy version = 1.11.1
• scipy version = 0.18.1
• pandas version = 0.18.1
• rdkit version = 2015.09.1

Installation

We recommend you to use Anaconda for convenient installing packages. Right now, LightChem has been tested for Python 2.7 under OS X and linux Ubuntu Server 16.04.

1. Download 64-bit Python 2.7 version of Anaconda for linux/OS X here and follow the instruction. After you installed Anaconda, you will have most of the dependencies ready.

2. Install git if do not have:
   Linux Ubuntu:

   sudo yum install git-all

3. Install sklearn:

   conda install scikit-learn=0.18

4. Install xgboost documentation
   Linux Ubuntu: Build xgboost shared library

   git clone --recursive https://github.com/dmlc/xgboost
   cd xgboost; make -j4

   Or if OSX: Build xgboost shared library

   git clone --recursive https://github.com/dmlc/xgboost
   cd xgboost; cp make/minimum.mk ./config.mk; make -j4

   After shared library is built, build xgboost python package

   cd python-package; python setup.py develop --user

5. Install rdkit Note: rdkit is only used to transform SMILE string into fingerprint.

   conda install -c omnia rdkit

6. Clone the lightchem github repository:

   git clone https://github.com/haozhenWu/lightchem.git

   cd into lightchem directory and execute

   pip install -e .
   

Benchmark Results

Model Abbreviation:
GbtreeLogistic: Gradient boosted tree with binary logistic as objective
GbtreeRegression: Gradient boosted tree with linear regression as objective
GblinearLogistic: Gradient boosted linear with binary logistic as objective
GblinearRegression: Gradient boosted linear with linear regression as objective

Evaluation results report below are mean/median of k-fold cross-validation results among all targets in a given dataset.

Concatenated Version:

Molecules of targets from same dataset are concatenated together into a single big dataframe, which leads to many missing values in the labels. Molecules without a lab result for a given target assume having a negative label. This approach might benefit multi-task learning.
LightChem so far only supports single-task learning, which builds separate models for each target. Will have Self-contained molecules version later.

Classification

Stratified Split
Cross Validation: 3 folds, 75% of the data
Test: Remaining 25% of the data

ROC-AUC

Cross Validation

Dataset	Layer	Label	Feature	Model	Mean	Median
tox21	Layer1	Binary	ECFP1024	GbtreeLogistic	0.783+-0.022	0.779+-0.017
				GblinearLogistic	0.729+-0.021	0.747+-0.012
			MACCSkeys	GbtreeLogistic	0.798+-0.019	0.801+-0.016
				GblinearLogistic	0.766+-0.02	0.776+-0.023
	Layer2	Binary	Layer1 holdout predictions	GbtreeLogistic	0.790+-0.018	0.794+-0.014
				GblinearLogistic	0.809+-0.020	0.806+-0.017
PCBA128	Layer1	Binary	ECFP1024	GbtreeLogistic	0.843+-0.012	0.857+-0.007
			MACCSkeys	GbtreeLogistic	0.835+-0.012	0.847+-0.008
		Continuous	ECFP1024	GbtreeRegression	0.863+-0.013	0.882+-0.008
			MACCSkeys	GbtreeRegression	0.852+-0.013	0.865+-0.009
	Layer2	Binary	Layer1 holdout predictions	GbtreeLogistic	0.880+-0.011	0.894+-0.007
				GblinearLogistic	0.879+-0.015	0.893+-0.006
muv	Layer1	Binary	ECFP1024	GbtreeLogistic	0.799+-0.053	0.793+-0.048
				GblinearLogistic	0.756+-0.055	0.805+-0.044
			MACCSkeys	GbtreeLogistic	0.810+-0.049	0.820+-0.046
				GblinearLogistic	0.755+-0.054	0.756+-0.059
	Layer2	Binary	Layer1 holdout predictions	GbtreeLogistic	0.787+-0.053	0.811+-0.053
				GblinearLogistic	0.687+-0.155	0.669+-0.109

Test:

Dataset	Layer	Label	Feature	Model	Mean	Median
tox21	Layer1	Binary	ECFP1024	GbtreeLogistic	0.776	0.748
				GblinearLogistic	0.734	0.738
			MACCSkeys	GbtreeLogistic	0.805	0.797
				GblinearLogistic	0.765	0.769
	Layer2	Binary	Layer1 holdout predictions	GbtreeLogistic	0.795	0.783
				GblinearLogistic	0.808	0.788
PCBA128	Layer1	Binary	ECFP1024	GbtreeLogistic	0.850	0.866
			MACCSkeys	GbtreeLogistic	0.843	0.857
		Continuous	ECFP1024	GbtreeRegression	0.872	0.890
			MACCSkeys	GbtreeRegression	0.856	0.868
	Layer2	Binary	Layer1 holdout predictions	GbtreeLogistic	0.883	0.903
				GblinearLogistic	0.884	0.901
muv	Layer1	Binary	ECFP1024	GbtreeLogistic	0.744	0.712
				GblinearLogistic	0.671	0.681
			MACCSkeys	GbtreeLogistic	0.774	0.824
				GblinearLogistic	0.751	0.782
	Layer2	Binary	Layer1 holdout predictions	GbtreeLogistic	0.663	0.689
				GblinearLogistic	0.552	0.605

EFR1

Cross Validation

Dataset	Layer	Label	Feature	Model	Mean	Median
tox21	Layer2	Binary	Layer1 holdout predictions	GbtreeLogistic	15.724+-1.103	18.408+-0.726
				GblinearLogistic	15.162+-1.702	18.352+-1.122
PCBA128	Layer2	Binary	Layer1 holdout predictions	GbtreeLogistic	33.873+-2.736	32.822+-1.682
				GblinearLogistic	32.665+-3.816	33.053+-1.869
muv	Layer2	Binary	Layer1 holdout predictions	GbtreeLogistic	21.580+-9.621	18.452+-7.192
				GblinearLogistic	7.178+-5.489	0+-0

Test:

Dataset	Layer	Label	Feature	Model	Mean	Median
tox21	Layer2	Binary	Layer1 holdout predictions	GbtreeLogistic	15.612	12.360
				GblinearLogistic	14.774	11.842
PCBA128	Layer2	Binary	Layer1 holdout predictions	GbtreeLogistic	34.413	33.333
				GblinearLogistic	33.703	32.727
muv	Layer2	Binary	Layer1 holdout predictions	GbtreeLogistic	11.905	0
				GblinearLogistic	0	0

FAQ

1. When I import lightchem, the following error shows up version GLIBCXX_3.4.20 not found:
   Try:

   conda install libgcc

   Source

Reference

1. [DeepChem] (https://github.com/deepchem/deepchem): Deep-learning models for Drug Discovery and Quantum Chemistry