takshan/mtl-qsar

Java source code for MTL-QSAR paper

MTL-QSAR

This repository has the Java source code for MTL-QSAR paper Multi-Task Learning with a Natural Metric for Quantitative Structure Activity Relationship Learning

License and Citation

The code in this repository is free and can as desired. However, please cite this paper if you find this code useful in your research or you use it for any other purpose:

@Article{Sadawi2019,
author="Sadawi, Noureddin and Olier, Ivan and Vanschoren, Joaquin and van Rijn, Jan N. and Besnard, Jeremy and Bickerton, Richard and Grosan, Crina and Soldatova, Larisa and King, Ross D.",
title="Multi-task learning with a natural metric for quantitative structure activity relationship learning",
journal="Journal of Cheminformatics",
year="2019",
month="Nov",
day="12",
volume="11",
number="1",
pages="68",
abstract="The goal of quantitative structure activity relationship (QSAR) learning is to learn a function that, given the structure of a small molecule (a potential drug), outputs the predicted activity of the compound. We employed multi-task learning (MTL) to exploit commonalities in drug targets and assays. We used datasets containing curated records about the activity of specific compounds on drug targets provided by ChEMBL. Totally, 1091 assays have been analysed. As a baseline, a single task learning approach that trains random forest to predict drug activity for each drug target individually was considered. We then carried out feature-based and instance-based MTL to predict drug activities. We introduced a natural metric of evolutionary distance between drug targets as a measure of tasks relatedness. Instance-based MTL significantly outperformed both, feature-based MTL and the base learner, on 741 drug targets out of 1091. Feature-based MTL won on 179 occasions and the base learner performed best on 171 drug targets. We conclude that MTL QSAR is improved by incorporating the evolutionary distance between targets. These results indicate that QSAR learning can be performed effectively, even if little data is available for specific drug targets, by leveraging what is known about similar drug targets.",
issn="1758-2946",
doi="10.1186/s13321-019-0392-1",
url="https://doi.org/10.1186/s13321-019-0392-1"
} 

Important Details

Please watch this video to see how to run the code: MTL-QSAR Video

The datasets used in this experiments are available on MTL-QSAR Datasets on OpenML

• Please download the datasets and similarity values available on google drive here: MTL-QSAR Data on GDrive

• Unzip the ECFP6_1024.zip file and make sure the resulting folder replaces the empty folder ECFP6_1024/ which is inside MTL-QSAR/ project folder

• Also, unzip the ChEMBL20.zip file and make sure the resulting folder replaces the empty folder ChEMBL20/ which is inside MTL-QSAR/ project folder. The ChEMBL20/ folder contains the ChEMBL groups and classes of drug targets as explained in the paper. It also includes the similarity values of drug targets and the sequences used to compute these similarity values.

• Important Notice The Feature-based MTL approach explained in the paper is implemented in the MTL java class and The Instance-based MTL approach explained in the paper is implemented in the PIDSimMTL java class.

• Currently the provided java project contains one sample Drug Target Class (under ChEMBL20\L5).

  • The AGC_protein_kinase_NDR_family.txt file contains a list of corresponding datasets that belong to drug targets which happen to be under the group/class AGC_protein_kinase_NDR_family.

  • The file AGC_protein_kinase_NDR_family_sequences.seq contains the sequences of each of the drug targets and the file AGC_protein_kinase_NDR_family_pid.pid contains the percentage ID similarity values for those drug targets.

  • The results of running standard RandomForest, Feature-based MTL and Instance-based MTL approaches as explained in the paper are in the folders: Results\StdRF, Results\MTL and Results\PIDSimMTL respectively.

  • The folder ECFP6_1024\ contains the corresponding datasets (list of ccompounds and their activity values on the drug target) for these drug targets. Notice the dataset name is the drug target ID.

  • The features are molecular fingerprints.

• The full data with all Drug Target Classes and their datasets and similarity values is available on the GDrive mentioned above.

• As part of validation our results, we have performed a randomisation procedure by shuffling the similarity values in the SimMTL approach. We have randomly selected 24 level 5 classes (the total number of drug targets is 120) and randomised their similarity values 1000 times. Each time we randomise we run SimMTL and compute RSME for each drug target. The file Group-24-TID-100-Results-1000reps.csv available under the MTL-QSAR\Results folder has the results. The first two rows are headers (group and drug target TID), the third row is the RSME value for that particular TID without randomising similarity values, and the remaining 1000 rows contain the RSME values after randomisation.

• Also, to validate results, we have performed a randomisation procedure by shuffling the Y-variable in the MTL approach. We have randomised it 1000 times and computed RSME for each drug target each iteration. The file MTL-YRandom1000-StdRF.csv available under the MTL-QSAR\Results folder has the results. The first row is the header (iteration number and drug target TIDs), the second row is the RSME value for that particular TID when performing STL (i.e. standard randomforest), and the remaining 1000 rows contain the RSME values for MTL after randomisation. It can be seen that the values in the standard randomforest row are always smaller than those of the other rows. This has been explained in the paper.

• The file benefited-not-benefited.pdf contains lists of drug target classes that fully or not fully benefit from our settings. This is explained in more detail in the manuscript.