Companion paper: Yap XH, Raymer M. Toxicity Prediction using Locality-Sensitive Deep Learner. Comput. Toxicol. (Under review)
Locality-sensitive deep learner (LSDL) uses attention mechanism to learn locality of a chemical in a dataset. We hypothesize that toxicity data has a locally-linear data structure: local regions with linear feature-target relationship.
Repository is oraganized as follows:
- Processed toxicity datasets (Tox21, AChEi, CoMPARA, AOT; [1-4]) using CDK and PaDEL descriptors and Morgan2 (ECFP2) fingerprints (/data/processed)
- python scripts with helper functions (/0_code)
- Synthetic (locally-linear) data generation (/1_Notebooks/SynthData/SynthData_10dim_generate_dataset.py)
- Synthetic (locally-linear) experiments (1_Notebooks/SynthData/SynthData_10dim_clusternoise)
- python scripts for each dataset (Dataset names: Tox21_CDKPaDEL, AChEi, CoMPARA, AOT_wFP) (/1_Notebooks/{dataset_name}). Note that several of these scripts will need to load trained tensorflow models. The trained models are too large to be included in git repository (>0.1-1GB per model depending on dataset size), but will be cloned onto lab computer (bio-cs). Also please check file paths in code files because files have been moved around/re-organized and paths may be broken.
For Tox21_CDKPaDEL (/1_Notebooks/Tox21_CDKPaDEL/), further preprocessing and training of COSA feature weights was applied in DataProcessing.ipynb. The python scripts Tox21_experiments_{modelName}_singlelabel.py were used to train models in a few variations: dense (feed-forward neural network), LS (Locality-sensitive model without feature weighting), LSwFW (Locality-sensitive model with COSA feature weighting), LSwFW_ones (Locality-sensitive model with naive feature weighting initialized to ones), xgboost (XGBoost).
For AChEi (/1_Notebooks/AChEi/), further preprocessing and training of COSA feature weights was applied in DataProcessing.ipynb. The models were trained using 210503_AChEi_ExtendedDataset_allFeatures.py. The combined model was evaluated in Exploration/ExtendedDatasetAnalyses.ipynb, under 1.2.2 Plots across all folds and 1.3 Repeat analysis with humanOnly.
An additional model using fingerprints only (Exploration/AChEi_smiles_humanOnly) was trained and used to demonstrate the weighting of chemotypes/functional groups on Soman in the thesis and paper.
For AOT_wFP (/1_Notebooks/AOT_wFP), AOT_binary.py, AOT_multiclass.py, AOT_regression.py were used to generate binary/multi-class/regression deep learners with naming convention similar to Tox21_CDKPaDEL (dense, LS, LSwFW, LSwFW_ones, xgboost).
XGBoost models are generated from the files AOT_binary_xgb.py, AOT_multiclass_xgb.py, and AOT_regression.py.
Combined models (Averaged predictions of LSDL with feed-forward neural network) were generated by 1) saving predictions of base classifiers (Results/AOT_predict/modelAOTtraintest_predict); 2) Running (GetCombinedModel.py) to evaluate performance of combined model. This jupyter notebook also outputs predictions of combined model.
Stacking models are obtained by first tuning the XGBoost classifier with (Stacking_tuning.py), then training and evaluating the model(s) with (Stacking.py).
For CoMPARA (/1_Notebooks/CoMPARA), only the binary targets were used. 210427_CoMPARA_binary.py was used to train deep learners and xgboost. The same workflow as AOT_wFP was used: 1) Get predictions (saved in Results/Predicts) and evaluate combined models (GetCombinedModel.py).; 2) Tune Stacking classifier with (Stacking_tuning.py) and train Stacking classifier with (Stacking.py).
[5-6]
Datasets: Tox21: [1] R. Huang, M. Xia, Editorial: Tox21 Challenge to Build Predictive Models of Nuclear Receptor and Stress Response Pathways as Mediated by Exposure to Environmental Toxicants and Drugs, Front. Environ. Sci. 5 (2017). https://doi.org/10.3389/fenvs.2017.00003. AChEi: [2] C.D. Ruark, C.E. Hack, P.J. Robinson, P.E. Anderson, J.M. Gearhart, Quantitative structure-activity relationships for organophosphates binding to acetylcholinesterase, Arch. Toxicol. 87 (2013) 281–289. https://doi.org/10.1007/s00204-012-0934-z. CoMPARA: [3] K. Mansouri, N. Kleinstreuer, A.M. Abdelaziz, D. Alberga, V.M. Alves, P.L. Andersson, C.H. Andrade, F. Bai, I. Balabin, D. Ballabio, E. Benfenati, B. Bhhatarai, S. Boyer, J. Chen, V. Consonni, S. Farag, D. Fourches, A.T. García-Sosa, P. Gramatica, F. Grisoni, C.M. Grulke, H. Hong, D. Horvath, X. Hu, R. Huang, N. Jeliazkova, J. Li, X. Li, H. Liu, S. Manganelli, G.F. Mangiatordi, U. Maran, G. Marcou, T. Martin, E. Muratov, D.T. Nguyen, O. Nicolotti, N.G. Nikolov, U. Norinder, E. Papa, M. Petitjean, G. Piir, P. Pogodin, V. Poroikov, X. Qiao, A.M. Richard, A. Roncaglioni, P. Ruiz, C. Rupakheti, S. Sakkiah, A. Sangion, K.W. Schramm, C. Selvaraj, I. Shah, S. Sild, L. Sun, O. Taboureau, Y. Tang, I. V. Tetko, R. Todeschini, W. Tong, D. Trisciuzzi, A. Tropsha, G. Van Den Driessche, A. Varnek, Z. Wang, E.B. Wedebye, A.J. Williams, H. Xie, A. V. Zakharov, Z. Zheng, R.S. Judson, Compara: Collaborative modeling project for androgen receptor activity, Environ. Health Perspect. 128 (2020) 1–17. https://doi.org/10.1289/EHP5580. AOT: [4] N.C. Kleinstreuer, A.L. Karmaus, K. Mansouri, D.G. Allen, J.M. Fitzpatrick, G. Patlewicz, Predictive models for acute oral systemic toxicity: A workshop to bridge the gap from research to regulation, Comput. Toxicol. 8 (2018) 21–24. https://doi.org/10.1016/j.comtox.2018.08.002.
COSA: [5] J.H. Friedman, J.J. Meulman, Clustering objects on subsets of attributes, J. R. Stat. Soc. Ser. B Stat. Methodol. 66 (2004) 815–839. https://doi.org/10.1111/j.1467-9868.2004.02059.x. COSA: [6] M.M. Kampert, J.J. Meulman, J.H. Friedman, rCOSA: A Software Package for Clustering Objects on Subsets of Attributes, J. Classif. 34 (2017) 514–547. https://doi.org/10.1007/s00357-017.