Science Inventory

Universal LD50 predictions using deep learning


Sayre, R. AND Chris Grulke. Universal LD50 predictions using deep learning. ICCVAM -Predictive Models for Acute Oral Systemic Toxicity, Bethesda, MD, April 11 - 12, 2018.


EPA’s National Center for Computational Toxicology (NCCT) develops and advances new alternative methods (NAMs) to support evaluation of the toxicity of thousands of chemicals to which Americans could be exposed. One such method is QSAR (quantitative structure-activity relationship); rigorous QSAR models can be used to predict the activity of new substances. Purpose: Develop an approach using deep learning to create a QSAR to describe a set of oral rat LD50 values (a common metric used by risk assessors to determine the hazard posed by a chemical) provided as part of the Predictive Models for Acute Oral Systemic Toxicity project hosted by the ICCVAM Acute Toxicity Workgroup with to support hazard characterization for a broad range of chemicals.


NICEATM Predictive Models for Acute Oral Systemic Toxicity LD50 entry Risa R. Sayre ( & Christopher M. Grulke Our approach uses an ensemble of multilayer perceptron regressions to predict rat acute oral LD50 values from chemical features. Features were generated from QSAR-ready SMILES with MOE 2016.08 (all 2D fingerprints, over 200 values), CORINA Symphony 14698 (Toxprint chemotypes), and RDKit 2016.09.4 (more than 70 numeric or vector fingerprints). SMILES that did not generate mol files in MOE and RDKit, and duplicate SMILES were discarded. LD50 values were log transformed, then scaled from 0 to 1, and feature values were standardized to zero mean with unit variance. All numeric fingerprints from MOE and RDKit were aggregated into single vectors. We used the modeling environment of Keras 2.1.1 in Python 3.5 with a Tensorflow 1.4.0 backend. Each one-hidden-layer Sequential model built per descriptor was optimized with a grid search of the following hyperparameters: (batch size, optimization function, loss function, learning rate, number of hidden dimensions, hidden layer activation function, and output layer activation function). Learning rate reduction and early stopping prevented overfitting. The sum of the distances from the mean of each feature’s training set defines the applicability domain index. The instantiation with the highest validation R2 for a given feature set was tested against a predictivity threshold of training R2 > 0.5, validation R2 > 0.45, RMSE < RMSE using the mean LD50 as a prediction, and a Spearman’s rho over 0.6 with p < 0.01. Models were further validated by comparing y-randomized results. The mean of predictions for each model above the threshold created the ensemble prediction value, which had a validation set R2 of 0.62. This approach does not necessarily reflect US EPA policy.




Record Details:

Product Published Date: 04/12/2018
Record Last Revised: 05/16/2018
OMB Category: Other
Record ID: 340732