1. Introduction

Towards an MIE Atlas Tools for Toxicity Prediction

Timothy E. H. Allen1, Sonia Liggi1, Jonathan M. Goodman1,

Paul J. Russell2, and Steve Gutsell2

1. Centre for Molecular Informatics, Department of Chemistry, University of Cambridge, Lensfield Road, Cambridge, CB2 1EW

2. Safety & Environmental Assurance Centre, Unilever, Colworth Science Park, Sharnbrook, Bedfordshire, MK44 1LQ

5. Expanding the Atlas

Acknowledgements

Unilever Centre for Molecular Informatics, University of Cambridge St John’s College, Cambridge

2. Finding the MIE

References

Consumer and environmental safety decisions are based on exposure and hazard data interpreted us-

ing risk assessment approaches. The adverse outcome pathway (AOP) conceptual framework [1] has

been presented as a logical sequence of events or processes within biological systems which can be

used to understand adverse effects and refine the current risk assessment practice.

The molecular initiating event (MIE) can be thought of as a gateway to the AOP—the initial chemical

interaction [2]. Chemistry is key to understanding the MIE. What is it about these molecules that al-

low them to do this?

Using knowledge of the chemical characteristics that govern these interactions, a greater understand-

ing of why chemicals cause toxic effects can be gained.

6. Conclusions

Literature searches were used to gain an understanding of MIEs, and how fragments can be used to

describe them.

This led to the development of a unified MIE definition.

Our approach has been applied to the NIH Tox Data Challenge and ChEMBL data to increase our un-

derstanding of MIEs associated with pharmacologically important targets.

We aim to build this understanding into an SAR tool for toxicity screening.

Acetaminophen (Paracetamol) is a widely used, mild analgesic. Acetaminophen overdoses can cause

potentially fatal liver failure, via an oxidised metabolite - N-acetyl-p-benzoquinone imine (NAPQI). Ac-

etaminophen is converted to NAPQI by a P450 oxidation process in the liver, mediated by the enzyme

CYP2E1 [3]. This activation is the MIE. NAPQI then binds to the reactive oxygen species (ROS) scaven-

ger glutathione, depleting this vital protection against oxidative stress [3]. This processes may be

thought of as the MIE for NAPQI. When the glutathione defence is depleted the excess NAPQI binds

to cellular proteins, lipids and nucleic acids, activates calpains, and generates ROS [4,5]. The resulting

oxidative stress is the main culprit of acetaminophen hepatotoxicity.

The para-aminophenol fragment was identified as key for this activity. This was matched to other ex-

ample compounds using the Derek Nexus [6].

As well as acetaminophen, Derek identified phenacetin and amiodiaquine as examples containing the

fragment. Phenacetin is known to cause hepatotoxicity in rats and mice, even when not metabolised

to acetaminophen [7]. Its toxicity in humans however seems to be minimal. Amiodiaquine has several

reported cases of hepatotoxicity in humans, including some fatalities [8].

In an attempt to apply this methodology to a larger dataset a number of appropriate receptors were

chosen, based on our previous research and Bowes 2012 paper on pharmacological anti-targets [12].

Receptor binding data was obtained from the open source database ChEMBL [13].

Some 2D fragments elucidated for the histamine H1 receptor are shown below.

These fragments show little structural similarity to the endogenous ligand, histamine. Some specific

H1 binders shown below show diphenylmethane fragments to be common. A crystal structure was

found for the H1 receptor bound to doxepin [14], furthering our understanding of this MIE, and allow-

ing SMARTS to be written to describe it.

1. Ankley G. T. et al., Environ. Toxicol. Chem., (2000) 29, 730-741.

2. Gutsell, S., Russell, P.J., Toxicol. Res., (2013) 2, 299-307.

3. Black, M., Ann. Rev. Med., (1984) 35, 577-593.

4. McGill, M. R. et. al., Hepatology (2011) 53, 974-982.

5. Jaeschke, H., Bajit, M. L., Toxicol. Sci. (2006) 89, 31-41.

6. Lhasa Limited. Derek Nexus v.3.0.1. Copyright 2012.

7. Prescott, L. F., Br. J. Clin. Pharmac., (1980) 10, 373S-379S.

8. Neftel, K. A. et. al., Br. Med. J., (1986) 292, 721-723.

9. AOP-Wiki: aopkb.org/aopwiki/index.php

10. Allen T. E. H. et. al., Chem. Res. Toxicol., (2014) 27, 2100-2112.

11. NIH Tox Data Challenge: www.tripod.nih.gov/tox21/challenge/about.jsp

12. Bowes J. et. al., Drug Discov., (2012) 11, 909-922.

13. ChEMBL Database: www.ebi.ac.uk/chembl

14. Shimamura T. et. al., Nature, (2011) 475, 65-70.

Ankley's conceptual diagram of an AOP, including the MIE. Image adapted from Ankley 2010 [1].

MIE MIE

To prove the concept of predictive toxicology using the chemistry of molecules, models needed to be

developed and tested using the knowledge we had gained so far. Our aim was to build clear models

using fragmental structural alerts in 2D, and a range of chemically sound molecular descriptors. This

will allow understanding to be gained about the MIE itself and help regulatory acceptance.

In summer 2014 the NIH Tox Data Challenge [11] was lunched to evaluate computational models for

toxicity prediction, providing data for nuclear receptors. We analyzed data for the androgen recep-

tor—ligand binding domain and found a number of steroidal and non-steroidal fragments associated

with positive activity, and several fragments associated with negative activity.

The following model was found to perform best: requiring a molecule to contain a fragment from Set

A, or no negative fragments and a fragment from Set B to be predicted positive.

Our models show a high overall quality (Q), which may be misleading due to the skewed nature of

the data sets. The Matthews Correlation Coefficient (MCC) shows against a balanced test set our

models provide between 65 and 85% correct prediction. Furthermore our balanced accuracy of 62.3

would have placed us second on the NIH’s leaderboard for the evaluation of their test set.

4. The NIH Tox Data Challenge

As a key anchor for the AOP, and a commonly used term, an understanding of what an MIE is and

how it should be defined is required. As AOPs become more prevalent, so will MIEs, particularly with

the development of AOP maps and open source systems [9].

A single definition for the MIE has yet to reach general acceptance. Different definitions stem from

different fields that have focused on a specific type of interaction. With experience from the atlas of

many interactions, we combined the best features of existing definitions to form a unified definition:

Our unified definition defines the MIE as the initial interaction between a molecule and a bio-

molecule or biosystem that can be causally linked to an outcome via a pathway [10].

We believe this to be a good definition, as it focuses on the initial interaction, relates the interaction

to a measurable outcome, includes a multitude of different interactions, and does not focus the term

entirely in toxicological research.

3. Defining the MIE

Set Model TP FN TN FP SE SP Q MCC

Train (SET A) Or (NO NEG + SET B) 188 115 8223 73 62.0 99.1 97.8 0.657

Test (int.) (SET A) Or (NO NEG + SET B) 12 7 816 8 63.2 99.0 98.2 0.606

NIH Test (SET A) Or (NO NEG + SET B) 1 3 247 1 25.0 99.6 98.4 0.346

SET B Positive Negative

SET A

Existing Binders Crystal Structure SMARTS

Charge-

Charge

Interaction

63/499

Hydrophobic Pocket