orcid.org/0000-0003-3863-9689 Building Scientific Confidence in the Development and Evaluation of Read-Across GenRA: Evaluating local validity for read-across prediction using chemical and biological information PRIORITIZATION ToxPi 1 1 1 1 1 1 1 1 CHEMICAL SUMMARY Chemical CASRN A B C D E F G H Name 80-05-7 Bisphenol A 80-05-1 Bisphenol B Studies 80-05-2 Bisphenol C 80-05-3 Bisphenol D 80-05-4 Bisphenol E 80-05-5 Bisphenol F 80-05-6 Bisphenol G 80-05-7 Bisphenol H 80-05-8 Bisphenol I 80-05-9 Bisphenol J 80-05-7 80-05-1 80-05-2 80-05-3 80-05-5 SCORING APPLY Interactive Chemical Safety Desc Summary Log for Sustainability Web Tool Tip Application Description of Assays (Data) or Prioritization Mode whatever is being hovered over TOXCAST iCSS v0.5 Grace Patlewicz National Center for Computational Toxicology The views expressed in this presentation are those of the author and do not necessarily reflect the views or policies of the U.S. EPA National Center for Computational Toxicology
Outline • Definitions • Workflow for category development and read- across • Identifying the sources of uncertainties associated with read-across and practical strategies to address these • Quantifying uncertainties and assessing performance of read-across • From research to implementation • Summary National Center for Computational Toxicology
Definitions: Read-across Known information on the property of a substance (source chemical) is used to make a prediction of the same property for another substance (target chemical) that is considered “similar” i.e. Endpoint & often study specific Source Target Reliable data chemical chemical Missing data Property Acute fish toxicity? Predicted to Known to be harmful be harmful 2 National Center for Computational Toxicology
Chemical category and read-across: General Workflow 1. Decision context 2. Data gap analysis 3. Overarching hypothesis 4. Analogue identification 5. Analogue evaluation – Data gap filling 6. Uncertainty assessment National Center for Computational Toxicology
Chemical category and read-across: General Workflow 1. Decision context 2. Data gap analysis 3. Overarching hypothesis 4. Analogue identification 5. Analogue evaluation – Data gap filling 6. Uncertainty assessment National Center for Computational Toxicology
1. Decision context • Prioritisation e.g. PMN • Screening level hazard assessment • Risk Assessment e.g PPRTV • Different decision contexts will dictate the level of uncertainty that can be tolerated National Center for Computational Toxicology
6. Sources of Uncertainty • Analogue or category approach? (#analogues) • Data quality • Overarching hypothesis/Similarity rationale – how to identify similar analogues and justify their similarity for the endpoint of interest • Address the dissimilarities and whether these are significant from a toxicological standpoint • Presence vs absence of toxicity • Toxicokinetics – including Metabolism National Center for Computational Toxicology
Identifying Uncertainties • Several publications that guide the construction and assessment of categories and use of read-across SCIRADE – Guidance and examples (OECD, 2014; ECHA, 2008; ECETOC TR 116, 2012) – Frameworks for identifying analogues e.g. Wu et al, 2010, Patlewicz et al, 2013 – Frameworks for assessing read-across (ECHA – RAAF, Blackburn and Stuart, 2014, LERAT (Patlewicz et al, 2015) National Center for Computational Toxicology
Addressing uncertainties - 1 • Search and Selection of analogues • Using metabolism information • Presence or absence of toxicity • Using in vitro data such as HTS data to enhance read-across National Center for Computational Toxicology
Search and selection of analogues • Explored the use of different structure-based approaches (Pubchem, Chemotyper and MoSS MCSS with Tanimoto index as a measure of similarity) to identify hindered phenol analogues and evaluate their validity for reading across Estrogenicity • Make a read-across Estrogenicity prediction for each target hindered phenol National Center for Computational Toxicology
Read-across predictions Filtering 1 (Log P ow & MV) Filtering 2 (No. of Data Sources) See poster from P Pradeep National Center for Computational Toxicology
Case study conclusions • Initial selection of analogues based on different descriptor sets (for this example) was invariant to the read-across prediction performance • Evaluating analogue validity paying close attention to the quality of the underlying analogue data and relevant physchem properties did significantly improve read-across predictive performance National Center for Computational Toxicology
Metabolism - 1 • Do we always need to do a detailed assessment of metabolism for read-across? • Or can we identify sufficiently based on existing tools and data? • Skin sensitisation ~ 22-25% of skin sensitisers require some level of activation – Whether activation is by oxidation (pre) or due to metabolism (pro) is less well understood – Tools e.g.TIMES-SS, OECD Toolbox can be helpful in diagnosing whether a substance is direct acting or indirect acting (pre- or pro- hapten) – For a dataset of 127 substances, non-animal methods could correctly identify the majority of pre and pro- haptens See associated poster National Center for Computational Toxicology
Addressing uncertainities - 2 • Read-across acceptance is context dependent – based on subjective expert judgement assessment – potential lack of harmonised or reproducible decisions • No clear understanding of what constitutes success • Do we know what the performance of a read- across is really like on a more general level? Critical need is an objective measure of uncertainty in a read-across prediction National Center for Computational Toxicology
Quantifying uncertainty & Assessing performance of read- across • GenRA (Generalised Read-Across) is a “local validity” approach • Predicting toxicity as a similarity-weighted activity of nearest neighbours based on chemistry and bioactivity descriptors α Σ k tox s x j ij j tox y = i α Σ α k = {chm, bio bc} s , j ij Where tox , in this case, is the in vivo toxicity of chemical j x j Shah et al, (2016) • Initial focus relied on standard guideline studies • Toxicity effects recorded as binary outcomes National Center for Computational Toxicology
15 National Center for Computational Toxicology
GenRA: Nominal cluster Explore performance as a function of number of nearest neighbours or similarity index National Center for Computational Toxicology
Quantifying uncertainty & Assessing performance of read-across • Tested and compared 1. Chemical descriptors 2. Bioactivity descriptors 3. Hybrid of chemical and bioactivity descriptors • No preselection of descriptors was performed • Bioactivity descriptors were often found to be more predictive of in vivo toxicity outcomes • The approach enabled a performance baseline for read-across predictions of specific study outcomes to be established • But still context dependent on the endpoint and the chemical neighbourhood under study National Center for Computational Toxicology
Quantifying uncertainty & Assessing performance of read- across Next steps in progress: • Use of other chemical descriptor sets that encode more expert knowledge of SARs • Incorporating TK information National Center for Computational Toxicology
Analysing local neighborhood of a chemical 1 0.8 s=0.81 k=4 1 1 1 19 National Center for Computational Toxicology
Analysing local neighborhood of a chemical 0.7 s=0.72 k=6 2 2 1 1 2 2 20 National Center for Computational Toxicology
Analysing local neighborhood of a chemical 0.63 3 s=0.63 k=10 1 1 2 3 2 3 21 National Center for Computational Toxicology
Analysing local neighborhood of a chemical 1 2 3 θ 1 2 3 22 National Center for Computational Toxicology
From Research to Implementation • Public accessible tool building on the iCSS Chemistry Dashboard under development 23 National Center for Computational Toxicology
From Research to Implementation 24 National Center for Computational Toxicology
Summary • Still many challenges remain in read-across • Quantifying the uncertainty of read-across prediction is a critical issue • Have illustrated a handful of the research directions being taken National Center for Computational Toxicology
Acknowledgements • Imran Shah • Tony Williams • Richard Judson • Rusty Thomas • Prachi Pradeep • Participants of the JRC Expert meeting on pre and pro haptens held Nov 10-11, 2015 National Center for Computational Toxicology
Recommend
More recommend
