View profile

Computational Toxicology Newsletter - Issue #3

Computational Toxicology Newsletter - Issue #3
By CompTox Team • Issue #3 • View online
The team of Comptox newsletter wishes our subscribers a great start to 2022! The field is moving fast and we are here to showcase some of this progress! We are also happy to announce that some episodes of the podcast series on the in3 project (which brought this team together) are available on spotify. Of special relevance to this newsletter is our episode on integration of in silico and in vitro assays.
Quick summary:
Machine learning Kinetics
PBK in vitro-in silico integration Kinetics
Kinetics
Uncertainty Deep learning
Happy reading!
— — — — — — — — — — — — — — — — — —
In Molecules by Toma C et al
Motivation
In the environmental risk assessment of chemicals, toxicological data for three trophic levels of aquatic organisms are needed to assess the acute and chronic toxicities to the aquatic environment. Typically, algae, crustaceans and fish are used to represent primary producers, primary consumers and secondary consumers, respectively. Computational models can be used to speed up the assessment process when there are many substances to be tested, while also avoiding the need for vertebrate (fish) testing.
🧠 Study aim
To develop QSAR models to predict the acute and chronic toxicity of chemicals toward algae (Raphidocelis subcapitata, previously known as Pseudokirchneriella subcapitata), crustaceans (Daphnia magna), and fish (several species).
💡 Key highlights
  • A battery of random forest models QSAR models was developed for Raphidocelis subcapitataDaphnia magna, and fish for both acute and chronic toxicities.
  • The QSARs were developed using a random forest (the tree ensemble) method
  • The models are freely accessible via the VEGAHUB tool
☕ Available resources
🧭 Context of use
Risk assessment of the effects of chemicals on aquatic ecosystems
— — —
In Molecular Pharmaceutics by Miljković F et al
Motivation
Progression of a novel drug candidate toward the clinic requires characterisation of its pharmacokinetic (PK) behaviour to ensure sufficient drug exposure at the site of action. However, PK remains one of the main contributing factors of compound attrition in clinical development.
🧠 Study aim
To develop machine learning models for 12 human PK parameters:
-plasma AUC after IV and PO (oral), IV CL, PO CLrenal, PO Cmax in plasma, PO bioavailability, IV elimination half-life, time to reach Cmax PO, apparent volume of distribution at steady state, apparent volume of distribution after non-intravenous administration
💡 Key highlights
  • A data curation protocol was developed to obtain a range of human in vivo PK parameters covering 1001 unique compounds and corresponding to 4491 compound–dose combinations from US Food and Drug Administration and European Medicines Agency documents
  • The chemical space of the 1001 clinical drugs was overlaid with a subset of randomly selected 200 000 small bioactive molecules from ChEMBL
  • Machine learning was used to develop models of 12 human in vivo PK parameters using only chemical structure information and available doses
  • Regression models for all PK parameters were generated using random forest and eXtreme Gradient Boosting algorithms. Both methods were comparable for all tasks with a slight advantage of RF models.
  • Deep learning did not provide an advantage over simpler machine learning methods in predicting in vivo PK data.
☕ Available resources
None
🧭 Context of use
Prediction of human pharmocokinetics during drug discovery
— — —
In Toxics by Armitage JM et al
Motivation
In vitro bioactivity and toxicity data are increasingly becoming available to characterize the potential hazards and risks of chemicals, and to support in vitro to in vivo extrapolations. However, there are challenges in the interpretation of in vitro toxicity test data since the typical practice of using nominal (administered) bulk medium concentrations to report the dose associated with responses instead of more directly relevant metrics such as the freely-dissolved or cellular concentration, which are typically not measured. The authors therefore developed a mathematical model to simulate the mass balance distributions of chemicals applied to in vitro systems. This paper reports the updating of the 2014 IV-MBM v1.0 model (where MBM is mass balance model) to IV-MBM EQP v2.0 (where EQP is equivalent plasma concentration).
🧠 Study aim
To describe, apply and evaluate the IV-MBM EQP v2.0 model to demonstrate its use for interpreting in vitro bioactivity and toxicity data.
💡 Key highlights
  • The distribution of the chemical in the test system is simulated by the EQP v2.0 model as a function of partitioning ratios and volumes or surface areas of the various phases included (e.g., medium, cells, serum lipids, and albumin, vessel wall)
  • Model performance was better for neutral organic chemcals compared with ionizable organic chemicals
  • The model is applicable to “non-volatile” and “volatile” neutral organics if allowances for a leakage into the surrounding air and/or sorption to a well plate covering are included
  • Key characteristics of the test system (e.g., well plate size, volume of medium, cell seeding density, FBS volume fraction) can readily be changed by the user
  • The model was parameterised and evaluated using four independent data sets with measured ratios of bulk medium or freely-dissolved to initial nominal concentrations
  • The model was applied to several hundred chemicals on Canada’s Domestic Substances List (DSL) with nominal effects data (AC50s) reported for two in vitro assays
Available resources
🧭 Context of use
Extrapolation of in vitro bioactivity and toxicity data. Design of in vitro toxicity experiments.
— — —
In Computational Toxicology by Sweeney LM and Sterner TR
Motivation
The USA Air Force needs to make exposure assessment of volatile chemicals used in their workplace, such as the chemicals constituting jet propulsion fuel 8 (JP-8). Some of these chemicals are well characterized but others are novel chemicals. The use of QSARs to parameterize PBK model can help perform the exposure assessment of these novel chemicals in a more timely manner. Metabolism is a especially important and difficult kinetic parameter to predict. There are several QSARs for prediction of Vmax and Km in the literature but their predictions are very much dependent on their applicability domains and descriptors used. 
🧠 Study aim
Review QSARs for Km and Vmax in literature based on OECD guideline for QSARs to understand their applicability domain. Compare their applicability domain with the chemicals in JP-8 to understand the potential use of QSARs in PBK parameterization for volatile chemical exposure assessment.
💡 Key highlights
  • None of QSARs from literature were fully reproducible due to various reasons, e.g. incorrect input information, sources of experimental data not reported or discontinued software tools. Hence all QSAR were refined in this paper.
  • For all QSARs reviewed, Km predictions were worst than Vmax.
  • The QSAR focused on CYP substrates led to the highest Km/Vmax predictions.
  • There is little overlap of applicability domains between a QSAR for volatile chemicals and a QSAR focused on CYP substrates, with the first being more applicable for the JP-8 case-study chemicals.
  • To have more accurate applicability domains it is important that the some descriptors reflect mechanistic knowledge , e.g. the functional groups that are substrates of the enzyme in study.
Available resources
  • The full curation and refinement of the several QSAR evaluated in this paper is reported in supplementary data.
🧭 Context of use
PBK parameterization of intrinsic clearance for volatile organic chemicals.
— — —
In Toxicological Sciences by Punt A et al
Motivation
Several efforts have been made to derive parameters from in vitro and in silico tools to input in PBK models. These efforts resulted in the existence of several in vitro and in silico methods to derive intrinsic clearance, fraction unbound in plasma, partition coefficient to the different tissues and oral first order absorption.
🧠 Study aim
Evaluate the accuracy of different methods of PBK parameterization, and combination of the methods, to predict Cmax in plasma in rat after oral exposure. The parameters and methods evaluated are:
Intestinal uptake - i) CACO-2 cells derived Papp or ii) QSAR based on the topological surface area (TPSA) 
Tissue:plasma partition coefficients - i) Berezhkovskiy method, ii) Rodgers and Rowland or iii) Schmitt
Intrinsic clearance - i) primary heptocytes, ii) S9 microsomes or iii) ADMET predictor CYP clearance 
Fraction unbound in plasma based on i) in vitro with rapid equilibrium dialysis or ii) in silico based on logP, logD and pKa
Physicochemical characteristic (e.g. logP) - i) ADMET or ii) ChemAxon.
💡 Key highlights
A generic PBK model was used to run
  • Of the 44 chemicals, bisphenol A, curcumin, permethering and resmetherin have the widest range of estimations.
  • Input parameters related to metabolic clearance are the most influential (clearance but also fub and blood flow in liver), followed by parameters related to oral absorption and blood:plasma ratio.
  • No specific characteristic of the chemicals had a trend of mispredictions.
  • Best performance was obtained when the: i) hepatic clearance was parameterized based on in vitro (hepatocytes or liver S9) measured intrinsic clearance; ii) the method of Rodgers and Rowland for calculating partition coefficients; and iii) in silico calculated fraction unbound plasma and Papp values (the latter is especially true for very lipophilic compounds).
  • Models does not include transport mediated processes and extrahepatic metabolism which can add large uncertainty.
Available resources
🧭 Context of use
Predict Cmax in plasma in rat after oral exposure.
— — —
In Frontiers of Pharmacology by Loizou G et al
✨ Motivation
Currently two issues plague QIVIVE; 1) the point of departure derived in in vitro is based on nominal concentration which does not reflect the differences in fate and distribution in the in vitro system and the tissue in in vivo and 2) QIVIVE is performed usually without considering the PBK model structure uncertainty or parameters values uncertainty. The study is part of a series of QIVIVE studies that use a similar probabilistic framework to evaluate chemicals flagged as hazardous in humans.
🧠 Study aim
💡 Key highlights
Available resources
🧭 Context of use
— — —
In Computational Toxicology by Tice R et al
✨ Motivation
In silico carcinogenicity models have limited use in research and regulatory decision-making.
🧠 Study aim
In silico models for a series of ten key characteristics (KCs) of carcinogenicity were reviewed with the scope to indicate where experimental methods and in silico models currently exist for each KC to support the hazard assessment for carcinogenicity and, importantly, where gaps exist.
💡Key highlights
  • The ten key characteristics of carcinogenicity include: (KC1) is electrophilic or can be metabolically activated; (KC2) is genotoxic; (KC3) Alters DNA repair or causes genomic instability; (KC4) induces epigenetic alterations; (KC5) induces oxidative stress; (KC6) induces chronic inflammation; (KC7) is immunosuppressive; (KC8) modulates receptor-mediated effects; (KC9) causes immortalization; (KC10) alters cell proliferation, cell death, or nutrient supply.
  • The KC construct provides a pragmatic platform for collecting data and organising information regarding carcinogens and it has recently been used in combination with read-across principles to support a recommendation to include chemicals in the NTP Report on Carcinogens that did not have animal cancer data.
  • The 10 KCs are meant to comprehensively account for the different mechanisms involved in tumor initiation, promotion, and progression, regardless of prevalence.
  • The “is genotoxic” KC is the only characteristic where global in silico methods for predicting carcinogenicity hazard have been developed that are being used for making regulatory decisions.
  • In silico predictions of electrophilicity exist also but are not accepted for making regulatory decisions alone due to low specificity. Mechanistic challenges as well as the absence of relevant methods for measuring certain KCs and/or the lack of robust experimental data, along with their complex causal relationship to carcinogenesis, make the other nine KCs less tractable for the development and application of in silico prediction models.
☕ Available resources
NA
🧭 Context of use
Hazard identification and risk assessment
— — —
In Computational Toxicology by Chambers B and Shah I
✨ Motivation
Environmental chemicals disrupt cellular homeostasis, resulting in the activation of adaptive stress response pathways (SRPs) to recover from chemical perturbation. SRPs have been linked to drug-induced liver injury, the critical stages of diabetes, neurodegenerative diseases, and cancers. A comprehensive set of SRP classifiers paired with rich transcriptomic data could efficiently screen environmental chemicals for non-specific effects and potential drugs for off-target effects.
🧠 Study aim
Assessment of non-specific chemicals by elucidating their activation of SRPs using transcriptomic data.
💡Key highlights
  • The construction of SRP consensus signatures sets was completed in three steps. First, the authors constructed consensus signatures by merging and pruning relevant gene sets from the MSigDB v7.1 database. Second, they developed an independent gene expression validation set by identifying reference perturbagens from the literature and curating the transcriptomic profiles from publicly available sources. Third, they used gene set enrichment analysis (GSEA) to score matches between signatures and transcriptomic profiles. Lastly, they evaluated the performance of GSEA scores as classifiers of SRP activity within reference perturbagen transcriptomic profiles using ROC AUC analysis. The entire workflow and associated data are available from the authors upon request.
  • The mapped SRPs can also address the issue of crosstalk between SRPs that would mask the mechanisms of actions.
  • There is a need for a curated catalog of reference chemicals for SRPs.
☕ Available resources
NA
🧭 Context of use
Hazard and risk assessment
— — —
In Journal of Chemical Information and Modeling by Feinstein J et al
✨ Motivation
  • The data for studying the toxicities of perfluoroalkyl and polyfluoroalkyl substances (PFAS) are scarce.
  • There is a need for artificial intelligence (AI) that can replicate the human-like decision to say “I can’t answer” for low-confidence/high-risk scenarios.
🧠 Study aim
An uncertainty-informed transfer-learning approach for predicting and understanding PFAS toxicities is proposed, based on multiple machine learning (ML) methods, including random forests, deep neural networks (DNN), graph convolutional networks, and Gaussian processes, for predicting acute toxicity (e.g., median lethal dose, or LD50) of PFAS compounds.
💡Key highlights
  • Uncertainty is evaluated here as the ability of the chosen metric to capture the model error; in other words, a suitable measure for evaluating the efficacy of an uncertainty metric is the correlation of uncertainty with the model error.
  • Two approximations for the model uncertainty were evaluated: (1) deep ensemble and (2) latent space distance, as well as analyzing the best-performing mechanism within the context of our validation set.
  • Selective prediction is an ML paradigm where the goal is to learn a prediction model that knows when it does not know. A selective prediction model performs “learning with abstention” on its own. Thus, learning by abstention provides an automatic mechanism for converting uncertainty per prediction into model decisions.
☕ Available resources
NA
🧭 Context of use
Hazard identification
— — —
In Regulatory Toxicology and Pharmacology by Lowe K et al
✨ Motivation
Human health risks from chronic exposures to environmental chemicals are typically estimated from potential human exposure estimates and dose-response data obtained from repeated-dose animal toxicity studies.
🧠 Study aim
A review on 1) the resources and methods available to predict human exposure levels and the associated uncertainty and variability, and 2) the margin between predicted human exposure levels and the dose levels used in repeated-dose animal studies.
💡Key highlights
  • Multiple case studies, ranging from those based on extensive monitoring data to those that require limited exposure information, are presented to demonstrate how human exposure levels could be estimated and how they compare to doses used in animal toxicity studies to help our understanding on the use of a TK-based approach in top dose selections.
  • There are a number of exposure assessment resources available, including monitoring data and exposure models, with varying degrees of sophistication, complexity, and uncertainty for estimating exposure that is fit for purpose.
  • The expected human exposures to environmental chemicals are typically orders of magnitude lower than no-observed-adverse-effect levels/lowest-observed-adverse-effect levels (NOAELs/LOAELs) when available.
☕ Available resources
NA
🧭 Context of use
Human risk assessment based on repeated-dose animal studies
— — —
In Archives of Toxicology by Firman JW et al
✨ Motivation
A lack of confidence in the general applicability of NAM continues to exist. The authors postulate that there are at least two fundamental reasons behind the persistence of this sense of scepticism: potential flaws within the methodologies themselves, and concerns regarding the lack of a means towards validation.
🧠 Study aim
Exemplification of a “tiered assessment” approach, whereby evidence gathered through a sequential NAM testing strategy was exploited so to infer the properties of a compound of interest. Rat acute oral lethality was chosen as endpoint and Bayesian inference as a methodology to enable quantification of the degree of confidence that a substance might ultimately belong to one of five LD50-associated toxicity categories.
💡Key highlights
  • Three assessment tiers were formulated: Cramer classification (tier 0), in silico that included Random Forest, structural alerts and EPA TEST LD50 (tier 1), in vitro cytotoxicity (tier 2), and in vivo experimental outcome (tier 3).
  • The tiered predictive approach was subsequently applied to a representative selection of 50 compounds from out of the 8186 forming the working dataset. Each of the 50 compounds selected to form a balanced, representative sample of the larger dataset were passed through the series of assessment corresponding to the tier.
  • Bayesian predictive models were constructed based upon proportional odds logistic regression (POLR).
  • Analysis revealed that the introduction of a second tier of assessment (in silico: QSARs and Random Forest) imparted substantial improvement in the balance of category assignments, relative to that obtained from the Cramer classification.
  • By contrast, the in vitro cytotoxicity data contributed little significant information when integrated within the tiered approach—producing only a very moderate redistribution in category assignment.
  • The tier 3 in vivo would be considered necessary only if there was large uncertainty with respect to the final verdict, or there existed an imperative to otherwise challenge the assigned category.
☕ Available resources
Supplementary materials available
🧭 Context of use
Risk assessment
— — —
In Archives of Toxicology by Burgoon LD et al
✨ Motivation
The kinetically derived maximal dose (KMD) provides a toxicologically relevant upper range for the determination of chemical safety.
🧠 Study aim
To describe a new way of calculating the KMD that is based on sound Bayesian, theoretical, biochemical, and toxicokinetic principles, that avoids the problems of relying upon the area under the curve (AUC) approach that has often been used.
💡Key highlights
  • The approach proposed uses toxicokinetic data and is based on Michaelis–Menten mechanics, and then, a mathematical analysis called the “kneedle” algorithm is used to identify the point of “diminishing returns”—the point at which the change in slope clearly demarcates the curve being nearly indistinguishable from the asymptote. The region within which that point lies, also known as a knee or elbow in the curve, is the KMD.
  • Also, the approach applies Bayes Theorem to obtain a final (posterior) distribution of plausible Vmax and Km values by adjusting observed data based on prior knowledge.
  • The advantage of defining KMD as a region, rather than as an inflection point along the curve, is that a region reflects uncertainty and clarifies that there is no single point where the curve is expected to “break;” rather, there is a region where the curve begins to taper off as it approaches the asymptote (Vmax in the Michaelis–Menten equation).
☕ Available resources
NA
🧭 Context of use
Risk assessment, toxicokinetics

 

Did you enjoy this issue?
CompTox Team

Hi there 👋 and Welcome to this little virtual corner ✨
Our aim is to curate peer-reviewed publications that showcase novel strategies or good case studies of in silico models useful in drug discovery or chemical safety assessment. With this, we hope to help keep our readers updated on Computational Toxicology but also provide learning resources for people new to the area.
If you have recommendations or want to join this initiative, please get in touch with us at comptox.newsletter@gmail.com. You can also leave your feedback at https://forms.gle/ikdyXMLQM5gR82nH6. We will always be happy to hear from you!

In order to unsubscribe, click here.
If you were forwarded this newsletter and you like it, you can subscribe here.
Powered by Revue