The Use of Artificial Intelligence to Support Safety Predictions

Last year, we were happy to contribute to a new book describing the use of both supervised and unsupervised machine learning approaches to support the prediction of toxicity and is illustrated with case studies related to regulatory submissions.

The book chapter emphasized the importance of high-quality databases, containing toxicology and chemistry information. Such comprehensive databases are critical for both the performance and acceptance of the results and require the use of ontologies and controlled vocabularies to harmonize and integrate the data from multiple sources in a meaningful manner.

Alongside this, the chapter highlighted the importance of coupling effective machine learning approaches with chemical descriptors to ensure any assessment is transparent and insights are actionable, including any expert review of the results.

Another aspect underscored was the importance that any predictive model is fit-for-purpose and its use and performance documented (such as how such models adhere to OECD validation principles).

This book is currently being finalized and will be released soon.

If you would like to discuss this topic in more detail then do get in touch with myself via (Glenn Myatt;

Computational toxicology 2022 year-end review

The computational toxicology group here at Instem has had another busy year that has resulted in six publications and book chapters1-6, completion of existing and initiation of new collaborative working groups, as well as significant updates to our computational toxicology solutions.

We have been collaborating on a number of research topics related to the in silico assessment of extractables and leachables (E&L). This work is led by Dr. Candice Johnson at Instem and includes (1) a chemical analysis of a large databases of E&L reported by the ELSIE and PQRI organizations, (2) an evaluation of how sensitization in silico models support E&L assessments, and (3) an analysis of whether in silico approaches could be used as a screen to assess whether a leachable may react with a biomolecule API (such as a protein) resulting in possible safety, efficacy or quality concerns. This work was presented at the E&L Europe conference7 and three manuscripts have been submitted for publication.  

This year, we also initiated a new collaborative working group to develop a framework and a protocol to support the recently published addendum to the ICH S1B guideline. This addendum describes six weight of evidence factors that could be used as part of a carcinogenicity assessment to determine whether performing a rat study adds value. Dr. Arianna Bassan, who is coordinating this work, recently presented the project at the American College of Toxicology (ACT) meeting in Colorado.8

Progress is being made on a series of collaborative working groups related to N-nitrosamines, which are being led by Dr. Kevin Cross at Instem. This work has resulted in a new publication around the use of the Ames test2 and a series of deliverables on the Mutamind project, funded by the EMA9. He also chaired a workshop on addressing the global challenge of nitrosamine impurities in pharmaceutical drugs at the 2022 ACT meeting.8 This year we also introduced a new Predict™ service10 supporting carcinogenicity risk assessment of N-nitrosamines.

2022 also saw the completion of a Special Issue of the Journal of Computational Toxicology on the in silico toxicology protocol project, which included eight publications and an editorial authored by Instem.11

Earlier this year, we released a new version of the Leadscope Model Applier that includes some significant developments, including a new read-across tool. This application provides access to over 200,000 chemicals and 600,000 toxicology studies through an easy-to-use interface and is aligned with regulatory expectations. A new SaaS (Software as a Service) option for the Leadscope technology was also released. In addition, we introduced a series of new and updated expert alerts and QSAR models to assess a complete battery of acute toxicity endpoints12, updates to the bioactivation models to support the assessment of drug-drug interactions, version 9 of the expert alert for bacterial mutagenicity, as well as new endocrine activity QSAR models. Furthermore, through Instem’s Research Collaboration Agreement with the US FDA, we have updated the abuse liability assessments with a new QSAR model predicting blood-brain barrier penetration,1 as well as updated the cardiotoxicity models13.

We would like to thank all our collaborators, customers, partners, and colleagues for all their help and support over the last year and look forward to continuing our work together to further the acceptance and use of in silico approaches and support the 3Rs.

We all at Instem wish you a Happy Holidays!

Please get in touch with me (Glenn Myatt; if you are interested in discussing in silico approaches or any of the collaborative working groups.


  1. Faramarzi, S. et al., (2022) Development of QSAR models to predict blood-brain barrier permeability, Front. Pharmacol.
  2. Trejo-Martin, A. et al., (2022) Use of the Bacterial Reverse Mutation Assay to Predict Carcinogenicity of N-Nitrosamines, Regulatory Toxicology and Pharmacology, 105247.
  3. Zwickl, C. et al., (2022) Principles and Procedures for Assessment of Acute Toxicity Incorporating In Silico Methods, Computational Toxicology, 100237.
  4. Crofton, K. et al (2022) Current status and future directions for a neurotoxicity hazard assessment framework that integrates in silico approaches, Computational Toxicology, 22, 100223.

Update on the Mutamind N-nitrosamine project

Earlier this year, it was announced that Instem was part of the Mutamind industry and academic consortium of 9 beneficiaries, funded by EMA and led by Fraunhofer Institute for Toxicology and Experimental Medicine (ITEM)1. Instem is collaborating on this 2-year research project to further investigate the mutagenicity of different classes of N-nitrosamines to distinguish highly potent from less potent carcinogens.

The project is making progress and a number of the deliverables have been recently published,2,3,4 In addition, Dr. Kevin Cross has recorded a presentation describing the project.5

If you would like to discuss this project in more detail, please get in touch with Dr. Kevin Cross (


  2. The QSAR project protocol (QSAR for Nitrosamines EUPAS 46057) is available at:
  3. The in vitro project protocol (In Vitro NA Mutagenicity EUPAS 49355) is available at:
  4. The endogenous formation project protocol (GITox EUPAS 49089) is available at:

New paper discussing the assessment of abuse liability using blood-brain barrier QSAR models

Last year, we discussed the development of a new solution to support the assessment of abuse liability.1 This included a large database of over 4,000 chemicals from numerous sources including the US Drug Enforcement Agency (DEA) scheduled drugs and structural alerts to profile potential abuse liability.

This assessment has now been extended to include a new QSAR model to predict drug permeability across the blood-brain barrier. This model includes a training set of 921 chemicals with rodent data and was developed by the US FDA (under a Research Collaboration Agreement).

A research article titled “Development of QSAR models to predict blood-brain barrier permeability” has just been published describing how this model was developed and validated.2

Please get in touch if you are interested in discussing this approach to assessing abuse liability (Glenn Myatt;



November 2022 computational toxicology conferences

We have been busy preparing for a number of upcoming conferences in November1.

At the 2nd International Akademie Fresenius Gene-Tox Conference (November 3, 2022), Kevin Cross will discuss “N-Nitrosamine Impurities in Drugs – Introducing the EMA-Mutamind Project”.

As part of the Extractables & Leachables Europe 2022 Conference (November 7 – 8, 2022) Glenn Myatt will make a presentation on ‘Emerging applications of computational methods in the assessment of extractables and leachables”.

In addition, at the American College of Toxicology’s 43rd Annual Meeting (November 13 – 16, 2022), we will be presenting on “Strategic In Silico: Creating Powerful New Scientific Insights” (Tuesday, November 15, 12.00-12.55pm) as well as participating in Symposium 1 “ICH S1B(Rq): New Approaches in the Carcinogenicity Assessment of Pharmaceuticals” (Monday 14th November 9-12pm MST) and Workshop 4 “Addressing the Global Challenge of Nitrosamine Impurities in Pharmaceutical  Products” (Monday 14th November, 2-5pm MST).

Please let me know if you are interested in discussing any of these presentations (Glenn Myatt;



Acute Toxicity in silico Battery

We are happy to have completed, after many years of work, a full battery of in silico models to evaluate acute toxicity. These models predict acute lethality by three routes of exposure (oral, dermal, inhalation), as well as skin and eye irritation/corrosion, and skin sensitization. They aim to support a variety of applications, including health, safety, and environmental product stewardship.

The underlying databases have been constructed to support the development of the computational models and include 175,885 chemicals with lethality data, 3,775 chemicals with skin irritation/corrosion data, 6,619 chemicals with eye irritation/corrosion data, and 2,188 chemicals with skin sensitization data. In addition, overall grades have been assigned based on the reported studies for each chemical aligned with international standards such as the Globally Harmonized System of Classification and Labeling of Chemicals (GHS)1 or the European Centre for Ecotoxicology and Toxicology of Chemicals (ECETOC)2.

It is currently considered best practice to use multiple computational methodologies3 – expert rule-based and statistically-based. Hence, both methodologies have been made available in this battery to predict the six different graded endpoints (i.e., GHS or ECETOC classifications). Many of the expert-rule based methods incorporate knowledge from the literature, such as the BfR rules for skin and eye irritation/corrosion4.

The models support a thorough expert review by providing access to the information behind the prediction, such as chemical analogs and, whenever possible, mechanistic information detailed as part of the expert rules. In addition, the predictions are fully integrated with the new read-across tool5 which might be used to complement the predictions.

In a recent study based on proprietary data, the initial version of the rat oral acute models was able to predict 95% of the chemicals as either correct or in a more conservative category6.

We would like to thank all our colleagues in India, Italy, UK and USA for all their hard work making this important milestone in in silico toxicology possible.

If you would like to discuss these new models, please contact me (Glenn Myatt;


  1. Globally Harmonized System of Classification and Labelling of Chemicals (GHS) (“The Purple Book”), United Nations, 2005 First Revised Edition, available at
  2. ECETOC 2008. Potency Values the Local Lymph Node Assay: Application to Classification, Labelling and Risk Assessment ECETOC Document No. 46 Brussels, December 2008
  3. Myatt GJ, et al. (2018) In silico toxicology protocols. Regul Toxicol Pharmacol 96:1–17
  4. Tsakovska, I., et al., 2007. Evaluation of SARs for the prediction of eye irritation/corrosion potential–structural inclusion rules in the BfR decision support system†. SAR and QSAR in Environmental Research 18, 221–235.
  6. Bercu, J., et al., 2021. A cross-industry collaboration to assess if acute oral toxicity (Q)SAR models are fit-for-purpose for GHS classification and labelling. Regul. Toxicol. Pharmacol. 120, 104843.

New read-across tool

Over the last year we have been working hard developing a new easy-to-use read-across solution that can instantaneously access over 200,000 chemicals with over 600,000 toxicity studies. The platform is able to integrate this data with chemicals and studies added by the user to conduct a read-across evaluation.

It has been designed to directly support a variety of use cases including regulatory submissions (e.g., assessment of pharmaceutical impurities and extractables & leachables), occupational toxicology, classification and labelling, and product discovery.

It is aligned with regulatory expectation to identify and justify the use of chemical analogs based on their similarity that can be analyzed from different perspectives: (1) structural similarity (global and local), (2) physico-chemical similarity, and (3) biological similarity. The read-across is performed using a variety of search and visualization options that have been integrated within the tool. These tools are coupled with direct access to curated and graded data, methodologies for predicting relevant toxicological endpoints, and capabilities for predicting the underlying mechanism using profiling tools. This allows for a smooth integration of QSAR predictions in your read-across study when needed.

At the heart of the tools is a data matrix where all information is collected alongside functionalities to interrogate the results, such as explaining predictions. This simple data matrix can also be shared and reports generated by directly exporting to Excel.

We would like to thank all the beta testers for their extremely helpful suggestions to improve the tool.

If you would like to discuss this new tool, please contact me (Glenn Myatt;

ELSIE Webinar on a Framework for Sensitization Assessment for E&L and Practical Application

We are delighted to be participating in an upcoming webinar organized by the Extractables and Leachables Safety Information Exchange (ELSIE) Consortium on the topic of a framework for sensitization assessment for Extractables and Leachables (E&L) on 22 September 2022, 9:30 – 11:00 AM ET1.

We will be describing the development of a representative database of E&L from the combined ELSIE Consortium and the Product Quality Research Institute (PQRI) published datasets. An analysis using chemical structure-based clustering will be outlined to show the similarities and differences between the two datasets. The presentation will also describe classes containing chemicals that were flagged as potentially mutagenic as well as potent (strong or extreme) dermal sensitizers by the Leadscope in silico tools.

Please get in touch if you would like to discuss this work in more detail (Glenn Myatt;


1.            Home | ELSIE (

New paper on the use of the Ames assay to predict carcinogenicity of N-Nitrosamines

This week, we are pleased to welcome Dr. Kevin P. Cross, Instem’s Principal Investigator with U.S. FDA Collaborations and VP of Product Engineering and Production, as a guest contributor to the blog.

We were happy to contribute to a recent publication titled “Use of the Bacterial Reverse Mutation Assay to Predict Carcinogenicity of N-Nitrosamines”.1

N-Nitrosamines (NAs) are a cohort of concern in the ICH M7 guideline2 and have recently come under increased regulatory scrutiny, mainly because of their potential to be highly potent mutagenic carcinogens in rodent bioassays.

This publication performed an analysis of public databases with both bacterial reverse mutation (Ames) assay and rodent carcinogenicity data, and the sensitivity of the Ames assay to predict the carcinogenic outcome was examined. The paper demonstrated that the Ames assay conducted under OECD 471 guidelines is highly sensitive for detecting the carcinogenic hazards of NAs.   

Please feel free to contact me (Kevin Cross; if you would like any additional information.


  1. Alejandra Trejo-Martin et al., 2022, Use of the Bacterial Reverse Mutation Assay to Predict Carcinogenicity of N-Nitrosamines, Regulatory Toxicology and Pharmacology, 105247.
  2. ICH, M7 (R1) Assessment and control of DNA reactive (mutagenic) impurities in pharmaceuticals to limit potential carcinogenic risk, European Medicines Agency, 2017.

Publication of the special issue of computational toxicology on the in silico toxicology protocol initiative

In 2016, we initiated a new project to develop a standardized procedure for in silico experts performing assessments, similar to in vitro or in vivo test guidelines. This would streamline the application of in silico methods, ensure best practices are adopted, and defend its use to colleagues, peers, and regulators.

The general framework paper1 was published in 2018, followed by a protocol for genetic toxicology2 in 2019, and a protocol for skin sensitization3 in 2020.

Since then, there has been considerable activity developing additional position papers and protocols, and much of this work is now included in a special issue of computational toxicology “In silico toxicology protocol initiative”4. The special issue includes 8 papers along with an editorial.

This body of work reflects an enormous effort on the part of the hundreds of collaborators on this project and will support the further acceptance and routine use of in silico methods at the same time as avoiding unnecessary animal studies.

We wish to thank all the collaborators and congratulate them on the completion of this special issue.

If you are interested in discussing this project in more detail, please contact me (Glenn Myatt;


  1. G.J. Myatt et al., In silico toxicology protocols, Regul. Toxicol. Pharmacol. 96 (2018) 1–17.
  2. C. Hasselgren et al., Genetic toxicology in silico protocol, Regul. Toxicol. Pharmacol. 107 (2019) 104403.
  3. C. Johnson et al., Skin sensitization in silico protocol, Regul. Toxicol. Pharmacol. 116 (2020) 104688.