How can I predict toxicity in drug discovery? 

Author

  • Barry Wong, PhD

The importance of predicting toxicity early in drug development 

Highlighting safety and toxicity early in the drug development process is critical. Advancing problematic compounds further along the pipeline can lead to devastating costs when a late-stage (clinical) failure occurs.  

During early drug discovery, the focus is on the key pharmacokinetic and toxicological (ADMET) traits: Absorption, Distribution, Metabolism, Excretion, and Toxicity. Early ADMET profiling identifies potential issues before significant resources are committed to development. This enables medicinal chemists to easily pivot direction and concentrate on compounds with a higher likelihood of success. 

While it is impractical to test every experimental combination, computational predictive models ease the manual burden by prioritising compounds according to their characteristics prior to synthesis and testing. As a result, this improves research efficiency by reducing costs and saving time in drug development and high-throughput screening. Additionally, it can provide early warning indicators of toxicity that enable researchers to modify and optimise chemical structures. 

Recent advances in machine learning and computational chemistry have improved the speed and accuracy of toxicity predictions. 

What computational approaches are available for toxicity prediction? 

Machine learning and QSAR models 

Machine learning (ML) centres on creating algorithms and statistical models that learn from data and identify patterns to make predictions.  

Quantitative structure–activity relationship (QSAR) models are used in drug discovery to predict biological activity and physicochemical characteristics from molecular structure.  

Modern QSAR modelling frequently leverages ML methods to build these predictive models. When implemented correctly, ML-driven QSAR modelling can be more accurate, more adept at handling complex and high-dimensional data, and scale more effectively to meet the demands of modern drug discovery. 

If you’re looking to get started, you can follow our handy step-by-step guide: How to train your first QSAR model. 

When trained on high-quality data, QSAR models are most effective for making broad predictions and automating routine processes. 

However, there are considerations you need to make to ensure a successful QSAR model, for example ensuring that your training dataset is of sufficient quality. To learn more, watch our on-demand webinar: How to build a better QSAR model. 

Read-across or similarity-based methods 

These techniques estimate the properties of a compound by referencing data from similar compounds. In other words, if you understand how Compound A acts and Compound B shares a structural or chemical resemblance, you can deduce or “read across” the expected behaviour of Compound B, even without any available direct data.  

The read-across approach works best when you have similar analogues and limited information on the target compound.  

Although this is a quick and cost-effective method, it demands a strong justification for similarity. It can fail if minor structural changes lead to significant variations in activity, often referred to as “activity cliffs”.  

Knowledge-based and rule-based systems 

Knowledge-based and rule-based systems, like Derek Nexus, leverage expert knowledge, logical rules, and sometimes statistical or ML models to predict toxicity. They depend on established rules that come from human expertise and known toxicological data. For instance, “If a molecule has an X group on an aromatic ring, it could be mutagenic”. Hybrid systems merge rule-based reasoning with ML or statistical models, combining various types of evidence to enhance prediction accuracy. 

Knowledge-based and rule-based systems are most valuable when a mechanistic understanding is crucial, and transparency is required. They are straightforward to interpret and comprehend. The rules are founded on robust scientific principles and established toxicological mechanisms, and they are accepted by regulatory bodies for use in safety assessments.  

However, there are some drawbacks. Insufficient rule coverage might miss new toxic mechanisms. Additionally, if the rules are either too broad or too narrow, there is an increased risk of false positives or negatives.  

Integrated hybrid approaches 

A hybrid approach that combines computational predictions with experimental validation can be an effective way to improve outcomes. In this context, a specific set of experiments are performed to validate the model, and the findings are used to refine future models while managing development costs. 

Integrating toxicity predictions into discovery workflows 

Toxicity predictions can be applied across multiple stages of the drug discovery workflow: 

  • Virtual screening: Identify and eliminate likely toxic compounds early to minimise wasted costs. 
  • Lead optimisation: Highlight potentially toxic moieties or metabolites that could be toxic and optimise compound design.  
  • Regulatory submissions: Use validated models to support documentation and streamline approval process.  

Summary  

Utilising computational techniques to forecast safety (and other characteristics) is beneficial in drug discovery. It minimises the necessity for experimental testing, aids in prioritising compounds for development or further investigation, and, in the right context, can bolster regulatory submissions.  

Forecasting toxicity in StarDrop 

Use StarDrop’s Derek Nexus module to predict over 40 toxicity endpoints within your multi-parameter optimisation strategy, to identify compounds with the optimal property balance before synthesis. 

Combined with the Metabolism module, it evaluates toxicity for both parent compounds and metabolites, providing complete safety assessment throughout metabolic pathways. 

Integrate these with the visual Glowing Molecule designer, to highlight the regions of your molecule that will have the greatest influence on toxicity and other key properties. 

About the author

Barry Wong, PhD

Barry, our Application Support Scientist, has over 15 years experience in academia and industry. His PhD in Molecular and Cellular Biology from the University of Dundee focused on the mechanisms of cell migration, metabolism and signal transduction. He enjoys problem solving for drug discovery, bioscience and healthcare professionals, and showing users new tricks to make the most of their StarDrop software.

Barry Wong, Application Support Scientist

More about toxicity prediction