Digital twin approaches for interpretable side effect prediction in drug discovery

Area of Science:

• Computational chemistry
• Pharmacology
• Drug discovery

Background:

• Artificial intelligence (AI) is crucial in preclinical drug development, but predicting drug side effects early remains challenging.
• Current AI methods often rely on expert features or 'black box' models, limiting their interpretability and applicability in early discovery.
• Accurate prediction of drug side effects is essential for safe and efficient drug development.

Purpose of the Study:

• To propose a novel AI paradigm for predicting drug side effects by utilizing biologically meaningful intermediate representations.
• To enhance the interpretability and actionability of AI models in early-stage drug safety assessment.
• To provide a framework for informing rational polypharmacology and guiding secondary pharmacology assays.

Main Methods:

• Developing an AI approach that predicts intermediate biological representations, such as off-target proteins and their downstream cellular effects.
• Simulating these downstream effects within a cellular digital twin.
• Training simple, interpretable models on these biologically derived representations instead of direct chemical-to-side effect mapping.

Main Results:

• The proposed method generates interpretable and actionable insights into potential drug side effects.
• This approach moves beyond 'black box' models, offering a clearer understanding of predicted safety concerns.
• The biologically-informed representations facilitate a more nuanced safety assessment.

Conclusions:

• This AI paradigm shift offers a more interpretable and actionable method for early drug side effect prediction.
• The approach has the potential to improve drug safety assessments and guide drug design strategies.
• It can also inform polypharmacology and optimize the design of pharmacology assays.