Development of an In silico platform for the prediction of off-target binding

Human beings are constantly being exposed to a diverse array of chemical compounds, both intentionally and unintentionally. The rigorous assessment of chemical toxicity is therefore paramount for human health. Traditionally, evaluating small molecule-induced toxicity involves costly and time-consuming in vitro and in vivo tests. In many cases, toxic effects begin with off-target binding, an undesired interaction between a small molecule and a protein. In the pharmaceutical industry, off-binding assessment is often performed in late pre-clinical stages of drug development. However, neglecting off-target toxicity can lead to drug failure, resulting in significant financial loss and years of development time.
This thesis presents innovative computational tools designed for the early assessment of off-target liabilities. These tools offer a cost-effective and rapid alternative to traditional methods, enabling their application in the early stages of pharmaceutical development to guide the design of safe drugs. The thesis begins by examining the impact of dataset quality on deep learning-based predictions of drug-target interactions. It shows that correct data handling is critical and demonstrates how a detailed characterization of intermolecular interactions improves predictions. Leveraging this understanding, we introduce PanScreen, an online platform automating the prediction of off-target binding. PanScreen encompasses a portfolio of pharmaceutically and toxicologically relevant off-targets and provides qualitative and quantitative predictions, including binding poses and estimated affinities. To complement this structure-based approach, we developed a deep learning model that preserves molecular similarities in the form of Euclidean distances in latent space. This model enhances protein-structure-free screening of ultra-large databases, accelerating similarity-based searches by orders of magnitude. We show that it can be applied to different similarity metrics, including alignment-based 3D shape similarities.
These in silico tools hold promise for predicting off-target interactions in diverse applications, offering an inexpensive and fast option to complement traditional methods. Specifically, PanScreen represents a significant step in this direction. As the development of these tools is an ongoing process, we offer a roadmap that outlines avenues for further improvement, aiming to enhance their robustness and accuracy. Ultimately, we envision a future where chemical safety assessment is rapid, cost-effective, and does not involve animal testing.