Computational Advances in Structural Biology

Abstract:

The continuous evolution of artificial intelligence (AI) algorithms has been affecting multiple scientific and industrial fields, and drug discovery (DD) and development are not an exception. We present a case study of how deep learning-based (DL) approaches can be utilized at different stages of a real-life project with a particular focus on applying structural models predicted with AlphaFold (AF).

Despite outperforming other homology modeling techniques in the CASP14 challenge, our analysis of multiple protein models generated with AF revealed that this approach should be used cautiously for DD purposes. From this perspective, an accurately reconstructed binding pocket is of the highest importance. In order to address the problem of inaccurately, or even mistakenly, generated backbone and side chain conformations, we developed a graph neural network-based (GNN) approach capable of finding inconsistencies in the ATP-binding sites of kinase structural models and then adjusting them accordingly.

In the case study presented herein, we successfully applied our end-to-end Pharma.AI platform and a structural model produced with AF. The team focused on protein targets without available crystal structures to show the applicability of AI-based methods to solve structural biology problems and accelerate the drug discovery process. Using the Pandaomics platform, CDK20 was identified as a promising therapeutic target for treating hepatocellular carcinoma (HCC). Then the AF model of CDK20 was exploited to generate a series of small molecules with the Chemistry42 platform. We synthesized proposed compounds, evaluated them in biochemical and cell-based assays, and identified novel highly active lead compounds. The case study highlights the potential of combining structural modeling and advanced generative AI to discover novel chemical matter targeting dark target space.