When building structure-activity relationships (SARs) in early drug discovery, in many cases, the synthesis of compounds with novel and complex structures remains the rate-limiting step. However, the fast exploration of SAR is a cornerstone to improving the pharmacological activity and physicochemical properties of the lead structures. The late-stage functionalization (LSF) of complex molecules could potentially help to overcome synthetic bottlenecks, thereby contributing to increasing the efficiency of the overall drug discovery process.

The LSF concept has emerged as a powerful and step-economical approach to rapidly explore closely related chemical space by functionalizing C-H bonds in complex drug molecules. Only a few applications of LSF in drug discovery, mostly focusing on a single LSF reaction type, have been reported in the literature. Various functional groups and multiple types of C–H bonds with different bond strengths as well as electronic and steric properties pose challenges for straightforward LSF. Consequently, running a successful LSF campaign often requires time-consuming and resource-intensive experimentation.

To overcome this issue, we have developed a geometric deep learning approach connected with automated, high-throughput miniaturized reaction screening for identifying late-stage hits and lead diversification opportunities. Our novel, interdisciplinary approach harnesses the power of chemical data. Among others, we identify representative Medicinal Chemistry substrate examples through structural, agglomerative clustering or capturing reaction data using the comprehensive, machine-readable simple user-friendly reaction format (SURF).

Case studies with a focus on borylation and Minisci-type alkylation reactions have confirmed the practical applicability of our geometric deep learning platform with laboratory automation for bioorganic and medicinal chemistry. The designed and developed graph transformer neural networks (GTNNs) that learn on three-dimensional (3D) molecular structures delivered accurate predictions for binary reaction outcome, reaction yield, and regioselectivity. The models have demonstrated successful prospective applications for the identification of suitable substrates through in silico reaction screening.

Our approach is now routinely and successfully applied to assess binary reaction outcome, reaction yield, and regioselectivity of LSF transformations in drug discovery projects at Roche.

David Nippa

Appearances