Machine learning C-H activation reactions for efficient exploration of drug-relevant chemical space

Kenneth Atz, David F. Nippa, Alex T. Müller, Uwe Grether, Rainer E. Martin & Gisbert Schneider

The synthesis of novel chemical matter remains a crucial challenge in small molecule drug discovery, often being the critical bottleneck impacting time and cost. Especially C-H activation reactions on late-stage lead molecules often require high throughput experimentation (HTE) screening to identify suitable conditions and substrates. [1] Machine learning methodologies, especially methods that enable efficient learning on three-dimensional (3D) molecular structures, have proven instrumental in various domains of chemistry. [2,3] We demonstrate prospective applications of graph transformer neural networks (GTNN) to C-H alkylation and borylation reactions. We demonstrate the importance of 3D information in predicting regioselectivity and quantify the impact of electronic information on predictive accuracy (Figure 1). Moreover, we show how trained GTNNs enable efficient in silico library screening to identify suitable substrates. [5] The identification of synthesizable chemical matter with desired properties enables fast exploration of drug-relevant chemical compound space.

Figure 1. Left: Prospective application of regioselectivity prediction models to the drug nevirapine. Right: Performance of the investigated atomistic GNNs including 2D (blue) and 3D (orange) information.

1. D. C. Blakemore, et. al., Nat. Chem 2018, 10, 383–394.
2. O. A. v. Lilienfeld, K. R. Müller and A. Tkatchenko, Nat. Rev. Chem. 2020, 4, 347–358.
3. K. Atz, F. Grisoni and G. Schneider, Nat. Mach. Intell. 2021, 3, 1023–1032.
4. D. F. Nippa, K. Atz, et al., Nat. Chem. 2023, in press.
5. D. F. Nippa, K. Atz, et al., Communs. Chem. 2023, 6, 256.

 Kenneth Atz