Molecular design and synthesis lie at the heart of medicinal chemistry. Identifying novel starting points for small molecule drug discovery campaigns often necessitates high-throughput screening of libraries containing thousands to millions of molecules. Once novel starting points are found, their diversification remains an additional critical bottleneck that impacts time and costs during molecular optimization. Deep learning methodologies, particularly those that enable efficient learning from three-dimensional (3D) molecular structures, have proven to be instrumental in various domains of chemistry [1,2]. Here, we present a computational de novo design approach (i.e., deep interactome learning) for the zero-shot construction of compound libraries tailored to possess specific bioactivity, physicochemical properties, synthesizability, and structural novelty. We demonstrate how prospective application of trained deep interactome learning models yields the identification of potent and selective designs on unseen protein targets [3]. Second, we illustrate the applications of graph neural networks (GNNs) to complex reactions. We show how trained GNNs achieve high accuracy in regioselectivity prediction and in silico substrate screening [4]. Unifying the developed forward reaction prediction models with high-throughput experimentation (HTE) enables rapid structural diversification and exploration of structure-activity relationships (SAR) [5]. The successful outcomes of these studies demonstrate how advanced deep learning methods combined with miniaturised, automated experiments creating a self-improving “lab-in-the-loop” for small-molecule drug discovery.

References

[1] K. Atz, F. Grisoni and G. Schneider, Nat. Mach. Intell., 3, 1023–1032 (2021).

[2] C. Isert, K. Atz, and G. Schneider, Curr. Opin. Struct. Biol., 79, 102548 (2023).

[3] K. Atz, et. al. Nat. Commun., 15, 3408 (2024).

[4] D. F. Nippa, K. Atz, et al. Nat. Chem., 16, 239-248 (2024).

[5] D. F. Nippa, K. Atz, G. Schneider et. al. ChemRxiv (2025).