Hallucination is an effect that can occur when a generative AI tool learns non-existent patterns, due to noise or lack of relevant data. The resulting models generate products that are either clearly nonsensical, or, worse, plausible but actually inaccurate. In a drug design context this can manifest as predictions that overly rely on relationships in the data that are not physical or chemical in nature (e.g. drug activity is strongly correlated with the presence or absence of a particular element type in a molecule - more likely an indication of a lack of diversity in training data).

Synthetic accessibility of products is a pragmatic issue, but one that has serious implications for the development and testing cycle of drug candidates. A generative AI built around abstract representations of drug molecules can sample new molecules from a vast chemical space, but with very little chance of finding hits that can be made quickly or easily.

Our solution to these problems is (i) to deploy models based on physically meaningful representations of molecules (e.g. shape and charge, as opposed to atom counts or SMILES strings) and (ii) to develop a generative method for drug candidates which is constrained to produce molecules from in-stock compounds using known reactions - still a vast chemical space, but one which is readily accessible.

To assess the mutagenic and carcinogenic potential of different N-nitrosamines lacking robust carcinogenicity data, several approaches are in use including the published carcinogenic potency categorization approach (CPCA), the Enhanced Ames Test (EAT), in vivo mutagenicity studies as well as read-across to analogue molecules with robust carcinogenicity data.

We employ quantum chemical calculations (adapted from the work published by Wenzel et al. in 2022) as a pivotal tool providing insights into the likelihood of reactive ion formation and subsequent DNA alkylation for a selection of molecules including e,g., carcinogenic N-nitrosopiperazine (NPZ), N-nitrosopiperidine (NPIP), together with N-nitrosodimethylamine (NDMA) as well as non-carcinogenic N-nitrosomethyl-tert-butylamine (NTBA) and bis(butan-2-yl)(nitros)amine (BBNA). In addition, a series of nitroso-methylaminopyridines is compared side-by-side.

We draw comparisons between calculated reaction profiles for structures representing motifs common to NDSRIs and those of confirmed carcinogenic and non-carcinogenic molecules with in vivo data from cancer bioassays.

Furthermore, our approach enables insights into reactivity and relative stability of intermediate species that can be formed upon activation of several nitrosamines. Most notably, we reveal consistent differences between the free energy profiles of carcinogenic and non-carcinogenic molecules. For the former, the intermediate diazonium ions mostly react, kinetically controlled, to the more stable DNA adducts and less to the water adducts via transition-states of similar heights. Non-carcinogenic molecules yield stable carbocations as intermediates that, thermodynamically controlled, more likely form the statistically preferred water adducts. In conclusion, our data confirm that quantum chemical calculations can contribute to a weight of evidence approach for the risk assessment of nitrosamines.

Vicent Rodeschini, Hugues Lemoine & Didier Roche

Despite the re-emergence of combinatorial approaches to explore considerably larger chemical spaces, notably with DNA encoded libraries¹ (DEL) and affinity selection-mass spectrometry² (AS-MS), compound collections used for Hit-identification tackle a very small proportion of the 10⁶⁰ estimated size of the drug-like chemical space.³ Just looking at the diversity of drugs entering the clinic each year, one can easily understand that this diversity can hardly be matched with synthetic compound collections. New possibilities are offered today with generative machine learning⁴ or the use of non-enumerated massive combinatorial chemical spaces⁵ to bias compound design when structural information is available.

Since its inception 18 years ago, Edelris has designed and explored a diverse and biorelevant area of the chemical space, the Keymical Space^TM, comprising innovative synthetic NP-like, sp³-rich topologies.⁶

In our hand, this unbiased tangible chemical space has proved to be very successful for the rational design of innovative target modulators. We will share in this presentation our success stories as well as the limitations encountered in applying virtual screening towards hit identification.

1. Kunig, V., Potowski, M., Gohla, A. & Brunschweiger, A. DNA-encoded libraries – an efficient small molecule discovery technology for the biomedical sciences. Biol. Chem. 399, 691–710 (2018).

2. Prudent, R., Annis, D. A., Dandliker, P. J., Ortholand, J.-Y. & Roche, D. Exploring new targets and chemical space with affinity selection-mass spectrometry. Nat. Rev. Chem. 5, 62–71 (2020).

3. Reymond, J.L. The Chemical Space Project. Acc. Chem. Res. 48, 3, 722–730 (2015).

4. Meyers, J., Fabian, B. & Brown, N. De novo molecular design and generative models. Drug Discov. Today 26, 2707–2715 (2021).

5. Hoffmann, T. & Gastreich, M. The next level in chemical space navigation: going far beyond enumerable compound libraries. Drug Discov. Today 24, 1148–1156 (2019).

6. Nelson, A. & Roche, D. Innovative approaches to the design and synthesis of small molecule libraries. Bioorg. Med. Chem. 23, 2613 (2015).

Additional reading:

• An Algorithm to Identify Functional Groups in Organic Molecules, P.Ertl, J. Cheminf. 9:36 (2017).
• A Systematic Cheminformatics Analysis of Functional Groups Occurring in Natural Products, P. Ertl and T. Schuhmann, J. Nat. Prod. 82, 1258-1268 (2019).
• The Most Common Functional Groups in Bioactive Molecules and How Their Popularity Has Evolved over Time, P. Ertl, E. Altmann and J.M. McKenna, J. Med. Chem. 63, 8408-8418 (2020).