Using Bioisosteric Transformations for Compound Prediction
Bioisosteres are functional groups which have similar physical or chemical characteristics and hence similar biological effects.1 The principle of bioisosterism is widely used in drug discovery, where it is often desirable to maintain similar molecular interactions to retain target potency, while accessing new chemical space. This may be required in order to overcome issues related to absorption, distribution, metabolism, elimination or toxicity (ADMET) properties or to find novel chemical matter to expand or avoid the scope of a patent.
There are many examples of bioisosteric replacements in the literature, but it is challenging to identify those that may be applicable to a chemistry of interest and assess the potential results of making a similar modification.2
A recent development in cheminformatics is the concept of “idea generation” to speed up the exploration of chemical space around a compound of interest.3,4 These approaches apply “medicinal chemistry transformations” to an initial compound to generate new, related compound ideas. These transformations typically represent generally applicable compound modifications commonly considered by medicinal chemists in compound optimization.
These two concepts can be combined to automatically find and apply bioisosteric replacements to generate novel compound structures that are likely to preserve the required biological activities. The properties of the resulting compounds can be predicted and the most promising ideas prioritized for further consideration using multi-parameter optimization (MPO), to identify those that are most likely to have a good balance of the properties required in a high-quality drug.5
Creating bioisosteric transformations
To generate a database of molecular transformations, we used the BIOSTER database, a compilation of 27,366 pairs of molecules with bioisosteric substructures, manually curated from the scientific literature. Each BIOSTER record is in the form of a pseudo reaction relating the two compounds, with a manually designated reaction center indicating the bioisosteric substructures.
A transformation was generated for each record, representing the atoms in the bioisosteric substructures in Daylight’s SMIRKS notation (which appropriate software can apply to initial molecules to suggest new ones) using a customized version of Digital Chemistry’s MOLSMART program. A number of challenges needed to be overcome to generate good transformations. For example, substituent groups around the substructures are not necessarily identical in both molecules, and equivalent substitution positions on the reactant and product sides had to be identified heuristically and mapped appropriately.
In addition, in order to minimize the number of promiscuous transforms that could be applied inappropriately and generate structures that do not “make sense” from a medicinal chemistry perspective, substitution was permitted only where there was a substituent on at least one side of the reaction, and atoms were restricted to the ring/chain and aliphatic/aromatic environment found in the original molecules.
Of the 27,366 records in the latest version of the BIOSTER database SMIRKS were successfully generated for 22,547 (82.4 percent).
A high-quality lead or drug candidate must achieve a balance of many, often conflicting properties, including potency and ADMET properties. MPO methods integrate data on multiple properties to efficiently identify chemistries that are most likely to achieve an appropriate balance for a drug discovery project’s therapeutic objective.5
In the examples presented herein, the compound ideas generated were prioritized using a probabilistic scoring algorithm that assesses the properties of a compound against the ideal property profile for the project.8 In this way, the profile reflects the acceptable trade-offs between different properties. A score is calculated for each compound reflecting its likelihood of achieving the ideal property profile, taking into account the uncertainty in the individual property predictions.
MPO methods can be applied to both experimental and predicted compound data, but this application prioritizes virtual compound ideas. Therefore, the prioritization must be based on predictions from in silico models.
Predictive application of bioisosteric transformations
In the following two retrospective examples, the bioisosteric transformations were applied using the Nova module4 and prioritized using in silico models of ADMET properties and the probabilistic scoring method provided by the StarDrop software platform.9
Lead optimization: dipeptidyl peptidase IV inhibitor
The BIOSTER transformations were applied to the lead compound from the project that resulted in the discovery of the anti-diabetic dipeptidyl peptidase IV (DPP IV) inhibitor alogliptin.10 This resulted in the generation of 230 compounds; illustrative results are shown in Figure 1. It is notable that the product shown in the center of Figure 1 is a close analogue of alogliptin.
Fast follower: histamine H1 receptor antagonist
Application of the BIOSTER transformations to the antihistamine drug azatadine yielded a total of 89 compounds, including pKi against the histamine H1 receptor, predicted using a QSAR model. Some illustrative results are shown in Figure 2 and it is notable that the product on the right represents the core replacement that led to the candidate compound Hivenyl.11
Bioisosteric transformations are an excellent source of new ideas for compound design, providing access to increased chemical diversity while maintaining a high likelihood of biological activity. Automatically applying bioisisteric transformations from a large database of precedented replacements enables efficient exploration of new chemical space in the search for new optimization strategies. This may result in a large number of new ideas, which can be prioritized to highlight those most likely to succeed against a project’s objectives. Furthermore, links to the primary literature, from which the transformations were derived, make it easy to follow-up the most interesting ideas to find synthetic routes and investigate the underlying biological data.
This approach can be applied throughout the drug discovery process, including expansion around initial hits, exploring scaffold hopping opportunities in lead optimization and patent protection.
- Brown N, editor. Bioisosteres in Medicinal Chemistry. Vol 54. Weinheim: Wiley-VCH; 2012.
- Ujváry I, Hayward J. BIOSTER: A Database of Bioisosteres and Bioanalogues. In: Brown N, editor. Bioisosteres in Medicinal Chemistry. Vol 54. Weinheim: Wiley-VCH; 2012. p. 55-74.
- Stewart K, et al. Drug Guru: a computer software program for drug design using medicinal chemistry rules. Bioorg Med Chem. 2006;14:7011-22.
- Segall MD, et al. Applying medicinal chemistry transformations to guide the search for high quality leads and candidates. J Chem Inf Model. 2011;51(11):2967–2976.
- Segall MD. Multi-Parameter Optimization: Identifying high quality compounds with a balance of properties. Curr Pharm Des. 2012;18(9):1292-1310.
- Daylight Chemical Information Systems Inc. Daylight Theory Manual. Available from: www.daylight.com/dayhtml/doc/theory/theory.smirks.html. Accessed on July 29, 2013.
- Digital Chemistry Ltd. Chemical Query Conversion. Available from: http://www.digitalchemistry.co.uk/prod_chemicalquery.html. Accessed on July 29, 2013.
- Segall M, et al. Beyond profiling: Using ADMET models to guide decisions. Chem & Biodiv. 2009;6(11):2144-2151.
- Stardrop. Optibrium. Available from http://www.optibrium.com/stardrop. Accessed on July 29, 2013.
- Feng J, et al. Discovery of Alogliptin: A Potent, Selective, Bioavailable, and Efficacious Inhibitor of Dipeptidyl Peptidase IV. J Med Chem. 2007;50(10):2297-2300.
- Janssens F, et al. Norpiperidine Imidazoazepines as a New Class of Potent, Selective, and Nonsedative H1 Antihistamines. J Med Chem. 2005;48(6):2154-2166.