Consensus rank orderings of molecular fingerprints illustrate the ‘most genuine’ similarities between marketed drugs and small endogenous human metabolites, but highlight exogenous natural products as the most important ‘natural’ drug transporter substrates

Abstract

We compare several molecular fingerprint encodings for marketed, small molecule drugs, and assess how their rank order varies with the fingerprint in terms of the Tanimoto similarity to the most similar endogenous human metabolite as taken from Recon2. For the great majority of drugs, the rank order varies very greatly depending on the encoding used, and also somewhat when the Tanimoto similarity (TS) is replaced by the Tversky similarity. However, for a subset of such drugs, amounting to some 10% of the set and a Tanimoto similarity of ~0.8 or greater, the similarity coefficient is relatively robust to the encoding used. This leads to a metric that, while arbitrary, suggests that a Tanimoto similarity of 0.75-0.8 or greater genuinely does imply a considerable structural similarity of two molecules in the drug-endogenite space. Although comparatively few (<10% of) marketed drugs are, in this sense, robustly similar to an endogenite, there is often at least one encoding with which they are genuinely similar (e.g. TS > 0.75). This is referred to as the Take Your Pick Improved Cheminformatic Analytical Likeness or TYPICAL encoding, and on this basis some 66% of drugs are within a TS of 0.75 to an endogenite.We next explicitly recognise that natural evolution will have selected for the ability to transport dietary substances, including plant, animal and microbial 'secondary' metabolites, that are of benefit to the host. These should also be explored in terms of their closeness to marketed drugs. We thus compared the TS of marketed drugs with the contents of various databases of natural products. When this is done, we find that some 80% of marketed drugs are within a TS of 0.7 to a natural product, even using just the MACCS encoding. For patterned and TYPICAL encodings, 80% and 98% of drugs are within a TS of 0.8 to (an endogenite or) an exogenous natural product. This implies strongly that it is these exogeneous (dietary and medicinal) natural products that are more to be seen as the 'natural' substrates of drug transporters (as is recognised, for instance, for the solute carrier SLC22A4 and ergothioneine). This novel analysis casts an entirely different light on the kinds of natural molecules that are to be seen as most like marketed drugs, and hence potential transporter substrates, and further suggests that a renewed exploitation of natural products as drug scaffolds would be amply rewarded.

Novel 'Housekeeping' Genes and an Unusually Heterogeneous Distribution of Transporter Expression Profiles in Human Tissues and Cell Lines, Assessed Using the Gini Coefficient

“…This illustrates the point that despite the fact that their substrates are almost uniformly available via the bloodstream, and biochemistry textbooks and wallcharts largely show this, they clearly use substrates differentially (ergothioneine and the SLC22A4 transporter being a nice example (Gründemann, 2012;Gründemann et al, 2005)). It also implies strongly that in many cases we do not in fact know the natural substrates, many of which are clearly exogenous (O'Hagan and Kell, 2017b).…”

“…An overall theme is certainly that we do often lack knowledge of the 'natural' substrates of even well-studied drug transporters (e.g. (Gründemann et al, 2005;O'Hagan and Kell, 2017c;Perland and Fredriksson, 2017)).…”

“…That transporters are also responsible for the uptake of pharmaceutical drugs and xenobiotics into cells, and their efflux therefrom (Colas et al, 2016;Dobson and Kell, 2008;Giacomini and Huang, 2013;Giacomini et al, 2010;Kell, 2015;Kell, 2016;Kell et al, 2013;Kell et al, 2011;Kell and Oliver, 2014;Lin et al, 2015;Stanley et al, 2009), means that to understand drug distributions we must understand transporter distributions. In many cases, we do not know either the 'natural' (O'Hagan and Kell, 2017c;Perland and Fredriksson, 2017) or the pharmaceutical drug substrates of these transporters, and one clue to this may be to understand their differential tissue distribution. Fortunately, we now have available a variety of expression profiles at the level of both the transcriptome (Fagerberg et al, 2014) and the proteome (Thul et al, 2017;Uhlén et al, 2015) that allow us to assess these.…”

VAE-Sim: a novel molecular similarity measure based on a variational autoencoder

“…As with any kind of system of this type, the 'closeness' is a function of the weighting of any individual features, and it is perhaps not surprising that the different fingerprint methods give vastly different similarities, even when judged by rank order (e.g. [25] and above). One similarity measure that is independent of any fingerprint encoding is represented by the maximum common substructure (MCSS).…”

“…Following training, each molecule (SMILES) could be associated with a normalised vector of 100 dimensions, and the Euclidean distance between them could be calculated. As previously [25], we compared the similarities between all drugs and all metabolites using the datasets made available in [91]. We here focus on just the MACCS and Patterned encodings of RDKit, and compare them with the normalised Euclidean distances according to the latent vector obtained from the VAE.…”