What is Chemical Similarity, and how is it Useful? (by Edmund Duesbury)

In the final year of my PhD, I have been investigating different forms of alignment of chemicals, and seeing which method is best at predicting whether a chemical will be active against a particular drugtarget.

Similarity is subjective to a particular problem domain.  As an example, which two most objects are most similar – an apple, a pumpkin or a basketball?  All three are more or less spherical, but the pumpkin and apple have the similarity of being fruit, while the pumpkin and basketball are a similar size.

The same subjectivity exists in chemistry.  A common goal when searching for similarity in chemicals is to predict whether one compound will act in the same way as another compound, known to have useful pharmaceutical properties.  The desired “similarity” in this case, is a similarity of biological activity: something which, at present, is impossible to predict.  However, we can attempt to infer such a property from aspects of structural similarity.

Serotonin reuptake inhibitors is a group of chemicals that includes many useful antidepressants. Consider the examples below, of compounds that act as serotonin reuptake inhibitors.  In the first case (Figure 1), similarity is based on the largest common fragment (highlighted in bold).

Figure 1

The similarity here is obvious, the only difference being the Br atom.  However, the same technique fails to show the biological similarity between the two inhibitors in Figure 2.

Figure 2.

In this case, the approach of finding the largest fragment has failed to highlight the “similarity” between the two compounds.  A technique based on finding the maximum possible overlap of edges however, is more successful (Figure 3).

Figure 3.

This method, which seeks to find a set of common fragments emphasises a different “similarity” between these compounds.

Information Retrieval (but not as we know it)

Last month, James Wallace bravely volunteered to lead a discussion based on his project work.  In so doing, he proved the title of this post to be wrong.

He is the second member of the chemoinformatics group to post on this blog (after Ben Allen) and he did an excellent job of helping the rest of us to understand the kind of work the group does.

There are occasional signs of bemusement amongst members of the iSchool about the position of chemoinformatics.  It is not always obvious how its work relates to that of others in the school who deal with the organisation of libraries, and the management of information in businesses.  But James began his talk with a few words that fitted chemoinformatics firmly at the heart of the iSchool: words such as 'classification', 'organisation', 'information retrieval' and 'database'.

It quickly became clear that he was using tools with which we were all familiar, but applying them to a subject that most of didn't understand (the classification of molecular structures).  What was interesting though, was the extent to which the nature of information retrieval (IR) he was discussing seemed historical.  Anyone involved with IR pre-Google would have found themselves on familiar territory.  One of the questions raised was whether or not a Google-type approach would be relevant to chemoinformatics.  The feeling was that it probably wouldn't - in part because it may be necessary to have a critical mass of users before a system such as PageRank could be useful.  I don't know whether or not that's true, but if it is, it would be interesting to know what the critical mass might be, and what factors would cause it to vary.

"Using network and graph theory to influence drug design" (by J.E.A. Wallace)

Chemoinformatics (also known as ‘cheminformatics’) applies computational and information science techniques to a range of chemical problems.  These techniques are used, for example, to model compounds and to predict their properties.  Chemoinformatics has made it possible to simulate expensive operations such as high-throughput screening (the rapid testing of numerous compounds for desirable qualities).  A good starting point for those new to the field is the book ‘An Introduction to Chemoinformatics’ by Leach and Gillet.

My project is specifically concerned with the idea of de novo design (Latin for ‘from new’).  De novo design builds novel molecules in virtual space and evaluates them against some measure of functionality, such as similarity to a compound known to work well, or the complementarity between the shape of the molecule and the receptor it is to fill. The aim is to streamline the drug discovery process.  By referring to these virtual compounds, it may be possible to reduce the number of synthesis operations necessary. However, the sheer number of potential molecules that can be made renders it impossible to explore all of chemical space for solutions, even with modern levels of computing power. As a result, various predictive sampling processes (such as the Monte Carlo method) have been adapted for use in this field. These can give good results, but their simplistic approach to constructing molecules can lead to products being generated that are impossible to fabricate in real world situations, limiting the overall usefulness of the methods.

In order to work around such problems, methods have been developed within the chemoinformatics group that use genuine reaction data from literature to create generic rules (the reaction vectors).  These rules can be applied to a given starting material to generate new molecules. This provides a compromise between finding novel molecules and retaining a synthetic awareness, as each transformation is based on a literature precedent. However, this method has its own limitations.  Some structures for example, can be built up using multistep reaction sequences. In many cases, the intermediate steps in such sequences do not resemble either the end product or the starting material in terms of similarity to the target.  Consequently, many potentially useful molecules are never completed because steps en route to the end product score poorly.

One solution to this problem is to create a new rule format that represents all the reaction steps in one operation. This requires some initial identification of those sequences, which is where my project comes in. If we regard a chemical reaction as a transformation of one molecule into another, it becomes possible to connect these transformations together into a network.  Every path through the network will represent a multistep synthetic route for which each step has a known example in the literature.  Reaction properties such as ease of synthesis, cost of materials, yield of product etc. can then be added to the network to bias the selection of routes in accordance with a particular set of criteria.