Escape from flatland 2: complexity and promiscuity.

Lovering.  Med. Chem. Commun., 2013,4, 515-519

There are certain papers in this set that we believe everyone should read but for reasons that are not entirely positive. The follow up paper to “Escape from flatland” is one of those. This paper is “Escape from flatland 2: complexity and promiscuity”.

This is the first of a group of three papers in the bucket list that focus on bad behaviour by those analysing data – one paper illustrating some of the problems and two providing some guidance. It is very easy to fall into lots of traps when analysing large datasets and the reader will doubtless be able to trawl our own papers to find examples of bad practice! The key thing is that amongst the bucket list papers are several that should help you avoid many of the traps.

In flatland, complexity is defined as the fraction of sp3 carbons and the number of chiral centres – a rather limited conception of a concept that I am sure we could argue about for years. As for promiscuity, it turns out in flatland this is the number of assays in which a compound achieves inhibition greater than 50 % at a concentration of 10 µM in a 15 assay panel (a parallel analysis looks at a panel of CYP enzymes). This set of 15 assays is described as being a subset of the Cerep panel selected by a panel of “internal scientists”. Readers should already be concerned that a compound that achieves 49 % inhibition in 15 panels would score a promiscuity of 0 while one with 51 % inhibition in 1 would be infinitely more promiscuous – more on this in future bucket list papers. There is also the worrying problem that there will be plenty of compounds that are in the set which are not soluble at 10 µM.

As shown in the graph, this measure of promiscuity can be plotted against binned values of the fraction of sp3 carbon atoms. Both axes suffer from seen and unseen lines drawn in the sand in order to categorise continuous data. The dataset is not made available for readers to judge how strong the illustrated trends actually are – it is perfectly possible that such trends can be completely changed by moving the dividing lines between categories. In the graph, red objects are for “aminergic” compounds (containing amines) while blue are others. The author is trying to suggest that increasing “complexity” leads in general to a decrease in “promiscuity”. It is hard to know how much emphasis to give to this suggestion or to be able to translate it readily to helping solve problems in drug discovery projects. Even given the trend, the average promiscuity at the top of the peak is just about 0.3 suggesting that about 5 out of the 15 assays are hit – it is hard to know how promiscuous this really is without even knowing what the 15 assays are.

Escape from Flatland. Lovering, Bikker and Humblet.

J. Med. Chem.2009, 52, 6752-6756

The authors surveyed all of the compounds with MW<1000 disclosed since 1980 (in the GVKBIO database – remember that?) and compared and contrasted them with the subset of molecules had made it to each of the clinical phases and to registration as drugs. They were particularly concerned with understanding whether adding unsaturation into molecules improves their medicinal chemistry properties and assessed this by computing the fraction of carbon atoms that are sp3 hybridised and whether the molecule is chiral or not. As you can see, their analysis suggests that at each progressive stage of the drug development process, the average fraction of sp3 carbons increases and (not shown) the proportion that are chiral also increases but not in quite as clear a way. This paper provides a challenge to chemists to find new methods that don’t naturally increase the number of flat atoms in the molecule (as amide couplings and the various palladium catalysed aryl couplings).

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

Molecular Complexity and Its Impact on the Probability of Finding Leads for Drug Discovery. Hann, Leach and Harper

J . Chem. Inf. Comput. Sci. 2001, 41, 856-864

At the heart of this paper is a delightfully simple thought experiment with obvious relevance to medicinal chemistry. There is a tension involved in testing molecules: the more molecular recognition features they contain, the more ways there are for them NOT to make a good set of interactions with a binding site. However, it is only when a molecule contains enough molecular recognition features that it can make an interaction that is strong enough to a) be detected and b) be useful. These two trends are in opposition to one another and lead to a maximum likelihood of observing a “useful” interaction that is summarised by the orange line in the graph. It directly follows from this that there is an advantage to be had by using more sensitive measurement methods that can detect weaker binding (which pushes the red curve to the left). Being able to detect these “lower complexity” molecules should increase the hit rate – the paradigm of fragment screening – now common terminology that is not used in the paper itself.

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

In the same theme of benchmarking compounds, the question what is the most potent you should expect a compound to be was addressed by Kuntz and Kollman. Indirectly this led to the whole ligand efficiency debate and the attempt to put fragments and larger molecules on the same scale. It also provides a useful null model in the QSAR and AI fields – if your model doesn’t do better than correlation with atom count, you’ve not contributed much.

“The maximal affinity of ligands” Kuntz, Kollman et al:
PNAS August 31, 1999 96 (18) 9997-10002

On a much larger scale Reynolds et al looked at the same problem 8 years later:

Reynolds CH, Bembenek SD, Touge BA. The role of molecular size in ligand efficiency. Bioorg Med Chem Lett. 2007 17(15):4258-61.

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

Peter Andrews strength of functional groups analysis

Functional Group Contributions to Drug-Receptor Interactions

J. Med. Chem. (1984), 27, 1648-1657

On the path from traditional QSAR and Topliss decision trees to matched molecular pair analysis, fragment based lead generation and ligand efficiency, Peter Andrews early attempt to calculate what contribution different functional groups make to binding is an important milestone. Collating 200 compounds with their data was a heroic effort in the 1980’s it’s worth considering how far we have come with access to data. Still a pivotal question for medicinal chemists is  “how well is my ligand binding compared to what I should expect” .

Applying a similar approach to a huge data set the results become more nuanced but the learning is significant.

Hajduk and Sauer , “Statistical Analysis of the Effects of Common Chemical Substituents on Ligand Potency”

J. Med. Chem. (2008), 51, 3553-564

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

The advent of high throughput screening brought a new problem for the medicinal chemist: false positives. These are compounds that appear active in an assay but rather than giving a useful mode of inhibition are actually interfering with the assay technology. Such compounds can be a huge resource sink and distraction for chemists and biologists, so across Pharma chemists developed rules for removing them. One of the highest impact sets of rules is Baell and Holloway’s Pan Assay INterference or PAINS set. However as with any set of rules they’re highly controversial as obviously excluding a compound right at the start of a programme can make a significant impact. The debate about their use rages on.

“New Substructure Filters for Removal of Pan Assay Interference Compounds (PAINS) from Screening Libraries and for Their Exclusion in Bioassays”  Baell and Holloway: J. Med. Chem. 2010, 53, 2719–2740

And for the debate about the filters to use and their selection:

Tropsha:. J. Chem. Inf. Model. 2017, 57, 417−427

Kenny: J. Chem. Inf. Model. 2017, 57, 2640−2645

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

Having addressed potency issues in 1972, John Topliss’s contribution to understanding bioavailability is also a classic. Bioavailability is a highly complex property covering absorption, metabolism, protein binding and excretion, nevertheless Topliss showed how a rational approach could be applied to attempt to predict classes of bioavailability. An area of considerable continuing interest.

“QSAR Model for Drug Human Oral Bioavailability” Yoshida & Topliss:  J. Med. Chem. 2000, 43, 2575-2585

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

“Utilization of Operational Schemes for Analog Synthesis in Drug Design“

Although Hansch and Leo had demonstrated the value of analysing compound potency using steric and electronic descriptors, it took John Topliss to reframe the approach as a decision tree that chemists could rapidly apply to explore chemical series. Beyond the well remembered ‘aryl substitution tree’ there is also an alkyl side chain tree and 3 example series.

It is interesting to note Topliss’s reflection on the use of statistical methods by chemists:

“Another problem in the utilization of the standard Hansch method is the reluctance on the part of some medicinal chemists to become involved with mathematics, statistical procedures, and computers. For these individuals a nonmathematical utilization of the Hansch approach might be of considerable interest.“

Topliss: J. Med. Chem. 1972, 15, 1006 – 1011

Almost all chemists use computers now, but still many remain resistant to applying mathematics or statistics to their design process

For a recent review see: http://dx.doi.org/10.1021/acs.jcim.7b00195

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

Any paper that starts with: “It usually comes as a surprise to students to learn that some (perhaps most) published articles belong in the bin, and should certainly not be used to inform practice.” Is clearly grounded in experience.

Medicinal Chemistry touches on a huge number of other disciplines, and with most chemist’s primary training in synthetic chemistry, developing the skills to read other disciplines papers intelligently is essential to rapidly filter the vital from the fatally flawed. This short publication elegantly captures some of these critical skills coming from a clinical perspective.

“How to read a paper : Getting your bearings (deciding what the paper is about)”

Greenhalgh,  BMJ 1997;315:243

The book of the same name also has excellent sections on statistics for the non-statistician, assessing methodology and assessing review papers.

Greenhalgh, “How to read a paper”

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

The combinatorial chemistry and HTS boom of the 1990’s led to some shocking compounds being put into development.  Analysis of their failure identified pharmacokinetics as a key issue.  After the publication of Lipinski’s work, a number of other group turned to analysing their own ADME datasets asking the question :“what are the molecular properties of compounds with good and bad bioavailability” with the hope of designing in better pharmacokinetics.

“Molecular Properties That Influence the Oral Bioavailability of Drug Candidates”

Veber et al, J. Med. Chem. 2002, 45, 2615-2623

For a retrospective on the properties of oral drugs and their analysis 20 years on, Shultz’s recent mini review gives another perspective:

Shultz , J. Med. Chem. 2019, 62 , 1701-1714.

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

It can be argued that BucketListPapers should be pointing out the “must read papers” from the literature. Kuhn, Mohr and Stahl do it again with another comprehensive review tailored for the drug and agrochemical hunter. Given the interest in “beyond rule of 5” compounds, this paper is even more relevant.

“Intramolecular Hydrogen Bonding in Medicinal Chemistry” Kuhn, Mohr and Stahl

J. Med. Chem. 2010, 53, 2601–2611

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

Bissantz, Kuhn and Stahl produced a comprehensive guide to molecular interaction of molecules in biological systems. A must read for all medicinal and agrochemical compound designers.

“A Medicinal Chemist’s Guide to Molecular Interactions” Bissantz, Kuhn and Stahl

J.Med.Chem. 2010, 53, 5061-5084

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

In molecular design we often refer to key ‘groups’ in our current best molecules and their effects. The concept of an isostere is a group that can replace another in molecular and retain most if not all of the properties. Better still is to improve one or two properties and keeping everything else the same.

“Isosterism and molecular modification in drug design”, Thornbar

Chem. Soc. Rev., 1979,8, 563-580

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

Optimising compound absorption through cell membranes is often an issue in drug discovery. A major improvement can be breakthrough moment in drug discovery. Absorption was known to correlate with Polar Surface Area (PSA), a calculated value estimating the amount of the surface of the molecule that is not hydrophobic. Too high a PSA and absorption through a non-polar membrane is harder as a design principle. The challenge was calculating these quickly to meet the demands of modern drug discovery. The method described by Ertl, Rohde and Selzer yields a Topological Polar Surface Area calculation, which today we take for granted as a key descriptor of molecules.

“Fast Calculation of Molecular Polar Surface Area as a Sum of Fragment-Based Contributions and Its Application to the Prediction of Drug Transport Properties”: Ertl, Rohde, Selzer

J.Med.Chem. 2000, 43,  3714-3717

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

In compound design, and particularly drug design, the concept of lipophilicity is key. A measured partition co-efficient between water and octanol serves as a predictor of further properties and drug developability. Leo, Hansch and Elkins set out to compile multiple measurements from the literature to form the basis of further understanding of molecules interacting with biological systems. Without this “ClogP”, a computer calculated value we take for granted so much, would not exist. A job very well done.

“Partition Coefficients and their Uses”: Leo, Hansch and Elkins:

Chem Rev. 1971;71(6):525–616

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

At the heart of handling chemical information within computers are methods to store complex structures accurately, uniquely and in a searchable manner. This early paper from Morgan describes one of the techniques that allowed the registration of compounds to CAS. Computers clearly deal with numbers quickly, so the further encoding of a structure into a bit number allows very fast comparison. The algorithm described within became the basis of the very widely used Morgan fingerprints and now is at the heart developments in convolutional neural networks.

The Generation of a Unique Machine Description for Chemical Structures-A Technique Developed at Chemical Abstracts Service. H. L. Morgan

J. Chem. Doc.1965,5,2,107-113

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

Five years after Roughley and Jordan’s seminal approach looking at medicinal chemists preferred reactions by hand, the folks at NextMove and Novartis used automated natural language processing to analyse >200,000 patents and extracted over 1.1 million unique reactions.  Using the Roughley and Jordan reaction typing they then classified the reactions.  With this much larger data set they could analyse the evolution of reaction types, for example with carbon-carbon bond formations they see the switching from phosphorus ylid to palladium catalysed cross couplings as the Suzuki and Negishi reactions have been applied in drug hunting research. Still however alkylation and acylation of heteroatoms remains a key process. They also analysed the properties of the products of reactions where unsurprisingly compounds grow in size and rigidity over the 40 year period reviewed.

This scale of work would never have been possible without automation and now it’s hard to see why anyone would ever go back.

“Big Data from Pharmaceutical Patents: A Computational Analysis of Medicinal Chemists’ Bread and Butter” by Schneider, Lowe, Sayle, Tarselli & Landrum

J. Med. Chem.(2016), 59, 9 ,4385-4402

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

The authors surveyed the publications in three medicinal chemistry journals in 2008 covering 139 papers describing the synthesis of 3566 compounds and employing 7315 different reactions. They categorised the reactions that had been used and identified surprising trends such as the frequency of C-C bond forming reactions being about 10 %.  They highlight the 10 most frequently employed reaction types (Table below) and that an average medicinal chemistry synthesis used 4.8 steps per compound (Graph).  They finish with some challenges to chemists working in industry and academia.

For what happens when AI comes to reaction analysis see our next post…..

“The medicinal chemist’s toolbox: an analysis of reactions used in the pursuit of drug candidates” by Roughley and Jordan.

J. Med. Chem.(2011),54,10, 3451-3479

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

This survey of the impact of fluorine on medicinal chemistry highlights that fluorinated molecules have accounted for about 5-15 % of approved drugs over the course of decades.  Fluorine has often replaced hydrogen or oxygen in earlier lead compounds and retained effectiveness.  The ability of fluorine to make interactions with proteins and to affect pKas is discussed and some of the methods for introducing fluorine into a lead molecule are highlighted.  Some case studies of drugs that benefit from a fluorine (either by improved pharmacokinetics or potency) are showcased and a final section suggests that introducing fluorine could reduce metabolism sufficiently to make drugs that are excreted intact into the environment; our own findings suggest that this latter effect is not really to be expected – fluorination increases metabolism as often as it decreases it when a comprehensive survey is made using matched molecular pairs: https://pubs.acs.org/doi/10.1021/jm0605233

“The many roles for fluorine in medicinal chemistry” by Hagmann.

J. Med. Chem.(2008)51, 15, 4359-4369

DOI: 10.1021/jm800219f

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

The paper presented an advance in docking technology that has been at the heart of Schrödinger’s software suite since. Standard Precision Glide (SP) is introduced in detail and includes terms for hydrophobic interactions, hydrogen bonds of charged and neutral types, metal interactions, a rotational bond penalty (for conformational entropy), coulombic and van der Waals terms.  In addition a solvation term considers water molecule presence. This glidescore is used to predict binding free energy and a combination of glidescore with molecular mechanics interaction energy and ligand strain energy to select docked poses. It is curious that it was necessary to scale the van der Waals’ radii (effectively shrink the atoms) to fit some known actives into their cognate binding sites. The effectiveness of the approach for correctly positioning ligands in the crystal structure they came from was compared to that with GOLD and FlexX. The test set used included those used for these other approaches and involved 282 complexes taken from crystal structures. To avoid bias from the ligand geometry, these were all created from scratch from 2D structures.

“Glide: A New Approach for Rapid, Accurate Docking and Scoring. 1. Method and Assessment of Docking Accuracy”  by Friesner, Banks, Murphy, Halgren, Klicic, Mainz, Repasky, Knoll, Shelley, Perry, Shaw, Francis and Shenken.

DOI: 10.1021/jm0306430

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

What else could be at number 5? In this article, the authors consider the properties that influence the solubility and permeability of drugs. This was partly a response to the introduction of high-throughput screening to drug discovery which yielded compounds that had never been tested in cells or in vivo and consequently had properties that need not be consistent with good solubility and permeability.  They showed that there had been an increase in higher molecular weight compounds and an increase in logP of compounds in the previous decade. The properties of a set of over 2000 drugs were considered (with a selection criteria for clinical exposure likely to favour orally bioavailable drugs).  Of these, about 10% was found to have a clogP above 5, about 11% a molecular weight above 500; such compounds are likely to have poor solubility.  A crude estimate of the number of hydrogen bond donors (number of OH + number of NH) and even cruder estimate of the number of acceptors (number of O + number of N) were also examined; only 8% of the drugs have donor count above 5 and only 12 % an acceptor count above 10.  These latter groups likely suffer from poor permeability. The period since this paper has seen a host of attempts to derive “rules” for other types of compounds (fragments, leads etc) although more nuanced views have also emerged.  Personally, we think that the prompt to consider the importance of solubility and permeability was a good one but the introduction of the idea of “rules” about what a drug looks like was a misstep.

“Experimental and computational approaches to estimate solubility and permeability in drug discovery and development settings” by Lipinski, Lombardo, Dominy and Feeney.

https://www.sciencedirect.com/science/article/pii/S0169409X96004231

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

The sulphonamides exert their bacteriostatic effect by competing with p-aminobenzoic acid.  To gain a deeper understanding of the structure-activity relationship, the activity and pKa of 50 sulfonamides were measured.  This required measurements that were as free from confounding influences as possible – an approach that prefigured much of the work undertaken in target-based drug discovery using all the latest tools of molecular biology since the 1990s. It was found (see graph) that the sulphonamides show a maximum in activity at pKas around 6. This is explained by a requirement for the sulphonamide to be in its ionised form (as might be expected for competition with p-aminobenzoic acid) but that the anionic form must not see the charge so stabilised by delocalisation that the SO₂ does not carry a maximal negative charge. These are the kernel of key ideas about understanding and optimising molecular interactions with a binding site that underpin much modern drug design.

“Studies in Chemotherapy. VII. A Theory of the Relation of Structure to Activity of Sulfanilamide Type Compounds” by Bell and Roblin  J. Am. Chem. Soc.(1942), 64, , 2905-2917

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

Various scientists had been gathering experimental data from which to extract descriptors that could be used to rationalise electronic (sigma values), size (molar refraction, MR and molecular weight, MW) and hydrophobic (pi values) effects caused by substituents, particularly on aromatic rings.  Hansch and co-workers aimed to bring all of these descriptors together in one place as a resource for those aiming to explore quantitative structure-activity relationships; an example of 4 groups that span the different types (polar vs hydrophobic, electron-donating vs withdrawing) is shown in the table.  Hansch et al.’s compilation of values remains a good starting point to understand the effects that various substituent types can exert, although Hansch and Leo went on to compile an even more comprehensive set: https://pubs.acs.org/doi/abs/10.1021/jm950902o. Some of our own work in matched molecular pairs has seen us effectively computing analogous values for the effect of aromatic substituents on solubility, plasma protein binding and other properties:  https://pubs.acs.org/doi/10.1021/jm0605233

“Aromatic Substituent Constants for Structure-Activity Correlations” by Hansch, Leo, Unger, Kim, Nikaitani and Lien.

DOI: 10.1021/jm00269a003

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

In this paper Free and Wilson analyse the variation in biological activity caused by substituents at various positions in a series of molecules.  They apply a simple least squares fitting to attribute a contribution that each substituent is making to the observed activity. This analysis of variance has become a fundamental tool for medicinal chemistry.  It is assumed that the SAR is additive although the possibility of additive and non-additive groups is discussed. The method should always be considered whenever the question is asked “have we made the best combination of substituents yet?” and as such should be used in almost every drug discovery project.  Recent developments building on the Free and Wilson approach have explored how many compounds must be included in the data set to ensure satisfactory values for the contributions of each group, for example:

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

https://pubs.acs.org/doi/abs/10.1021/jm801070q

“A Mathematical Contribution to Structure-Activity Studies” by Free and Wilson. DOI:10.1021/jm00334a001

Our first BucketListPaper is from 1899 and introduced the idea that the activity of anesthetics might correlate with  lipophilicity (as measured by partition into oil from water).  This first hint that lipophilicity might be a key property, as found by Overton and Meyer, kick-started the field of QSAR. The papers formed the basis of a theory of anesthesia, which still intrigues and stimulates new scientific studies.

#BucketListPapers #DrugDiscovery #MedicinalChemistry #BeTheBestChemistYouCanBe

1. Overton E. Studien über die Narkose Zugleich ein Beitrag zur Allgemeinen Pharmakologie. Jena, Germany: Verlag von Gustav Fisher, 1901. Meyer H. Zur Theorie der Alkoholnarkose. Arch Exp Pathol Pharmakol 1899;42: 109-18
2. Graph abstracted by Campagna et al.: https://www.nejm.org/doi/full/10.1056/NEJMra021261)

https://pubs.acs.org/doi/10.1021/acs.chemrev.8b00366.