Multi-parameter optimization: The delicate balancing act of drug discovery

Dealing with complex, multi-parameter data with high uncertainty is an enormous challenge. One must consider the overall balance of properties as early as possible in the process to focus on chemistries with the best chance of quick progress and downstream success.

Register for free to listen to this article
Listen with Speechify
A safe and efficacious drug has a balance of manyproperties. Potency against the therapeutic target is essential and appropriatephysicochemical and absorption, distribution, metabolism and elimination (ADME)properties are also required in order to achieve suitable in-vivo disposition. Furthermore, selectivity against off targetsand an absence of non-specific and idiosyncratic toxicities are necessary toachieve acceptable side effect and safety profiles. Unfortunately, theserequirements are often conflicting, with improvements in one property leadingto detrimental changes in another.
The challenge of successfully achieving this delicatebalancing act is illustrated by a historical view of pharmaceutical R&D, apicture dominated by increasing costs and low success rates. The causes of thehigh attrition rates for clinical candidates have changed over the years. In particular,the previously high failure rate due to poor pharmacokinetics (PK) has beenreduced, while the failures due to toxicity have increased. The reduction in PKfailures has been achieved through the introduction of early in-vitro screens to filter out compoundswith potential ADME issues, and similar efforts are underway to develop earlytests for toxicities. However, the overall success rate has not improved, andthe hidden cost of missed opportunities due to good compounds incorrectlyeliminated is also likely to be high. This suggests that an alternativeapproach is required, taking a holistic approach to designing compounds with agood balance of properties as early as possible in the process.
This realization has led to a recent surge in interest inmethods for simultaneously optimizing multiple factors, described asmulti-parameter optimization (MPO). Many MPO methods have been developed infields such as engineering, economics and quality control that may be readilyadapted to the drug discovery environment and in the context of drug discoverya good MPO method needs to satisfy the following requirements:
  • Interpretability: The output must provide intuitive guidanceon the impact of individual properties on the quality of a compound and how itcould be improved. 
  • Flexibility: There is no "one-size-fits-all" definition of a perfect drug. A project team should be able to define the profile of properties they require, depending on the therapeutic and commercial objectives of the project, historical data and experience.
  • Weighting: Not all properties are equally important. Compromise is usually necessary, and it should be possible for a project team to define the acceptable trade-offs by weighting individual property criteria.
  • Uncertainty: Unlike engineering disciplines, where it is often possible to simulate the properties of a design within a fraction of a percent, it is common for computational models in drug discovery to have uncertainties of an order of magnitude, and even experimental measurements may have uncertainties of a factor of two or more. Therefore, when selecting and designing compounds, it is important to consider the confidence with which we can choose between compounds in order to avoid missed opportunities.
A number of MPO methods are being applied in drug discovery.Here, we will briefly discuss some key, illustrative examples.
The most famous example of a rule-of-thumb is Lipinski'sRule of Five (RoF), which relates the molecular weight, lipophilicity andhydrogen bonding characteristics of a compound to its likelihood of achievinggood oral absorption. The RoF has since been joined in the medicinal chemists'armory by many other rules-of-thumb that relate the biological properties, developmentpotential or safety of compounds to simple characteristics, including polarsurface area, flexibility, number of sp3 carbons and number of aromatic rings.
The enormous popularity of rules-of-thumb derives from theease with which they may be applied and interpreted. They provide clearguidelines that help chemists to focus on factors that will increase thepotential to identify high quality compounds while optimizing potency andeliminate chemistries with a low chance of success early in the process.
Rules-of-thumb have been derived from extensive statisticalanalysis of historical data and most relate to a specific objective. Therefore,it is important to apply them in an appropriate context; for example, the RoFwas developed as a guide to improving the chance of achieving oral absorption,but it is frequently applied as a definition of "drug-likeness," despite thefact that the requirements for a compound intended for other routes ofadministration such as intravenous or inhalation are quite different,potentially leading to inappropriate decisions.
It is also important not to be too rigid when applyingrules-of-thumb, particularly when options are limited. The correlations betweenthese simple characteristics and the biological properties of a compound arenot strong. Is there a significant difference in the chance of oral absorptionfor a compound with a molecular weight of 501 Da versus one with 499 Da?
Probably the most common approach used in an attempt toidentify compounds with an appropriate property profile is to filter out thosecompounds that fail to meet each criterion in turn, with the hope that one ormore "ideal" compounds will emerge at the end, having satisfied all of thecriteria. Early in a project, the criteria corresponding to one of therules-of-thumb may be applied as filters, but predicted and experimentallymeasured properties are often compared against a target product profile in thisway.
The apparent simplicity of a filtering approach hides anumber of dangers. Achieving a balance of properties often requires compromise,as the requirements often conflict and it is very common for no compounds toemerge from the sequence of filters. Hard cut-offs introduce artificially harshdistinctions between options, and this is exacerbated by the effect of theuncertainty in the underlying data; combining multiple uncertain filtersaccumulates error and dramatically increases the chance of incorrectlydiscarding a good compound. As a simple illustration, if we apply 10 filtersthat are each 90 percent accurate in passing/failing a compound, theprobability of an ideal compound emerging, even if it was present in the setbeing filtered, is only 35 percent; we are more likely to throw away an idealcompound than accept it.
An alternative approach that avoids hard cut-offs and allowsacceptable trade-offs to be defined is provided by "desirability functions." Adesirability function maps a property value onto a scale between zero and onethat represents the desirability of a compound with that property value; anideal value will achieve a desirability score of one, while a completelyunacceptable value will receive a desirability score of zero. Desirabilityfunctions give excellent flexibility, as a desirability function can take anyshape. For example, a slope indicating increasing desirability as the propertyvalue approaches the ideal range or even non-linear relationships, such as asigmoid or bell curve. The desirability scores of individual properties canthen be easily combined into a "desirability index" to reflect the overallquality of a compound by adding them together or taking the average orgeometric mean.
Furthermore, the results may also be easily interpreted, as theimpact of each individual property to the overall desirability index can becalculated to guide strategies to improve the overall quality.
A group at Pfizer described an application of desirabilityfunctions to prioritize compounds with a greater chance of success against acentral nervous system (CNS) target. Its "CNS MPO" approach employs sixcalculated physicochemical parameters to calculate a desirability index in therange 0 to 6. They found that 74 percent of a set of marketed drugs for CNStargets achieved a CNS MPO index of ≥ 4, compared with only 60 percent of thePfizer candidates, a statistically significant difference. They also found thathigh CNS MPO index correlated with positive outcomes for several key in-vitro ADME and toxicity endpointsincluding permeability, metabolic stability, active transport byP-glycoprotein, cytotoxicity and hERG inhibition.
The probabilistic scoring method builds on the flexibilityand interpretability of desirability functions by explicitly taking intoaccount the uncertainty in the underlying data. Thus, not only is a scorecalculated for each compound, representing the likelihood of achieving therequired profile of properties, but an uncertainty in each overall score isalso estimated. This provides much more information, as it becomes possible tosee when the data allows compounds to be confidently distinguished or,alternatively, when higher resolution data or another criterion is necessary tomake a clear choice. The objectivity this introduces allows projects to focustheir efforts on chemistries with the highest chance of success while notmissing potential opportunities. 
Dealing with complex, multi-parameter data with highuncertainty is an enormous challenge. The temptation is to focus on a singleproperty, often targeting potency, in the hope that any issues that arise canbe dealt later in the process. This is a risky strategy, as once locked into atight chemical series, it becomes difficult to break out if other necessaryproperties are not achievable. This leads to long optimization cycles andlate-stage failures, increasing the time and cost of drug discovery andreducing productivity. It is much better to consider the overall balance ofproperties as early as possible in the process to focus on chemistries with thebest chance of quick progress and downstream success.
Dr. Matthew Segall is director and CEO of Optibrium Ltd.Segall has a M.Sc. degree in computation from the Universityof Oxford and a Ph.D. in theoreticalphysics from the University of Cambridge. Asassociate director at Camitro, ArQule Inc. and then Inpharmatica, he led a teamdeveloping predictive ADME models and state-of-the-art intuitivedecision-support and visualization tools for drug discovery. In January 2006,he became responsible for management of Inpharmatica's ADME business, includingexperimental ADME services and the StarDrop software platform. Following acquisitionof Inpharmatica, Segall became senior director, where he was responsible forBioFocus DPI's ADMET division. In 2009, he led a management buyout of theStarDrop business to found Optibrium.

Subscribe to Newsletter
Subscribe to our eNewsletters

Stay connected with all of the latest from Drug Discovery News.

DDN Magazine May 2024

Latest Issue  

• Volume 20 • Issue 3 • May 2024

May 2024

May 2024 Issue