By expecting less, pharma can get more out of modeling.
By Vicki Glaser
Oct. 8, 2008 | Software packages described as decision-enabling or decision-support tools are gaining a higher profile in the drug discovery arena, yet how enabling are these predictive in silico approaches? Are medicinal chemists confident enough in their predictive capabilities to rely on them for compound design and lead selection?
The sorts of software packages being marketed as decision support tools typically offer either data integration packages, LIMS systems, or graphic visualization tools, but “do not make a real difference in the way decisions are made in drug discovery,” contends Matt Segall, senior director of ADMET at BioFocus DPI. In his view, in silico modeling software should enable “guided decision making—it’s about analyzing the data within the context of your objectives to identify the best route to take forward. Because of the complexity of the data now generated in drug discovery, that is a multi-component, multi-objective optimization process.”
Too often overlooked is the fact that “every individual parameter you are measuring has a significant degree of uncertainty or variability, and that needs to be taken into account when making decisions,” says Segall. He describes BioFocus’ technique of “probabilistic scoring” as a method that applies probability-based analysis to complex, uncertain data to identify areas of chemistry that have the highest likelihood of success.
The company designed its StarDrop platform for use by drug discovery scientists rather than computational experts, to help them define the criteria for success of a project. As key criteria may be in conflict—for example, driving potency up may sacrifice metabolic stability or other pharmacokinetic properties—part of this process involves defining the relative importance of individual properties and generating a target product profile.
The software helps identify and balance these inevitable trade-offs in multiples properties according to individual scientists’ priorities in a common and objective language. If those priorities change with future analyses and an expanding knowledge base, then the profile can be modified and the test compounds rescored to determine what effect the change has had on the decision-making process.
Although a target profile is itself subjective, Segall would argue that “it’s better to decide up-front what you are looking for—but be flexible—than to focus too much on one objective and hope you end up in an area of chemistry that will give you the other properties you need.”
Why the Bad Rap?
Mike Moyer, director of medicinal chemistry at the Stanley Center for Psychiatric Research, The Broad Institute of MIT and Harvard, spent 18 years at Pfizer before moving to academia. In his current environment the more limited resources available to synthesize and test compounds has put a premium on optimizing compound design early on. To achieve that, Moyer’s group has turned to in silico modeling techniques and has, in a sense, turned the model generation process on its head.
Moyer contends that building the most predictive in silico models while synthesizing the fewest number of actual compounds requires an up-front investment in model generation. This necessitates designing and synthesizing compounds specifically for the purpose of constructing and optimizing models.
“In silico models will only be as good as the datasets on which they are built,” says Moyer. Although he acknowledges that “no medicinal chemist wants to make a compound simply for model making,” he believes this to be the most efficient path to compound optimization.
The initial compounds made in many medicinal chemistry labs do not necessarily yield optimal models. Thus, when these models are then used to drive compound design and optimization, the results may be disappointing, which, Moyer states, is at least part of the reason in silico modeling may not be as predictive as desired and has garnered a bad reputation.
Unrealistic expectations have also plagued the field. In silico modeling programs “are not perfect and they will never be perfect,” says Paul Wyatt, director of drug discovery in the College of Life Sciences at the University of Dundee. “In the world of [drug metabolism and pharmacokinetics] (DMPK), there is never an exact figure”—a drug does not behave exactly the same way in two different people.
After a 25-year career in pharma and biotech, including stints at SmithKline Beecham and Glaxo, Wyatt moved to Dundee in 2006 to set up a drug discovery group focused on neglected diseases, mainly tropical parasitic diseases. Based on his experience he is convinced that “in silico must be the way to go for predicting the properties and activities of molecules.”
The value of in silico modeling, asserts Moyer, lies in the ability to perform multidimensional computational analysis and to strike a balance between competing properties and identify the best possible combination of features of an individual compound designed for a specific target.
Moyer values the StarDrop program for its ability to look at the trade-offs between potency and ADME properties and “find ways to get to the sweet spot” as quickly as possible. “It has a way of explicitly calculating the uncertainty in the model and making that visible to the user. It gives you a value along with an associated error and makes clear the limitations in the model.”
Defining the Anti-Drug
Perhaps the greater benefit of in silico modeling lies not in identifying promising compounds, but rather in ruling out compound sets or chemistries that would likely lead down a development path doomed to failure.
“These programs can give you a pretty good idea of which compounds not to make,” says Wyatt. Pharma has been operating under the faulty assumption that if it takes 10,000 compounds to find a drug, then there must be a suitable drug candidate in a library of 10,000 or more compounds, and it is just a matter of making the compounds and screening them against the target to find the drug. But actually synthesizing large compound libraries eats up valuable time and resources and is no guarantee of success.
One of the limitations Wyatt has found with many of the modeling and decision support tools available on the market is that they are standalone programs that do not allow for communication and data sharing between applications and users. With StarDrop, Wyatt’s group is able to incorporate its own data and scoring functions, to compare sets of compounds, and to generate compound profiles across a broad range of parameters.
StarDrop incorporates in silico ADME and QSAR models, quantum mechanical P450 models, probabilistic scoring techniques, data analysis tools, the Glowing Molecule visualization and interactive compound design tools, and the Auto-Modeler for developing models based on users’ own data. Earlier this year, BioFocus DPI made several of its predictive drug discovery databases publicly available online by transferring them to EMBL’s European Bioinformatics Institute. The StarDrop software platform was not part of this transaction, which did include DrugStore, StARLITe, Strudle, and Kinase and GPCR SARfari. In August, Galapagos acquired Sareum Holdings’ drug discovery services business, adding structural biology capabilities to BioFocus’ pre-clinical drug discovery services.
This article appeared in Bio-IT World Magazine.
Subscriptions are free for qualifying individuals. Apply Today.