At CHI’s “World Pharmaceutical Congress” held recently in Philadelphia, there was no shortage of good ideas for ways to optimize the drug discovery pipeline. From retrospective studies to ground-breaking methodologies, this topic elicited some lively discussion.
“Historically, drug discovery has been much like the stock market in that, over the long term, steady positive progress has been made,” said Josef Scheiber, Ph.D., postdoctoral fellow, lead discovery informatics/safety profiling, Novartis Institutes for BioMedical Research. As evidenced by the presentations, however, there is plenty of ground left to be broken.
Dr. Scheiber spoke at the meeting about a novel method to predict adverse side effects from the chemical structure only. He believes that, by employing clinical databases along with chemical information about the drugs used for treatment, the chemical features that are most likely the reason for a certain side effect can be identified. This information can then be employed in early drug discovery. “It is possible to avoid chemical moieties regarded as unsafe in the lead-optimization process,” Dr. Scheiber reported.
“We use information about common chemical space between adverse drug reactions and target interaction of a compound to identify overlaps. Ultimately that leads to an understanding of the big picture of adverse reactions. Where possible, we can link adverse drug reactions to targets and compute a kind of map.”
Another aspect of his talk focused on finding additional uses for compounds. “Look at Viagra and how that originated,” Dr. Scheiber continued. “There are many more successful examples, like the recent repurposing of Ropinirole for restless leg syndrome. It’s becoming clear that you can use data from clinical studies and early discovery to design better compounds. You can use information from late-stage drug development so we can understand and develop better compounds.”
The Challenge of Finding Good Space
There are a few key challenges to library design. In addition to the vast number of monomers in random order, there is the question as to whether they are in fact available in addition to making an objective assessment of the design space.
“The question you need to ask is: what is the drug-like channel space for my series?” said Robert Maguire, senior principal scientist, cardiovascular, metabolic and endocrine diseases at Pfizer, “and can it actually work?”
Maguire described how an analysis of the physical properties of Pfizer lead series and candidates can be used to influence the design, selection criteria, and optimization for library sets in early programs. “Our approach uses readily available physical-property calculations to define design zones that are subsequently populated with target molecules to shape the library and increase the probability of identifying lead compounds with drug-like physical properties and room for more facile optimization.”
Good design requires a number of considerations, according to Maguire. Starting small and building out is more efficient. “And keep in mind that the farther away you are from a good space, the harder it is to get into a good space,” he said. “Beware of twilight zone compounds—these are the ones that look good but require a lot of optimism and a lot of time.”
Maguire’s talk included advice on seeking a reality check if you are spending too much time on a compound or making too many changes to make a compound work. “You always need to evaluate your project as a whole. Big molecules involve bigger changes to influence their physicochemical profile. The greater the changes needed to influence compound profile, the more time and effort is required to maintain the properties that attracted you to that compound in the first place.”
“Often, you will have access to a training data set and would like to find additional active molecules,” said S. Stanley Young, assistant director of bioinformatics, National Institute of Statistical Sciences (www.niss.org). His talk focused on ChemModLab, a free web service that provides sophisticated QSAR analysis with 80 different models.
“The models from ChemModLab can be used as an ensemble and will virtually screen over 10 million commercially available compounds,” he said. “The resulting list can be diversity selected and used for a lead-hopping campaign.”
Young explained that the training data used was from PubChem. “There is a large number of HTS data sets, which are reasonable training sets that are available publicly so you can test your methods. Rounding out the technical components in his study were powerMV (www.niss.com/powermv) and ChemSpider (www.chemspider.com).
In the ChemSpider demo, they put two million compounds into 42 batches of about 50,000 compounds each. Each batch was processed in less than three minutes—42 batches on 12 processors takes about 12 minutes, Young reported.
The predictions were consolidated, and the top 1,000 compounds were selected, explained Young. “The results were inspected using PowerMV, and it’s as good as anything out there. It’s pretty tough to beat good 2-D descriptors, especially early in the process and especially with regard to the accuracy of the data sets, and there is plenty of data to support that.”
A recurring question throughout all the talks was that of intellectual property rights and how that affects the targeting and developing of compounds. “Whether two molecules are (dis)similar is in the eye of the beholder,” Young responded to one query. “Scientists look to fool the receptor—but you really want to fool the patent office.”
The goal, explained Mee Shelley, senior applications scientist at Schrödinger, is to discover bioactive compounds with novel chemotypes. Her group is doing this is by modifying the core of a lead compound to create structurally distinct chemotypes that have similar or improved binding properties using the receptor information as an important guide.
“The majority of scaffold-hopping methods in the literature use ligand-based descriptors in a search for core replacements,” she said. “However, the search can be focused on a more relevant chemical space by utilizing structural information from the receptor.”
Shelley described a structure-based core-hopping method in CombiGlide, which can systematically evaluate and rank a large collection of candidate core structures in the presence of the receptor.
“We have created a method for evaluating a collection of core candidates starting from a lead compound (template) that is well positioned in a receptor. Candidate structures are evaluated based on geometric alignments with the template as well as the interactions with the receptor.”
Each scaffold aligned with the template must fit also into the receptor, Shelley explained. After side chains are added from the template and the position and conformation are refined, researchers can then rank core-hopped structures based on docking scores. The result is a flexible platform on which multiple applications using core hopping are possible.
“You can find replacement rings or linkers. You can also expand the core region and search with larger candidate structures, e.g., fragments generated using RECAP rules. Another area we have been exploring recently is to use our core-hopping method in conjunction with fragment-based drug discovery—all these are possible on this platform,” said Shelley.
Scaffold hopping, or lead hopping, can be defined as the identification of isofunctional molecular structures with significantly different molecular backbones. Virtual screening and de novo design are two computational techniques that have been applied successfully in the past to generate new hits and lead candidates in early drug discovery.
Ingo Mügge, senior associate director, CADD, Boehringer Ingelheim Pharmaceuticals, discussed how these well-known computational approaches can guide researchers toward new scaffolds with more favorable properties beginning from known ligands.
“A central premise of medicinal chemistry is that structurally similar molecules have similar biologic chemistry,” said Mügge. “A counterpremise is that structurally different molecules can produce similar biologic activities, thus producing the premise for scaffold hopping.”
Looking at the statistics, Mügge noted, the more pharmacophore restraints that are set, the greater the scaffolds reduction will be. “Exhaustive ring replacement works well if many pharmacophoric constraints are known. If no constraints are set for even small ring systems, the number of solutions may be overwhelming.”
Mügge described de novo design with BIBuilder. “The goals are to design high-quality molecular structures using tractable synthesis routes. Results should be drug-like and devoid of unwanted structures and/or circumvent intellectual property constraints. In addition, results will represent prospective new lead series that will require further optimization.”
Two case studies in which BIBuilder was utilized were reviewed by Mügge. His group prepared a set of ligands representing unique chemotypes for seven diverse drug targets, prepared negative controls from MDDR, calculated various descriptors (2-D topological and 3-D), performed a structure-based virtual screen, and then ranked the compounds along several parameters and compared the results.
In the first case, there was no topological bias, as actives were as (dis)similar to each other as they were to negative controls. However, in the second case, actives were more similar to each other than they were to negative controls, creating a topological bias.
“Based on whether you want to conservatively replace the core or whether you want to create a completely new molecule, you can adjust this tool accordingly,” said Mügge. “The more you know about the nature of the core, the more specific the parameters you can place.
“Once you produce a number of compound ideas, that is the time to go to the chemist and pick their brains and make final synthesis decisions” said Mügge. “This is a conservative—but effective—approach.”