Pharma researchers mistrust in silico models of ADMET, Han van de Waterbeemd, of Pfizer Global research and development, told an audience at Drug Discovery Technology Europe earlier this month.
Experimental ADMET “feels better”, and computational prediction “feels scary,” according to van de Waterbeemd. But combining the two is likely to lead to more efficient and effective toxicity prediction for drugs in humans, and computational ADMET approaches, he went on to say, are much needed.
Then, to emphasize his point, he flashed a PowerPoint that read: “Do we have all the ADMET predictive tools we need? NO!”
Recently, companies like UK-based Amedis have heeded this call for new ADMET programs. Amedis uses several algorithms based on genetic programming to break down compounds into their atomic components and then compare these components to the original data.
The payoff can be huge for big pharma and developers of the software. At a recent presentation, Boston Consulting Group’s Detleve Biniszkiewicz said such software could add $35 million in net value to a $600 million drug project.
But what exactly does pharma need in this area?
According to van de Waterbeemd, key questions that ADMET software can address involve not only the QSAR (quantitative structure activity relationships), but the QSPR, or quantitative structure-property relationships, with the properties including absorption, distribution, physiochemistry, and toxicity.
Studying the factors that affect absorption and bioavailability is a data-intensive process, which involves calculating physiochemical properties of the compound, such as Log P, Log D, water solubility, pKa, and permeability. Tools that model the structure of the compound and how it is affected by these properties, he said, would be helpful, especially when married to models of efflux and passive diffusion.
Another issue, blood-brain-barrier and skin permeation, is ripe for in silico modeling, because “the data is the issue,” said van de Waterbeemd. In particular, data on cerebrospinal fluid and how it interacts with the compound under study may be more important than brain data, he noted.
Of course, metabolism is worth predicting in silico too, he said. Although there are numerous options in this area, such as MetabolExpert, M-CASE, Meteor, and Metabolite, the quality of these programs is “far from perfect,” he said. Any program should look at the rate of metabolism, stability, turnover, inhibition of p450s and other enzymes, site of metabolism, enzyme specificity, and metabolic pathways.
The good news is, programs for all stages of the ADMET process — from metabolic stability to intestinal absorption, blood-brain-barriers, and microsomal binding, are multiplying. Van de Waterbeemd noted that there were 50 ADMET posters presented at Euro QSAR 2002. One poster featured PreAdme, a web-based solution being developed by Kyoung Tai No and his team at the Computer Aided Molecular Design Research Center in Seoul, South Korea. (A beta version is available for download at http://camd.ssu.ac.kr/adme/).
In general, van de Waterbeemd said that more global probabilistic approaches are helpful at the earlier hit triage stage in drug discovery, while local mechanistic ones are more useful at the candidate selection phase. Since these solutions will do more than signal that “something is wrong,” and actually answer the question of why something is wrong, they may enable wet-lab researchers to then modify a lead to specifically address the ADMET issue.