Since the first confirmed appearance of the human immunodeficiency virus more than 20 years ago, researchers developing drugs against the virus have been confounded by its ability to mutate and develop resistance to potential treatments.
But researchers in Sweden have developed a protein-based statistical model they believe may enable drug makers to predict how effectively new drug candidates will retard HIV strains, which in turn can lead to new drugs against which mutant HIV strains are less prone to develop resistance.
Their method, called proteochemometric modeling, is designed to analyze interactions of a series of proteins with a series ofligands. It consists of bioinformatics methods that are applied over groups of proteins, and can be used to create models that offer researchers a better understanding of molecular recognition processes, said Jarl Wikberg, a professor of pharmaceutical pharmacology at Uppsala University in Sweden, and the developer of the technology.
Wikberg developed proteochemometric modeling around six years ago. A study describing the application of the technology to HIV appears in the March issue of PLoS Computation Biology. Beyond HIV, Wikberg said that the model can be applied to many diseases.
“Internally, we have used it on targets like protein kinases and ion-channel receptors and antibodies. It’s a completely general technology,” he told ProteoMonitor last week. “The advantage of this technology is that we can optimize [it] for several targets at the same time.”
If the approach is used to develop new HIV treatments, “we predict it will be substantially more difficult for the virus to escape a retardant …compared to presently used retardants,” he said.
Mapping Mutant Strains
For their HIV protease work, Wikberg and his colleagues used a multivariate analysis of the sequence-based physiochemical descriptions of 61 retroviral proteases comprising wild-type proteases, natural mutants, and drug-resistant proteases from nine different viral species. Through analysis of the proteases based on their ability to cleave 299 substrates, the researchers mapped out the physiochemical properties and cross dependencies of the amino acids of the proteases and their substrates.
By using this approach, they said they were able to analyze in detail the molecular-chemical mechanisms involved in substrate cleavage by retroviral proteases.
To compensate for the structural changes that retroviral proteases undergo in binding, Wikberg and his colleagues described each structurally aligned amino acid of the 61 retroviral proteases by their principal physiochemical properties rather than by their static 3D structures. They also described the retroviral protease substrates in the same manner by describing “the principal physiochemical properties of every single amino acid of the octapeptide sequence spanning the P4 to P4' position,” they say in their paper.
Because protease cleavage rates depend on the constituents of the experimental assay, additional assay descriptors were introduced. Cross terms — formed by multiplying two of the descriptors — can also be introduced into the multivariate modeling to account for the dynamic noncovalent and covalent interactions between the substrate and the enzyme that occur during substrate recognition and cleavage, the researchers say in the paper.
“The descriptors of the retroviral proteases, substrates, assays, and cross-terms were correlated to the experimentally determined substrate cleavage rates [Kcat/K m] using partial least squares regression modeling,” they say.
As a way of validating their model, they excluded data for eight retroviral strains one at a time in their entirety and predicted the excluded data using models constructed from the remaining data.
According to Wikberg, the analysis indicated the models could accurately predict the activities of the excluded retroviral proteases. Activity for the HIV-2 protease was predicted with 93 percent accuracy, while the predictive power for the HIV-1 protease mutants had an 86 percent rate of accuracy.
They then created a cleavability model to distinguish cleavable sequences from noncleavable sequences by “correlating substrate and protease descriptors and their cross terms to a vector representing cleavability or noncleavability.”
In work that has not been published, Wikberg and his colleagues tested their technology with existing HIV drugs and found that it could accurately predict the drugs’ ability to retard mutant HIV strains, Wikberg told ProteoMonitor.
Wikberg originally tested the proteochemometric technology on modeling multiple G-protein-coupled receptors. That work produced “great results,” Wikberg said, and about two years ago, he and his team started to test the technology on HIV proteases.
“For HIV proteases, there are 24,000 resistance mutations that are described, and that is impossible to manage. But we are doing now such studies with resistance mutations, and it’s possible to obtain models, and these models can be used to optimize the ligand for several resistance mutations, for example, at the same time.”
Since HIV was first diagnosed in the early 1980s, a cure has eluded researchers. While existing therapies have prolonged the life expectancies of HIV patients, the virus has also shown the ability to develop resistance to the drugs over time.
The number of deaths worldwide from HIV/AIDS has steadily risen every year since 1986 and in 2006 2.9 million people worldwide died from the virus, according to the Joint United Nations Program on HIV/AIDS and the World Health Organization.
In their paper, Wikberg and his colleagues say that the majority of protease inhibitors for the treatment of HIV have been peptide mimetics, especially ones designed against the HXB2 HIV-1 protease. However, such therapies have not been able to retard the replication of HIV strains that have mutated and become drug-resistant.
“Although efficiently hydrolysable protease substrates have served as excellent templates for peptide-mimetic inhibitor design, it is difficult to predict which combination of amino acids will make the best substrate over multiple proteases,” the authors say. “Analysis of protease mutations associated with drug resistance is also confounded by the existence of many viral subtypes carrying naturally occurring polymorphisms.”
The early models that Wikberg and his colleagues developed for HIV protease analysis were not very predictive, Wikberg said. That’s when his team decided to use data for other viruses that had been collected over 16 years and combine them into a single dataset. Using that dataset, they updated their model.
“With this technology, we can have an overview of everything that goes on for an unlimited number of targets,” Wikberg said. “For HIV proteases, there are 24,000 resistance mutations that are described, and that is impossible to manage. But we are doing now such studies with resistance mutations, and it’s possible to obtain models, and these models can be used to optimize the ligand for several resistance mutations, for example, at the same time.”
Because their technology is based on a general statistical model, Wikberg said it can be applied to predict similar mutations for other viral diseases. Wikberg said that several drug makers have shown interest in his technology, though he declined to elaborate.
“The problem with the pharmaceutical industry is the cost of developing a drug. [It’s not possible] with the present technology to develop a compound for a rare mutation,” he said. “But with this technology it may become feasible because it’s straightforward to do predictions. If one collects data systematically, the more accurate things become and the easier it becomes to do a new drug.
“The more data we have, the better the models become, the more predictive they are, and the cheaper it will be to make a variation of a compound, or a new compound,” he said.