The Wonderful Wizard of AZT: Predicting the Clinical Success of New Drugs

By Matthew Luchette 

January 7, 2013 | The pharmaceutical industry could use a fortuneteller. According to a March 2010 issue of Nature Reviews Drug Discovery, drug companies spend on average $1.8 billion dollars to bring a new therapy to market, and that number is estimated to increase in the coming years. This staggering figure is due in part to the lack of predictive algorithms to help pharmaceutical companies accurately choose successful candidate compounds before human testing begins: as few as 1 out of every 5 drugs that enter FDA clinical testing is eventually approved for treating patients, and the process that can take nearly a decade.

With such a high rate of attrition and with millions of dollars wasted on developing and testing unmarketable compounds, many drug companies are looking to improve the way they select and develop candidate drugs. A “crystal ball” algorithm for predicting with certainty the clinical success of a drug at little-to-no cost to the pharmaceutical company remains elusive. Yet a Harvard project published in Nature Medicine in September 2012 may come closer than ever before.

The project, led by Harvard graduate students Daniel Scholes Rosenbloom and Alison Hill, began by investigating a confusing observation in HIV research: if HIV patients fail treatment when the viral population develops resistance to a drug, as is commonly believed, why do patients taking protease inhibitors (PI) fail treatment without developing resistance mutations in the viral protease gene?

To answer this question, the team designed a model that incorporates three main factors known to affect the clinical outcome of HIV patients: the dose-response relationship of the patient’s medications (or how sensitive the viral population is to a change in dosage), the patient’s adherence to medications, and the patient’s viral load when therapy begins. The group then used the algorithm to calculate how variations in these three parameters affected the growth of wild-type and mutant, drug-resistant viral populations.

As Rosenbloom explained, “The goal of the paper was to show that the evolution of a viral population over time can be explained in models.”

By relating the dose-response curves of various anti-HIV drugs to the probability the viral population will develop resistance to the drug, the team could explain why patients taking PIs can fail treatment without developing viral protease mutations.

“Due to the sharp slope of protease inhibitor dose-response curves, even relatively strong protease inhibitor resistance mutations are selected only in a narrow range of drug concentrations,” the article explains. In other words, because PIs can cause dramatically different effects on the viral population over a small range of concentrations, PIs protect patients from the virus’s potential to develop resistance. This effect is increased further by the relatively short half-life of PIs. As a result, patients taking PIs are more susceptible to virologic failure without the virus needing to develop resistance mutations should they miss scheduled doses.

This model, the team reports, is “the first explanatory model of virologic failure in agreement with clinical trials.” Instead of fitting their model to clinical data, the team used biological principles and laboratory measurements of HIV replication to predict the outcome of a therapy.

Rosenbloom believes the project could lay the groundwork in helping pharmaceutical companies accurately choose compounds that will succeed in clinical trials, saving both time and money wasted testing compounds that fail FDA testing. “Given a suite of drugs that have been proposed, our work could help companies develop algorithms that will test these drugs cheaply and quickly,” Rosenbloom explained. The model could also help clinicians develop personalized HIV therapies for a patient based on his viral genome, viral load, and history of drug adherence.

The algorithm is a valuable contribution to a growing area of research, called “model-based medicine.” According to Rosenbloom, most pharmaceutical models fall into two categories: molecular modeling models drug-protein interactions on the molecular level, and pharmacometric simulations models the interactions between a drug and a patient. This new model falls into a third category: dynamic population modeling, which focuses on describing the effect of a drug on a viral population over time.

One of the most important aspects of the model, Rosenbloom said, is that it “introduces evolution of resistance to the growing field of model-based medicine.” In a diverse, rapidly changing viral quasi-species, the model could help HIV researchers predict the effect of therapeutic interventions on a complex disease, helping drug companies have a better sense of a compound’s clinical success earlier in drug development.