At last week’s annual LabAutomation Conference, pharmacogenomics was a sizzling buzzword. It was also one of those areas where many attendees said they wanted to get into, but few professed to know much about.
Still, by the time the conference wound up last Wednesday, few attendees hung around to hear Chris Herold, of San Diego startup Prediction Sciences, discuss the pharmacogenomics technology he hopes will turn the world of blockbuster antidepressants on its head.
Those who left early might just regret not staying.
Herold, who directs Prediction Sciences’ CNS-diagnostic group, discussed the development of GeneRx, a predictive diagnostic the company is developing to help doctors predict in advance which antidepressant will work best.
The diagnostic is being designed to look for signature haplotypes of non-responders to common antidepressants — the selective serotonin reuptake inhibitors, which include Prozac, Zoloft, Paxil, and Celexa.
The basic idea behind the technology, which aims to combine haplotyping data from a patient’s genotyped blood sample with clinical data on response, is to find signatures that predict response to a particular drug. (Herold did not say who is providing the company with patient samples.)
According to Herold, the technology relies on a supervised learning algorithm, called a neural net, which can learn to distinguish the haplotype of responder versus non-responder, or a patient likely to experience toxic side effects versus those who can safely metabolize the drug.
It would be nice to find a few “golden SNPs,” said Herold, but in most cases a handful of SNPs does not contain sufficient predictive ability, so it is necessary to look at a large block of SNPs, or haplotype. The problem, he said, is when the difference between the responder and non-responder, or toxic versus safe haplotypes, is not clear.
This is where the neural network comes in. A neural network — which has applications in widely diverse areas of computing from signal processing to financial engineering to face recognition — includes an input layer, a hidden layer and the output layer.
A training set of data with known inputs and outputs — SNPs and response/non-response — is input into the neural net in a serial fashion so the net can learn to detect the outcome based on the input. Then a test set, for which the outcome is known in advance by the researcher, is put into the input layer without giving the neural net the output. Based on the training set, the neural net is supposed to be able to predict the outcome in the test set.
In a project funded by a $100,000 NSF SBIR grant, Herold and his colleagues have applied this combination of haplotype, clinical data, and neural networks to dev- elop a predictive diagnostic for response to the antidepressant Celexa, made by Forest Pharmaceuticals.
First, the researchers haplotyped samples from 105 patients given Celexa for four weeks. They had data on which patients were responders and which were non-responders The haplotypes focused on 91 SNPs, most of which were related to serotonin transporters and receptors, and which were located near intron-exon boundaries. Then the researchers took 67 of the patients, evenly divided between responders and non-responders, and used them as the training set for their neural network.
After training the network to correctly correlate haplotype with patient response in 75.3 percent of the training set, the group then used the algorithm on 29 patients in a test. They were able to correctly predict response to the drug in 72.2 percent of the cases. Finally, they used an evaluation set of nine other patients, in which they were able to predict response in 77.8 percent of the cases.
Herold acknowledged that the neural network has a way to go as a prediction device. “This is phase one,” he said. Next, the company hopes to obtain a larger dataset in order to refine the neural net. “If we can get to 85 to 90 percent, [this] definitely will be worthwhile as a predictive diagnostic.”
Meanwhile, the company is pressing on with work on GeneRx for other drugs, including antihypertensives and other psychotropics. Earlier this month, the firm won a $100,000 California BioSTAR grant to develop a diagnostic to predict response to hypertension drugs. The study is being conducted by Daniel O’Connor, a professor of medicine at the University of California, San Diego, one of the company’s cadre of academic collaborators.