Building better classifiers for reproducible science

RENCI’s Clark Jeffries recently presented a webinar for Orion Bionetworks called “Seeking Best Practices in Classifier Construction and Testing.”

Jeffries is a PhD-level bioinformatics specialist with an interest in interpreting neuroscientific information to better understand and treat psychiatric and neurological conditions. For years he has worked with researchers in the School of Medicine at UNC-Chapel Hill to analyze data and better understand debilitating diseases like schizophrenia. 

Research that involves data from patients with medical conditions always poses challenges. The researchers much keep the data safe and confidential, and they must construct experiments carefully and concisely in order to assure that other scientists can reproduce their results.

According to Jeffries, three factors affect reproducibility: case/control and selection of subjects; the consistency of lab assays; and consistent construction and testing of classifiers. Jeffries has focused much of his recent work on the third factor, developing stringent testing of classifiers to ensure that samples are classified correctly and are useful research tools.

You can read about Jeffries recent work with a team led by Diana O. Perkins, a professor of psychiatry in the UNC School of Medicine, here.  For more depth about building and testing classifiers, check out the Orion Bionetworks webinar here.

Classification methods for biological samples might sound a bit arcane to those of us who don’t spend our days in life sciences labs. But it’s that attention to detail that helps bring about breakthroughs – for example, a blood test that could help doctors determine if a patient showing psychiatric systems is likely to develop psychosis.