Current mathematical techniques have reached their limits, but deep learning promises to totally transform how analyses are done, writes chemistry professor Jean-François Masson.
Analytical chemistry, which involves identifying and quantifying molecules in lab samples, is used in a wide range of fields, including medicine, environmental analysis, drug testing and pharmaceutical quality control.
There are hundreds of thousands or even millions of types of molecules. Analyzing them is essential for understanding how illnesses work at the molecular level and find new treatments. Modern techniques such as chromatography, mass spectrometry and spectroscopy are used to separate and identify molecules based on their structures and chemical bonds. However, when molecules are similar, they can be hard to distinguish. This hinders our understanding of the molecular mechanisms of disease, which in turn prevents us from finding effective treatment options.
A revolution on the horizon
Deep learning could transform the way we analyze chemicals. The techniques used in analytical chemistry create spectra or chromatograms – essentially, photographs of molecules – from which information is extracted. The problem, however, is that the images are highly complex and the mathematical techniques currently used to extract data have reached their limit. In fact, no major improvements have been made in the last decade.
Now the hope is that deep learning will enhance analytical chemistry techniques and allow better quality data to be extracted. In so doing, it will change the process for discovering which molecules cause illness and help us better describe the molecular composition of diseased cells and tissue. By improving our understanding of how illnesses disrupt chemical composition, we may be able to develop more effective treatments – and possibly even personalized treatments.
New neurochemical analysis
Deep learning has already changed our research methods. My team at UdeM worked with neuroscientific researchers to develop a new neurochemical analysis technique called optophysiology. It detects neurotransmitters by spectroscopically measuring their vibrations. Each neurotransmitter has a unique footprint that conventional mathematical techniques were able to process with only modest success.
Deep learning has vastly improved the results we get from optophysiology. We are now able to identify a dozen neurotransmitters in a single experiment, while traditional techniques used in the neurosciences can only identify one or two.
Thanks to artificial intelligence, neuroscientists are now able to obtain better results from analytic screening tools and gain a better understanding of how the brain works. We are hopeful that this will someday also lead to better drug treatments for neurodegenerative and psychiatric illnesses.