AI augments humans to lead them through the (protein) crystal maze

I spy with my little AI, something hidden under the microscope


AI can help scientists spot tiny folding protein crystals, and thus one day potentially assist eggheads in designing new drugs, according to a paper published in PLOS One.

To demonstrate this form of boffinry is possible, a large team of researchers from academia and industry, including bods at Duke University in the US and Google and British pharmaceutical giant GlaxoSmithKline, built a convolutional neural network to recognize microscopic protein crystals.

The project was launched by the Machine Recognition of Crystallization Outcomes (MARCO) initiative, an international effort to collect pictures of protein crystals from X-ray crystallography experiments.

Over time, 493,214 images of protein crystals were harvested, and the dataset was shared with Google researchers to train and test the neural network. After testing it on 50,284 images, the system was found to be 94 per cent accurate in detecting the presence of protein crystals in a solution.

Proteins are complex molecules made up of a string of amino acids. They play a vital role in carrying out bodily functions such as the transmission of hormones and protecting it against viruses.

One crucial step in studying and understanding proteins, and using this to develop medication, involves turning these molecules into crystals. The trouble is, the crystals are not easy to spot in lab conditions, hence the need for a machine-learning system to help detect them.


Boffins build neural networks fashioned out of DNA molecules


“Once a protein structure is known, its biological function can be understood, which may lead to the development of new therapeutic treatments,” according to the MARCO initiative. "Obtaining protein crystals is thus an important first step to uncovering the origin of life and to curing diseases."

This is a tedious task. In order to classify proteins, scientists first have to isolate the proteins and then grow them just right in the correct recipe of liquid solutions. It’s a rare chance that more will form in the concoction, and crystallographers can often miss them when studying drops of the liquid under a microscope.

This is where the neural network comes in. It helps scientists study the images of the drops, and sorts them out into four classes: crystals, precipitate, clear, and other. An autonomous algorithm, ideally, reduces errors, and speeds up research.

“Every time you miss a protein crystal, because they are so rare, you risk missing on an important biomedical discovery,” said Patrick Charbonneau, an associate professor of chemistry at Duke University and leader researcher of the MARCO initiative, at the end of last month.

The data has been made public, and the team has open sourced its model. It was written using TensorFlow, and required 50 Nvidia K80 GPUs to train the network for 19 hours. ®

Biting the hand that feeds IT © 1998–2018