Prof Dirk Opperman, Associate Professor in the Department of Microbiology and Biochemistry at the University of the Free State (UFS), in collaboration with Dr Ana Ebrecht (a former postdoc in the same department) and Prof Albie van Dijk from the Department of Biochemistry at the North-West University (NWU), was part of an international collaboration of researchers who participated in solving an intricate problem in science – accurate protein structure prediction.

The team of researchers recently contributed to an influential paper describing new methods in protein structure prediction using machine learning. The paper was published in the prestigious scientific journal, Science.

“These new prediction methods can be a game changer,” believes Prof Opperman.

“As some proteins simply do not crystalise, this could be the closest we get to a three-dimensional view of the protein. Accurate enough prediction of proteins, each with its own unique three-dimensional shape, can also be used in molecular replacement (MR) instead of laborious techniques such as incorporating heavy metals into the protein structure or replacing sulphur atoms with selenium,” he says.

Having insight into the three-dimensional structure of a protein has the potential to enable more advanced drug discovery, and subsequently, managing diseases.

Exploring several avenues …

According to Prof Opperman, protein structure prediction has been available for many years in the form of traditional homological modelling; however, there was a big possibility of erroneous prediction, especially if no closely related protein structures are known.

Besides limited complementary techniques such as nuclear magnetic resonance (NMR) and electron microscopy (Cryo-EM), he explains that the only way around this is to experimentally determine the structure of the protein through crystallisation and X-ray diffraction. “But it is a quite laborious and long technique,” he says.

Prof Opperman adds that with X-ray diffraction, one also has to deal with what is known in X-ray crystallography as the ‘phase problem’ – solving the protein structure even after you have crystallised the protein and obtained good X-ray diffraction data, as some information is lost.

He states that the phase problem can be overcome if another similar-looking protein has already been determined.

This indeed proved to be a major stumbling block in the determination of bovine glycine N-acyltransferase (GLYAT), a protein crystallised in Prof Opperman’s research group by Dr Ebrecht, currently a postdoc in Prof Van Dijk’s group at the NWU, as no close structural homologous proteins were available.

“The collaboration with Prof Opperman’s research group has allowed us to continue with this research that has been on hold for almost 16 years,” says Prof Van Dijk, who believes the UFS has the resources and facilities for structural research that not many universities in Africa can account for.

The research was conducted under the Synchrotron Techniques for African Research and Technology (START) initiative, funded by the Global Challenges Research Fund (GCRF). After a year and multiple data collections at a specialised facility, Diamond Light Source (synchrotron) in the United Kingdom, the team was still unable to solve the structure.

Dr Carmien Tolmie, a colleague from the UFS Department of Microbiology and Biochemistry, also organised a Collaborative Computational Project Number 4 (CCP4) workshop, attended by several well-known experts in the field. Still, the experts who usually participate in helping students and researchers in structural biology to solve the most complex cases, were stumped by this problem.

Working with artificial intelligence

“We ultimately decided to turn to a technique called sulphur single-wavelength anomalous dispersion (S-SAD), only available at specialised beam-lines at synchrotrons, to solve the phase problem, says Prof Opperman.

Meanwhile, Prof Randy Read from the University of Cambridge, who lectured at the workshop hosted by Dr Tolmie, was aware of the difficulties in solving the GLYAT structure. He also knew of the Baker Lab at the University of Washington, which is working on a new way to predict protein structures; they developed RoseTTAaFold to predict the folding of proteins by only using the amino acid sequence as starting point.

RoseTTAaFold, inspired by AlphaFold 2, the programme of DeepMind (a company that develops general-purpose artificial intelligence (AGI) technology), uses deep learning artificial intelligence (AI) to generate the ‘most-likely’ model. “This turned out to be a win-win situation, as they could accurately enough predict the protein structure for the UFS, and the UFS in turn could validate their predictions,” explains Prof Opperman.

A few days after the predictions from the Baker Lab, the S-SAD experiments at Diamond Light Source confirmed the solution to the problem when they came up with the same answer.

Stunning results in a short time

“Although Baker’s group based their development on the DeepMind programme, the way the software works is not completely the same,” says Dr Ebrecht. “In fact, AlphaFold 2 has a slightly better prediction accuracy. Both, however, came with stunningly good results in an incredibly short time (a few minutes to a few hours),” she says.

Both codes are now freely available, which will accelerate improvements in the field even more. Any researcher can now use that code to develop new software. In addition, RoseTTAFold is offered on a platform accessible to any researcher, even if they lack knowledge in coding and AI.