Texas Advanced Computing Centre researchers established an AI-augmented pipeline using Frontera, resulting in a tenfold increase in target binding success.

Deep learning approaches were utilized to complement existing energy-based physical models in 'do novo' or from-the-ground-up computational protein design, resulting in a 10-fold increase in lab-verified success rates for binding a created protein with its target protein. The findings will aid scientists in developing better medications to combat diseases such as cancer and COVID-19.

The secret to comprehending proteins, including those that regulate cancer, COVID-19, and other disorders, is relatively elementary. Determine their molecular makeup and the proteins that bind to them. However, there is a caveat.

"The search space for proteins is enormous," said Brian Coventry, a research scientist at the Institute for Protein Design at the University of Washington and The Howard Hughes Medical Institute. 

Designed protein

Considering that there are 20 possible amino acids for each location in a protein and that there usually are 65 amino acids in a protein investigated in his lab, there are more binding combinations than atoms in the universe. The researchers improved the success rate of binding a designed protein with its target protein in the lab by a factor of ten by using deep learning approaches to complement existing energy-based physical models in 'do novo' or from scratch computational protein creation. 

Deep learning applications

Popular deep-learning applications that readers may know include the language model ChatGPT and the picture generator DALL-E. Deep learning uses computer algorithms to analyze and infer patterns in data, layering the algorithms to extract higher-level characteristics from raw input. Deep learning algorithms were employed in the study to learn repetitive transformations of the protein sequence and potential structure that fast settle on accurate models.

The authors' deep learning-augmented de novo protein binder design approach used the machine learning software tools AlphaFold 2 and the RoseTTA fold created at the Institute for Protein Design. The study's principal author, head of the Institute for Protein Design, and Howard Hughes Medical Institute investigator David Baker had access to the Texas Advanced Computing Centre's NSF-funded Frontera supercomputer.

Since the protein design trajectories are entirely separate, no data had to be passed between them while the compute jobs were executing, making the study problem a good candidate for parallelization on Frontera. 

Conclusion

According to Coventry, there is still a long way to go, even if the study results demonstrated a 10-fold improvement in the success rate of created structures to bind on their target protein. There are still three orders of magnitude for us to overcome. He said that future studies will focus on further improving that rate and moving on to a new class of much more challenging targets. Cancer T-cell receptors and viruses are two good examples. Further optimization of the software tools or additional sampling will lead to better computationally generated proteins.