Ivan Tanasijević1*, Branka Rakić1
1Institute for Artificial Intelligence R&D, Fruskogorska 1, 21000 Novi Sad, Serbia
ivan.tanasijevic [at] ivi.ac.rs
Abstract
Protein design plays a pivotal role in various scientific and industrial applications, such as drug development and biotechnology. With the advancement of computational methods, new tools and algorithms have emerged to facilitate the generation of novel protein designs. This study presents a comparative analysis of Pepspec and RFdiffusion, two prominent methods in protein design, to evaluate their effectiveness in designing peptides with desired properties. Mainly, we aim to design peptides that bind with high affinity and specificity to a desired protein target.
Pepspec is an application native to the Rosetta software package. It relies on Monte Carlo sampling of backbone confirmations and residue mutations and a stochastic optimization based on the Rosetta score – a measure approximating the binding free-energy of the complex.
On the other hand, a recently developed tool, RFdiffusion, is a denoising diffusion probabilistic model based on an existing artificial neural network, RoseTTAFold, developed for protein structure estimation. It is trained to remove noise from protein structures on a large database of protein complexes to ultimately be able to generate novel binder designs based on the target structure.
In this study, we aim to compare the efficiency of these two design tools. As it is common in generative ML algorithms, the comparison will be made by evaluating both the design quality and design versatility. The quality will be assessed by using the well-known AlphaFold2 Machine learning tool to estimate the binding affinity of the peptide-protein complex while the versatility will be measured using standard sequence based statistical methods.
RFdiffusion and Pepspec offer distinct approaches to protein design. By assessing the strengths and limitations of each method in this study, we aim to deepen the understanding of these methods and allow leveraging these tools effectively in designing peptides with desired characteristics, contributing to advancements in the field of protein engineering and biotechnology.
Keywords: Rational protein design, AI/ML in biology and medicine, Computational bioengineering
