Challenge: Protein-protein interactions (PPI) are well recognized as promising therapeutic targets. Although interfering peptides capable of inhibiting PPIs are receiving increased attention, the identification of those that possess high biological activity is very challenging due to their enormous diversity. There is thus a need to use bioinformatics and machine learning to effectively predict if a compound could be efficacious, which could lead to a decrease in the cost associated with the drug development process.

Solution: Using state-of-the-art machine learning algorithms, the team improved virtual screening of combinatorial linear and cyclic peptide libraries. The researchers assessed in silico the activity of millions of peptides against important PPI targets of human pathogens such as S. aureus (bacterial), C. albicans (fungal) and HIV-1 (viral). In this approach, the sequences of the best hundred inhibitory peptides are entered in a multi-cycle feedback approach combining machine learning screening, informed combinatorial chemistry and in vitro assays to further improve the properties of these peptides. Such learning loop (in silico selection combined with in vitro activity assessment) is then repeated up to three times to select peptides with the highest activity against the targeted PPIs.

Achievements/Impact: The team developed a new software application based on robust machine learning principles to accelerate the design and generation of optimized bioactive linear and cyclic peptides. The algorithm also accounted for the multiple potential binding sites present on a given protein. The team further validated their process by making an efficient inhibitor for the NEF protein of HIV1.