Predicting Protein-Protein Interactions using Machine Learning (AI)

Proteins are the workforce and main building component of the human body. However, instead of working individually, proteins commonly collaborate to fulfill their tasks. Therefore, to fully understand how proteins function, it is crucial to investigate how they interact. Numerous experimental techniques exist to explore interactions between proteins. The problem with these methods is that they either look at very few interactions at a time or generate many false positives (falsely predict non-existing interactions). Different experimental methods may also generate contradictory findings. Due to these shortcomings, it is necessary to develop complementary computational methods. The break-through in this field came when Deepmind released the machine learning method AlphaFold2. Despite the excellent performance of AlphaFold2 there is still work to be done. AlphaFold2 has problems when predicting unstable interactions as well as antibody-antigen interactions. AlphaFold2 is also slow and computationally expensive making it infeasible when investigating millions of interactions. The research in Arne Elofsson’s group consists of both improving AlphaFold2 and developing novel algorithms in the areas where AlphaFold2 is currently lacking.