Performance Analysis of Deep Learning Models on Chemokines Protein Group Using Structure-Based Pattern Detection

Abstract

The recent success of deep learning-based models such as AlphaFold 2 on the prediction of structure from protein sequences has made structural assessment much easier. Structural prediction from these models can be used in many protein-based types of research like identifying misfolded protein, protein docking, etc. much quicker than traditional methods. Recent studies include benchmark studies of nanobodies, empirical analysis on Cytokines protein groups, and so on. These studies compare different protein folding models against each other to evaluate which models perform the best. However, no further investigation has been done to detect anomalies or protein that consistently performs poorly across all models. Additionally, no similar research has been done on the Chemokines protein group. In this work, we have assessed the performance of AlphaFold 2 on the Chemokines protein group and attempted to detect patterns in which the model performs very poorly. The result highlights protein groups with patterns in which AlphaFold 2 performs poorly.