A Novel Epitope Dataset: Performance of the MCL-Based Algorithms to Generate Dataset for Graph Learning Model

Naturally, the epitope dataset can be presented as a graph. Dataset preparation in the previous methods is part of model development. There are many graph-based classification and regression methods. Still, it is difficult to identify their performance on the conformational epitope prediction model because datasets in a suitable format are unavailable. This research aims to build a dataset in a suitable format to evaluate kernel graph and graph convolution network. This dataset, which results from graph clustering on graph antigens, can be used to identify the performance of many graph neural network-based algorithms for conformational epitope prediction. The Ag-Ab complexes that meet the criteria for forming a conformational epitope prediction dataset from previous studies were downloaded from the Protein Data Bank. Raw datasets in the form of specific exposed antigen chain residues are labeled as epitope or non-epitope based on their proximity to the paratope. The engineering features in the raw dataset are derived from the structure of the antigen-antibody complex and the propensity score. Aggregating atomic-level interactions into residual levels create an initial graph of the antigen chain. The MCL, MLR-MCL, and PS-MCL are graph clustering algorithms to obtain labeled sub-clusters from the initial graph. A balance factor parameter is set to several values to identify the optimal dataset formation based on minimal fragmentation. The output of the MCL algorithm is used as a baseline. As a result of the fragmentation analysis that occurs, the MLR-MCL algorithm gives the best model performance at a balance factor equal to 2. PS-MCL gives the best performance at a value of 0.9. Based on the minimum fragmentation, the MLR-MCL algorithm provides the best model performance compared to MCL and PS-MCL. The dataset in a format according to benchmarking dataset can be used to identify the characteristics of antigen subgraphs formed from the graph clustering process and to explore the performance of graph-based learning conformational epitope prediction models such as graph convolution networks.