Flexibility-aware graph-based algorithm improves antigen epitopes identification

Abstract

Epitopes of an antigen are the surface residues in the spatial proximity that can be recognized by antibodies. Identifying such residues has shown promising potentiality in vaccine design, drug development and chemotherapy, thus attracting extensive endeavors. Although great efforts have been made, the epitope prediction performance is still unsatisfactory. One possible issue accounting to this poor performance could be the ignorance of structural flexibility of antigens. Flexibility is a natural characteristic of antigens, which has been widely reported. However, this property has never been used by existing models. To this end, we propose a novel flexibility-aware graph-based computational model to identify epitopes. Unlike existing graph-based approaches that take the static structures of antigens as input, we consider all possible variations of the side chains in graph construction. These flexibility-aware graphs, of which the edges are highly enriched, are further partitioned into subgraphs by using a graph clustering algorithm. These clusters are subsequently expanded into larger graphs for detecting overlapping residues between epitopes if exist. Finally, the expanded graphs are classified as epitopes or non-epitopes via a newly designed graph convolutional network. Experimental results show that our flexibility-aware model markedly outperforms existing approaches and promotes the F1-score to 0.656. Comparing to the state-of-the-art, our approach makes an increment of F1-score by 16.3%. Further in-depth analysis demonstrates that the flexibility-aware strategy contributes the most to the improvement. The source codes of the proposed model is freely available at https://github.com/lzhlab/epitope. Author summaryEpitope prediction is helpful to many biomedical applications so that dozens of models have been proposed aiming at improving prediction efficiency and accuracy. However, the performances are still unsatisfactory due to its complicated nature, particularly the noteworthy flexible structures, which makes the precise prediction even more challenging. The existing approaches have overlooked the flexibility during model construction. To this end, we propose a graph model with flexibility heavily involved. Our model is mainly composed of three parts: i) flexibility-aware graph construction; ii) overlapping subgraph clustering; iii) graph convolutional network-based subgraph classification. Experimental results show that our newly proposed model markedly outperforms the existing best ones, making an increment of F1-score by 16.3%.