Integrating machine learning to advance epitope mapping

Abstract

Identifying epitopes, or the segments of a protein that bind to antibodies, is critical for the development of a variety of immunotherapeutics and diagnostics. In vaccine design, the intent is to identify the minimal epitope of an antigen that can elicit an immune response and avoid off-target effects. For prognostics and diagnostics, the epitope-antibody interaction is exploited to measure antigens associated with disease outcomes. Experimental methods such as X-ray crystallography, cryo-electron microscopy, and peptide arrays are used widely to map epitopes but vary in accuracy, throughput, cost, and feasibility. By comparing machine learning epitope mapping tools, we discuss the importance of data selection, feature design, and algorithm choice in determining the specificity and prediction accuracy of an algorithm. This review discusses limitations of current methods and the potential for machine learning to deepen interpretation and increase feasibility of these methods. We also propose how machine learning can be employed to refine epitope prediction to address the apparent promiscuity of polyreactive antibodies and the challenge of defining conformational epitopes. We highlight the impact of machine learning on our current understanding of epitopes and its potential to guide the design of therapeutic interventions with more predictable outcomes.

Keywords: B-cell; algorithm; databases; epitope; features; machine learning; toolboxes; vaccine.

Figure 1

Visualization of protein analysis for feature design. The protein structure of leporine serum albumin was visualized in ‘new cartoon’ with VMD (33). Coil and loop regions of the secondary structure are colored yellow and helical regions are colored purple. Certain features are derived from the analysis of the whole protein, smaller subsections, and sequence-based analysis. Features can broadly be categorized as structural, physicochemical, or sequence-based. The features provided below each of the categories are select examples that belong to each category. The secondary structure used as referenced PDB ID 4F5V (34).