The use of machine learning to understand the relationship between IgE to specific allergens and asthma

In a study recently published in PLOS Medicine, Custovic and colleagues report an elegant analysis of the results of specific Immunoglobulin E (IgE) assays on sera from a birth cohort that they have studied in detail over the last 15 years. The IgE assays used a microchip with >100 purified allergens, and their machine learning approach used a hypothesis-free statistical approach to group children with similar allergen sensitization profiles. The primary objective was to evaluate the relationship of the results to the diagnosis of current asthma. The goal of employing a “machine learning” approach is both as a form of data reduction to avoid spuriously identified associations resulting from multiple comparisons and to identify biologically relevant groupings that may not be recognized with conventional approaches. The use of “hypothesis free” clustering methods with individual allergen components in itself is not novel. However, the authors have employed a new network analysis method, which is improved over previous methods because it can capture nonlinear relationships and does not rely on assumption of a parametric probability. As the authors show, this approach can lead to improved sensitivity and specificity. With this model, they were able to show that pairings of allergens (e.g., the cat allergen Fel d 1 and the peanut allergen Ara h 1) were more predictive of the child having asthma than individual components, although we must be cautious in the generalizability of the specific patterns identified, in this single cohort in the United Kingdom, because the pairing may represent host susceptibility to making an IgE response or common co-exposures, which could be allergens or adjuvants. Still, the concept of complicated connections between individual allergenic proteins and allergic disease are likely to apply more broadly and are important to consider in future studies