Description of clinical trial study population. We enrolled 41 patients with CMA from March 2018 to October 2019 (Figure 1), as detailed previously (25). Eleven were ineligible, as they failed to react during their screening DBPC food challenge (DBPCFC) to 444 mg (<1 tablespoon) of BM (26). Thirty patients met all eligibility criteria and were randomized 1:1 to BM or placebo OIT during the first year of treatment (Figure 1 and Figure 2). Twenty-eight participants completed the 12-month BM DBPCFC. All were then invited to receive an additional 12 months of open-label BMOIT. Twenty-seven started open-label BMOIT, with 24 completing the month 24 (end of treatment) BM DBPCFC and 22 completing the month 24 UM DBPCFC. The last participant completed their final DBPCFC in February 2022.

CONSORT diagram. Flowchart of participants’ disposition throughout the study.

Trial design and time points for outcomes. Timeline showing baked milk (BMOIT) and placebo oral immunotherapy (OIT) groups and timing of blood draw (mechanistic studies), skin prick tests (SPT), and double-blind placebo-controlled food challenge (DBPCFC) to baked (BM) and unheated (UM) milk.

Demographic and other baseline characteristics of the 30 randomized participants are summarized in Supplemental Table 1 (supplemental material available online with this article; https://doi.org/10.1172/jci.insight.184301DS1) and were previously published (25). Sixteen (53%) were male and 14 (47%) were female. The median age at enrollment was 11 years (range: 3–18 years). All had a prior reaction history to some form of milk (63% UM and BM, 27% UM only, and 10% BM only). Twenty-six (87%) had multiple food allergies. There was high prevalence of other atopic diseases at enrollment, with 53% having atopic dermatitis, 67% asthma, and 70% allergic rhinitis.

During year 1, all 30 randomized participants completed the initial dose escalation (IDE) and tolerated the required minimum dose of 3 mg of BM/placebo. Participants then began daily OIT home dosing. Two participants withdrew during buildup in year 1 (related to starting college and family reasons) and hence 28 of 30 (93%) underwent the 12-month BM DBPCFC (25). After the 12-month DBPCFC, the groups were unblinded and all participants were offered open-label BMOIT for an additional 1 year. One participant in the active group withdrew after the 12-month BBPCFC, prior to starting year 2 due to taste/not wanting to eat more muffins; therefore, 27 participants received open-label BMOIT during the second year of treatment (14 initial placebo group, 13 initial active group). In year 2, two participants withdrew from the placebo crossover group (1 due to GI symptoms, 1 due to moving) and 1 withdrew from the initial active dose group due to symptoms with dosing and starting college.

Twenty-four participants (12 of each group) completed the end of treatment, month 24 BM DBPCFC. One participant in each group (n = 2) did not tolerate 2000 mg of BM and were not eligible for the UM DBPCFC.

Clinical trial efficacy outcomes. At baseline, the median maximum tolerated dose (MTD) was 44 mg of BM protein (approximately one-quarter teaspoon) in the active group and 144 mg (approximately 1 teaspoon) in the placebo group. At month 12 and month 24, participants underwent a DBPCFC to up to a cumulative dose of 4044 mg of BM protein (approximately a half cup). In the month 12 BM DBPCFC, 11 of 15 (73%) in the active group compared with 0 of 15 in the placebo group tolerated 4044 mg of BM (P < 0.001) (Figure 3) (25). In the month 24 BM DBPCFC, 9 of 15 (60%) of the initial active group and 10 of 15 (67%) of the placebo crossover group tolerated 4044 mg of BM, with no significant difference between groups (Figure 3 and Supplemental Table 2). In the per-protocol (PP) analysis, the majority of participants in both groups tolerated 4044 mg of BM (9 of 12 [75%]) in the initial active group compared with 10 of 12 (83%) in the initial placebo group. Coincidently, when we collapsed the initial treatment groups and analyzed the PP population based on time on treatment (0, 12, and 24 months), we found that 21 of 26 (81%) tolerated 4044 mg of BM after 12 months of active BMOIT (Supplemental Figure 1). This was significant compared with the baseline, as was the comparison of 24 months to baseline for the individuals in the initial active BMOIT group (Supplemental Figure 1).

Efficacy outcomes. Percentage of each group tolerating predefined total cumulative dose of baked milk (months 12 and 24) and unheated milk (month 24) protein by treatment group for intent to treat (ITT) and per-protocol (PP) populations. There was a significant difference (P < 0.05) between groups only at month 12. For the ITT analysis, n = 30 (month 12 and month 24). For the PP analysis, n = 28 (month 12 baked), n = 24 (month 24 baked), and n = 22 (month 24 unheated). Statistical analyses were performed using Fisher’s exact test.

Twenty-two participants, 11 in each group, underwent the 24-month UM DBPCFC, with a possible maximum cumulative dose of 8030 mg of milk protein (approximately 1 cup). Eight of 15 (53%) in the initial active group compared with 5 of 15 (33%) in the placebo crossover group tolerated 2000 mg or more cumulative UM protein (P = 0.46), while 4 of 15 (27%) versus 0 of 15, respectively, tolerated the maximum cumulative dose of 8030 mg of UM protein (P = 0.1) (Figure 3). In the PP analysis, in the initial active group, 8 of 11 (73%) tolerated 2000 mg or more cumulative UM protein and 4 of 11 (36%) tolerated 8000 mg compared with 5 of 11 (45%) and 0 of 11 in the initial placebo group, respectively (P = 0.39 and 0.0) (Figure 3).

We also evaluated the MTD of BM and UM (Figure 4A). At baseline, the median MTD of BM was 44 mg of CM protein (range: 4–144 mg) in the initial active group compared with 144 mg (range: 4–144 mg) in the initial placebo group (P = 0.25). At month 12, the initial active group had a median change in BM MTD from a baseline of 3900 mg compared with 0 in the placebo group (P = 0.0001) (25). At month 24, both groups had a median change in BM MTD from a baseline of 3900 mg, with no difference between groups (12 vs. 24 months of treatment) (Figure 4A). In those who underwent the 24-month UM DBPCFC, the median MTD of UM was 2780 mg, with a range of 430 mg to 8030 mg (Figures 4B and Supplemental Figure 2). There was a significant difference in the median MTD of UM protein after 24 versus 12 months of treatment based on the PP analysis (median MTD of 5530 mg initial active group vs. 1030 mg in the initial placebo group, P = 0.01), but this was not statistically significant in the intention-to-treat (ITT) analysis (3530 mg vs. 1030 mg, P = 0.11) (Figure 4B).

Maximum cumulative tolerated dose. (A) Median value of the maximum cumulative tolerated dose of baked milk at baseline, month 12, and month 24 split by initial treatment group in the intent-to-treat (ITT) population (n = 30). (B) Maximum tolerated dose of unheated milk at month 24 for each participant (dots) split by initial treatment group. Group medians are indicated by the bars. Left panel is ITT population (n = 30). Right is per-protocol (PP) analysis (n = 22). Statistical analyses were performed using Mann-Whitney U tests. *P < 0.05.

In addition, we reviewed the severity of symptoms during the 24-month DBPCFC to both BM and UM. Severity of symptoms during the DBPCFC to BM and UM was determined based on the CoFAR Grading Scale for Systemic Allergic Reactions version 3.0 (27), which has 5 levels of increasing severity, ranging from mild symptoms involving 1 organ system (Grade 1) to death (Grade 5). Five of 24 participants who underwent the 24-month BM oral food challenge (OFC) had a reaction with 2 Grade 1 reactions and 3 Grade 2 reactions (all Grade 2 due to mild symptoms in 2 organ systems). Eighteen of 22 participants had a reaction during the UM OFC. There were 6 Grade 1 reactions, 9 Grade 2 reactions, and 3 Grade 3 reactions. There were no Grade 4 or Grade 5 reactions in any of the OFCs.

Overall, these analyses of food challenge outcomes at the 24-month time point showed that BM OIT appears to be effective at desensitizing patients to both BM and some amount of UM. There were some trends, but no statistically significant differences for greater efficacy in the initial active group over the initial placebo group (24 vs. 12 months of treatment).

Clinical trial adverse events. Safety and adverse events (AEs) were assessed throughout the trial. Safety data from the first year of blinded treatment was previously reported (25). During the open-label period, 9 of 13 in the initial active group and 14 of 14 in the placebo crossover group had at least 1 AE (incidence rate ratio [IRR]: 0.69, 95% CI: 0.26–1.72) (Supplemental Table 3). The average number of AEs per person during the open-label treatment year was 66 in the initial active group (range: 0–379) and 19 in the placebo crossover group (range: 3–68).

Overall, during open-label treatment, dosing-related reactions were common, but typically mild. There were symptoms with 1043 of 8914 (12%) BMOIT doses (Table 1). Eight of 13 in the initial active group compared with 14 of 14 in the placebo crossover group had at least 1 dosing-related AE in year 2 (IRR: 0.62, 95% CI: 0.22–1.57) (Supplemental Table 3). Greater than 98% of dosing-related reactions were mild and there were no severe reactions. When considering all participants, 20 of 27 (74%) had any AE with the highest severity of mild and 26% had AEs with the highest severity of moderate. The most common symptoms were oropharyngeal (OP) and GI (Table 1). The initial active group reported more OP symptoms (IRR: 8.03, 95% CI: 6.40–10.15) and GI symptoms (IRR: 1.82, 95% CI: 1.36–2.45) than the placebo crossover group in year 2. No significant difference in skin or lower respiratory tract symptoms was found between groups.

Dosing-related adverse events (AEs) by treatment group (percentage of doses)

For the initial active group, the number of dosing-related reactions decreased over time, with fewer dosing-related reactions in year 2 compared with year 1, both overall (18.6% vs. 42% of doses; IRR: 0.44, 95% CI: 0.41–0.48) and in the maintenance phase (18.6% vs. 37% of doses; IRR: 0.50, 95% CI: 0.46–0.55) (Supplemental Table 4). Although OP and GI remained the most common, both significantly decreased compared with year 1 (OP-IRR: 0.54, 95% CI: 0.49–0.59; GI-IRR: 0.19, 95% CI: 0.16–0.23).

Within the placebo crossover group, there was a higher rate of dosing-related AEs during buildup compared with maintenance (IRR: 2.74, 95% CI: 1.9–4.0; median 10.5 AEs per person in buildup compared with 3 in maintenance) and more doses that required any medication during buildup (IRR: 2.21, 95% CI: 1.18–4.44) (Table 1).

Approximately 1% of doses (102 of 8914) required treatment for a dosing-related reaction, with the most common treatment being antihistamines. Compared with the placebo crossover group, the initial active group required fewer doses to be treated with any medication (IRR: 0.57, 95% CI: 0.37–0.87), antihistamines (IRR: 0.58, 95% CI: 0.37–0.92), or oral steroids (IRR: 0.13, 95% CI: 0.003–0.94), with no significant difference in use of albuterol or epinephrine. Epinephrine was used for 1 dosing-related reaction and a total of 4 times (3 during buildup and once in maintenance) by 3 individuals, all in the placebo crossover group (Supplemental Table 5).

Overall, we found that dosing-related reactions were common, with the most common symptoms being OP and GI; however, more than 98% of reactions were mild, and reactions appeared to decrease over time.

Measurement of immunologic outcomes longitudinally across clinical trial. Antigen-specific antibodies and CD4+ memory T cells were measured longitudinally throughout the trial (Figure 2). We sought to determine which immunologic measurements, if any, were associated with clinical outcomes. Data on skin tests and serologic biomarkers in year 1 have been previously reported (25). Here, we describe those measures for year 2 as well as detailed cellular immune studies from month 0 through month 24.

Selection of antigen-specific CD4+ T cell populations from flow cytometry and scRNA-Seq datasets to predict BMOIT treatment versus placebo while blinded for treatment groups. To examine the mechanisms of tolerance during BMOIT, we leveraged our previously published method (28) of isolating CM-specific CD4+ memory T cells expressing CD154 and/or CD137 (CM+). Briefly, we stimulated total PBMCs with a pool of T cell–reactive CM peptides for 6 hours followed by flow cytometry to examine surface markers and further sorting of CM+ memory CD4+ T cells (CD154+ and/or CD137+) for 10× Genomics scRNA-Seq (Figure 5A). We additionally sorted CM– (CD154–CD137–) cells from each sample as a control. We performed this assay on PBMCs from the 28 individuals longitudinally (5 time points; not all 28 individuals had PBMCs at every time point) across the BMOIT clinical trial (Figure 2). We included DMSO controls to show significant induction of the CM+ cells over background (Supplemental Figure 3, C and D).

Blinded selection of CD4+ T cell populations from flow cytometry and scRNA-Seq assays for prediction of BMOIT treatment group. (A) Graphical representation of assay to identify/isolate CM-specific T cells. (B) Description of population selection analysis steps. (C) Flow cytometry populations selected for blinded analysis. Flow plot example of CD4+ memory CM+ populations. (D) Bar plots showing paired analysis of each selected population from month 0 to 24. (E) scRNA-Seq populations selected for blinded analysis. UMAP plots of CM+ clusters and FOXP3 expression in those clusters. (F) Bar plots showing paired analysis of each selected population from month 0 to 24. We performed this assay on PBMCs from 28 participants. Statistical analyses were performed using paired Wilcoxon’s tests. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.

Given the high dimensionality of our dataset, it is important to avoid overfitting due to multiple hypothesis testing. Therefore, we focused our analysis on (a) predefined T cell populations that we had previously identified as modulated in a cross-sectional study of allergic versus non-allergic individuals (28), (b) populations that have been associated with tolerance, and (c) CM+ T cell populations and the scRNA-Seq clusters they are derived from (Supplemental Table 6). The initial analysis was performed in a blinded fashion to the initial treatment group of each individual. We checked each cell population for significant changes between baseline (month 0) (no individuals on treatment) and month 24 (all individuals on treatment) and only moved forward with those that had a significant change in either direction (Figure 2, Figure 5B, and Supplemental Table 6).

For the flow cytometry analysis, 7 predefined populations showed significant differences with BMOIT (Figure 5, C and D, Supplemental Table 6, and Supplemental Figure 3). These included CM+ memory CD4+ populations defined by expression of CD154 and CD137 (CM+, total CD154+, CD154+CD137–, and CD154+CD137+), which all decreased significantly from month 0 to 24 (Figure 5D). The gating for these populations was set to replicate the 10× sorted data, so are less strict and may include non–antigen-specific cells. In Supplemental Figure 3, we include a strict gating strategy to show percentages for these gates and show that these populations do not outperform the less strict gates in predicting outcomes (Supplemental Figure 3, E–G). We additionally included CM+ CD127–CD25+ cells, which we previously identified to be increased in CMA versus non-CMA individuals using the same methods (28). Interestingly, this population increased from month 0 to 24 (P = 0.0002; Figure 5D). As total CM+ cells decreased over the time period, we included a not-gated (NOT) sample from the CM+ CD127–CD25+ population and back-calculated the percentage of total CD4+ memory cells (Figure 5D). Non–antigen-specific Tregs were also gated by CD4+CD127–CD25+, as these cells have been previously associated with tolerance. These cells increased significantly from 0 to 24 months (P = 0.0104; Figure 5D).

The analysis of the scRNA-Seq dataset was performed following the same processing steps we previously described (28) and detailed in the Methods. Briefly, clustering was performed on CM+ and CM– sorted fractions together to determine overlap, and true CM+ cells were defined as those falling into clusters that were made up of more than 80% CM+ sorted cells. This analysis revealed 6 scRNA-Seq clusters with distinct gene expression profiles (Figure 5, E and F, Supplemental Table 6, and Supplemental Figure 4, A–E). Three of the CM+ clusters showed significant increases or decreases from 0 to 24 months (C3, C4, C5). CM+ cluster C3 was FOXP3+ MHCII markers and increased over time (P = 0.002). Clusters C4 and C5 were both FOXP3–, were distinguished by CCR7+ (C4) and Th1/Th17 marker expression (C5), and decreased over time (P = 0.009 and P = 0.0001, respectively; Figure 5, E and F, and Supplemental Figure 4E). CM+ cells expressing FOXP3 (CM+FOXP3+) were included based on our previous findings and significantly increased from 0 to 24 months (P = 0.0009; Figure 5, E and F). CM+ cells expressing pathogenic Th2/Tfh markers (CM+ Th2A: GATA3, IL4, IL5, IL13, IL9, CCL1, IL3, IL13, PLAC8, CSF2, HPGDS, CRLF2, PPARG, IL1RL1, PTGS2, IL17RB, KLRB1, HDC, and H2AFZ) previously shown to be associated with CMA (28) were included and significantly decreased over time (P = 0.008; Figure 5F and Supplemental Figure 4F). The CM+FOXP3+/Th2A cell ratio was calculated and showed a significant increase from 0 to 24 months (P = 0.003; Figure 5F). Detailed descriptions of all populations from the flow cytometry and scRNA-Seq analyses are listed in Supplemental Table 6.

CM+ CD4+ T cell populations defined by scRNA-Seq best determine BMOIT treatment group. With the final 7 flow cytometry and 6 scRNA-Seq T cell populations selected, we were unblinded to the groups to evaluate how well the selected populations could differentiate initial active BMOIT versus initial placebo after the first 12 months. We evaluated the performance by ROC analysis using the percentage of each flow cytometry (Figure 6A) or scRNA-Seq (Figure 6B) population per individual at the 12-month time point. None of the populations defined by flow cytometry had significant predictive power (AUC < 0.6; Figure 6A). In contrast, the scRNA-Seq populations had several predictors with AUC greater than 0.6, with the best being the CM+FOXP3+/Th2A ratio (AUC = 0.779; Figure 6B). Importantly, this measure outperformed the predictive power of antibody measurements taken at the same time point, where CM IgG4 performed best (AUC = 0.643; Figure 6C). The populations that performed best in the ROC analysis (CM+FOXP3+/Th2A, CM+ C3, CM+ Th2A) were also significantly different in a direct comparison between placebo and treatment groups at the 12-month time point, where higher percentages of CM+FOXP3+/Th2A and CM+ C3, and lower percentages of CM+ Th2A cells distinguished initial active versus initial placebo (Figure 6D). No other time point comparisons for any population reached significance (Supplemental Figure 5).

Evaluation of selected CD4+ T cell populations to determine BMOIT treatment group. (A–C) ROC analysis of (A) flow cytometry, (B) scRNA-Seq, and (C) antibody measurements to determine baked milk oral immunotherapy (BMOIT) treatment and placebo groups at the 12-month time point. Area under the curve (AUC) values are listed for each population under their respective plot. (D) Line plot analysis of placebo versus treatment group at 0-, 6-, and 12-month time points for selected populations. Each line represents a participant and the bolded line is the mean of the indicated placebo/treatment group. Statistical analyses were performed using Mann-Whitney U tests (n = 28). *P < 0.05.

CM+ CD4+ T cell populations redistribute with 12 months of BMOIT. We next analyzed changes in immune measurements after 12 months of BMOIT treatment, combining both treatment arms (initial active and initial placebo) to increase statistical power. With all 28 individuals, 12 months of BMOIT treatment resulted in significant increases or decreases in all of our selected CD4+ T cell (flow and scRNA-Seq) populations and almost all measured antibody titers (Figure 7A). From the flow cytometry analysis, we saw significant increases in Tregs and CM+ CD127–CD25+ cells and decreases in CM+, CM+ NOT, CD154+ total, CD154+CD137+, and CD154+CD137– populations with BMOIT (Figure 7B). For the scRNA-Seq populations, we saw significant increases in CM+FOXP3+, CM+ C3, and the CM+FOXP3+/Th2A ratio and decreases in CM+ Th2A, CM+ C4, and CM+ C5 (Figure 7C). For some of the populations including Tregs, the CM+FOXP3+/Th2A ratio, and CM+ Th2A, the change in population percentage was already significant at the 6-month time point (Figure 7, B and C). Of note, there was no significant change in any T cell population from 12 months on treatment to 24 months on treatment (Supplemental Figure 6).

CM-specific antibody and CD4+ T cell populations change significantly through 12 months of BMOIT. (A) Heatmap showing overall changes in immune measurements over time on treatment, where color represents median value. (B–D) Bar plots with lines connecting each participant showing flow cytometry populations (B), scRNA-Seq populations (C), or antibody measurements (D) across time on treatment. The bars are median values. Statistical analyses were performed using paired Wilcoxon’s tests (n = 28). *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001.

CM IgE, casein IgE, β-lactoglobulin IgE, and CM IgG4 antibody levels change significantly through 12 months of BMOIT. Antibody measurements also showed significant changes with BMOIT, with a significant increase in CM IgG4 and decreases in CM IgE, casein IgE, and β-lactoglobulin (bLac) IgE after 12 months (Figure 6D) of active treatment. The CM IgG/IgE ratio had a strong increase across OIT, with a significant increase with 6 months of BMOIT and then a further significant increase at 12 months (Figure 7D). We also found further significant decreases from 12 to 24 months on active treatment in CM and casein IgE (Supplemental Figure 6). There was no significant change in α-lactalbumin (aLac).

No significant change in milk skin prick test. From baseline to month 24, there was no significant change in the milk skin prick test (SPT) within or between groups. At screening, the median milk SPT wheal size was 13.5 mm (range: 7–25) for all participants. The initial BMOIT group had a decreased milk SPT wheal size from baseline (median: 14 mm, range: 7–23) to month 24 (median: 9, range: 0–19; P = 0.08). The placebo crossover group had an increase in milk SPT wheal size, with a change from a median of 13 mm at baseline (range: 7–25) to 17.5 mm (range: 6–24) at month 24 (P = 0.34). There was no significant difference in change in milk SPT wheal size over time between groups (P = 0.1). When grouping by time on treatment, a significant decrease in milk SPT was found after 6 months (P = 0.04), but not 12 months (P = 0.2) of active BMOIT (Supplemental Figure 7).

Antibody and scRNA-Seq T cell population measurements best correlate with tolerated dose of BM. We next sought to determine whether any of our immunologic measurements correlated directly with the tolerated dose of BM from OFC. Thus, we took the BMOFC tolerated doses at months 0, 12, and 24 (shown by time relative to treatment) and correlated them to their respective T cell population percentage or antibody measurement at that time point (Figure 8, A–C). The strongest correlations were observed for scRNA-Seq T cell populations and the antibody measurements. Within these, the CM IgG/IgE ratio was the most significant antibody correlation (Spearman’s r2 = 0.5, P = 2.89 × 10–6) (Figure 8A), and the CM+FOXP3+/Th2A cell ratio was the most significant T cell measurement (Spearman’s r2 = 0.29, P = 0.01) (Figure 8B). Both measures were positively correlated with the BMOFC MTD. We additionally correlated the UMOFC MTD with the immune measurements at the challenge time point and found no statistically significant results, likely due to the single time point for the UMOFC (Supplemental Figure 8A). However, the CM+ C3 and CM+ C5 scRNA-Seq T cell populations did show trending positive and negative correlations with UMOFC, respectively (Supplemental Figure 8A). We saw no correlation of either of the OFC outcomes with baseline tolerated BM doses, age, or sex (Supplemental Figure 8B).

Correlations of immunologic measurements with all OFC data. (A–C) Scatter plots showing correlations of (A) antibody measurements, (B) scRNA-Seq populations, and (C) flow cytometry populations with baked milk oral food challenge (BMOFC) doses (mg). Color of the dot represents treatment time point. Significant correlations are noted on the plots. (D) Dot plot showing Spearman’s correlations of pretreatment T cell/antibody measurements with BMOFC (1 year of treatment), unheated milk oral food challenge (UMOFC) (24-month time point), or fold-change in BMOFC (1 year to baseline). Size of the dot represents magnitude of the correlation coefficient and color shows direction of correlation. Statistical analyses were performed using Spearman’s δ for those completing a BMOFC after at least 12 months on treatment (n = 26). Significant P values are noted.

Baseline proportion of CM+ Th2A cells negatively correlates with induction of desensitization during BMOIT. To test whether any baseline immune measurements could predict the tolerated doses of BMOFC or UMOFC after treatment, we first performed a correlation analysis. There were no significant correlations with BMOFC, but baseline CM IgG was negatively correlated with UMOFC (P = 0.05) (Figure 8D). We also performed correlations with the fold-change of BMOFC (1-year treatment vs. baseline) and discovered a significant negative correlation of CM+ Th2A cells as well as a positive correlation with the CM+FOXP3+/Th2A cell ratio (P = 0.02 and P = 0.03, respectively) (Figure 8D).

Prediction models of OFC outcomes perform best when combining T cell and antibody measurements. Finally, we used a machine-learning approach to test the ability of the immunological measurements at baseline and challenge time points to predict desensitization using T cell, antibody, and clinical data features (Figure 9A and Supplemental Table 9). We tested this using 2 different outcome measurements based on either BM or UM OFC MTD. Cutoffs for each outcome were set to split the subjects into binary “pass”/“fail” groups for each OFC (4044 mg for BM and 2000 mg for UM). For each outcome, we tested baseline and challenge T cell and antibody measurements along with clinical features (age, sex, baseline BMOFC, and time on treatment for UMOFC only). We tested a number of different models (linear and nonlinear) with a 3-fold (BMOFC) or 5-fold (UMOFC) cross-validation approach (based on outcome imbalance) repeated 10 times. We then selected the best performing model (highest average AUC) for each outcome and set of features. We found that the highest performing model was a ridge regression that used the features at the 24-month time point to predict UMOFC outcome (mean AUC: 0.806) (Figure 9B).

Combined classification of OFC outcome. (A) Box-and-whisker plots showing the area under the curve (AUC) values for each CV fold of overall best model selected (highest mean) for each set of features and outcome. Feature time point, outcome type, best model, and mean AUC across folds are indicated on the bottom. (B) Feature importance coefficients per fold for each feature in the best performing model from E (24-month unheated milk oral food challenge [UMOFC]). (C) Logistic model equation to predict UMOFC outcome. In the box-and-whisker plots, the box represents the middle 50% of data, with the bottom and top lines of the box representing the 1st and 3rd quarterlies, respectively. The line in the box represents the median. The whiskers extend to the minimum and maximum values. Outliers are indicated with dots.

We then looked at the features with the highest absolute coefficients in this top performing model for UMOFC outcome and found that a mixture of scRNA-Seq T cell (CM+ C5, CM+ C3) and antibody (aLac IgE, bLac IgE, CM+ IgG/IgE ratio) features were consistently more important in the model after cross-fold validation (Figure 9B). We then trained a simple logistic regression model on the full dataset (equation shown in Figure 9C), which summarizes how different factors can be combined to derive an overall score that better separates individuals with UMOFC success versus failure (AUC = 0.91, SD = 0.17) (Figure 9C). This equation could be used to predict how a participant would respond to UMOFC using inputted immunological measurements of CM+ C5, aLac IgE, CM+ C3, CM IgG/IgE ratio, and bLac IgE where the weight of each is indicated by the multiplied coefficient value.