Refined measurement of immunological balance
Previously, we conducted cross-sectional plasma-induced transcriptional analyses of individuals recently diagnosed with type 1 diabetes, unrelated healthy control individuals lacking family history of autoimmunity and healthy autoantibody-negative siblings of individuals with type 1 diabetes possessing high-risk (DR3 and/or DR4) or low-risk (non-DR3/non-DR4) HLA haplotypes [22]. Among these cohorts we identified 1374 differentially induced probe sets. To quantitatively measure immune activity, we developed a gene-ontology-based composite inflammatory index (I.I.com), calculated by determining the ratio of the mean intensity of the induced inflammatory genes to the mean intensity of the induced regulatory genes [22]. High scores reflect greater inflammatory bias and low scores reflect greater regulatory bias.
Here, in an effort to simplify the I.I.com, we used random forest analysis to identify 359 probe sets that were optimal classifiers of the diabetic cohort relative to the three control cohorts. Among these 359 probe sets, 325 (90.5%) overlapped with the 1374 probe sets identified previously [22]. Similar to the original data, this subset captured increasing IL-10/TGF-β bias and decreasing IL-1/NFκB bias across the continuum created by the siblings with low HLA risk, individuals with type 1 diabetes, siblings with high HLA risk and unrelated control individuals (Fig. 1a). This is reflected by the increased induction of chemokine transcripts by plasma of siblings with low HLA risk and individuals with diabetes, and elevated induction of regulatory transcripts (IL2RA, CBLB, SMURF1, SMURF2, SKI) by plasma of siblings with high HLA risk and unrelated control individuals. As described for I.I.com [22], an inflammatory index based on these 359 probe sets (I.I.359) was calculated by dividing the average signal intensity of the 103 probe sets generally annotated as ‘inflammatory’ by the average signal intensity of the 256 probe sets generally annotated as ‘regulatory’. Similar to I.I.com, the average I.I.359 for the individuals with diabetes was significantly higher than the other cohorts (Fig. 1b, c). The receiver operator characteristic (ROC) analysis of this subset showed improved discrimination of the diabetic and control cohorts (Fig. 1d).
The relationship between baseline I.I.
359 and beta cell function
Next, we examined the relationship between the baseline inflammatory/regulatory balance (I.I.359) and the decline in beta cell function among TN-09 participants. We hypothesised that baseline I.I.359 might relate to disease activity and correlate with stimulated C-peptide AUC in untreated (placebo) participants. We further hypothesised that the therapeutic effect of CTLA4-Ig would alter the natural disease course, abolishing the relationship established in the placebo arm. A significant inverse relationship was observed between baseline I.I.359 and the per cent change from baseline C-peptide AUC at 3, 6, 12, 18 and 24 months in placebo-treated participants (p ≤ 0.015); this relationship was not observed among CTLA4-Ig-treated participants, consistent with immunomodulation altering disease progression (Fig. 2a–e). Similarly, baseline I.I.359 was inversely related to the rate (slope) of C-peptide decline over the 24 month study in the placebo arm but not in the CTLA4-Ig arm (Fig. 2f).
To determine whether baseline I.I.359 is predictive of the duration of the post-onset partial remission period, we investigated the relationship between I.I.359 and the duration of MMTT-stimulated C-peptide ≥0.2 nmol/l in the TN-09 placebo arm participants (Fig. 3a) and the duration of IDAA1c ≤9 in local individuals with newly diagnosed diabetes (Fig. 3b). In both cohorts, the baseline I.I.359 was inversely related to the duration of the post-onset partial remission period (p ≤ 0.012). Further, participants with I.I.359 above the median had a significantly shorter duration of persistent insulin secretion than those below the median (Fig. 3c, d), supporting the hypothesis that individuals with higher inflammation at onset will experience accelerated decline in beta cell function. Supporting this observation was the measurement of lower abundances of peripheral activated Tregs (CD4+/CD45RA−/FOXP3high) during the immediate post-onset period in the local participants that exhibited an IDAA1c ≤9 for less than 6 months (Fig. 3e, f; p = 0.016).
The baseline C-peptide AUC of the TN-09 participants was directly related to age at diagnosis, supporting a recognised relationship [31, 32]. However, baseline I.I.359 was independent of age in the TN-09 and local cohorts (Fig. 4), indicating that the more rapid decline of beta cell function in those individuals with higher inflammatory bias at baseline was independent of the age of clinical onset. Further, baseline I.I.359 was independent of time from clinical onset in both cohorts and, in TN-09 participants, there was no correlation between diabetes duration and the baseline stimulated C-peptide AUC.
Stratification of individuals fosters identification of a CTLA4-Ig therapeutic response signature
Our initial strategy for defining immunomodulation achieved by CTLA4-Ig in TN-09 followed that described in our analyses of the IL-1 antagonism trials [24]. Briefly, induced transcription for each participant at 3, 12 and 24 months was normalised with that of baseline, then differentially induced transcripts between the CTLA4-Ig and placebo arms were identified. A total of 427 differentially induced probe sets (log2 ratio >|0.263|, 1.2-fold; ANOVA p < 0.05) were identified between the trial arms at ≥1 time point (ESM Fig. 1). Unacceptably, no transcript exhibited an FDR <50% at any time point. We hypothesised that the analysis could be improved by matching treated and placebo participants for baseline I.I.359 and respectively focusing on those participants with the greatest therapeutic response or the most rapid disease progression.
To identify CTLA4-Ig-treated participants with the greatest therapeutic response, we applied two criteria. The first criterion utilised the placebo arm trend line generated when baseline I.I.359 was plotted against per cent change from baseline stimulated C-peptide AUC at 3, 6, 12, 18 and 24 months (Fig. 2a–e). Participants who were >1.5 SD above this regression line at ≥3 time points were identified. The second criterion utilised the placebo arm regression line when baseline I.I.359 was plotted against the rate of C-peptide decline over the 24 month study period (Fig. 2f). Participants who were >1 SD above the trend line were identified. Eight CTLA4-Ig treated participants met both criteria. These eight participants were then compared with placebo participants (n = 7) matched for I.I.359 and residing at or below the placebo regression line depicted in Fig. 2f.
As previously reported, the decrease in beta cell function among CTLA4-Ig-treated participants paralleled that observed in the placebo arm after 6 months [25]. Therefore, we hypothesised that the maximal plasma-induced transcriptional signature representing therapeutic response would be detected at 3 months post-enrolment. In this way, we identified 1509 differentially induced probe sets between the selected eight CTLA4-Ig- and seven placebo-treated participants (mean log2 ratio >|0.263|, 1.2-fold; ANOVA p < 0.02; FDR ≤30%). These data did not significantly overlap with the cross-sectional data set. Specifically, 50/1374 (3.6%, χ2 > 0.059) and 10/359 (2.8%, χ2 = 1) transcripts overlapped those used to define I.I.com and I.I.359, respectively.
On average, the signatures of the remaining placebo- and CTLA4-Ig-treated participants were intermediate to those of the selected placebo- and CTLA4-Ig-treated participants and not distinct from one another (Fig. 5a, left). As expected, hierarchical clustering of individual selected placebo- and CTLA4-Ig-treated participants using the 1509 probe sets resulted in distinct grouping (Fig. 5a, middle). In contrast, the remaining participants exhibited imperfect clustering, suggesting that this subgroup possessed both slow-progressing and treatment-non-responder participants (Fig. 5a, right). Consistent with the resumed decline in stimulated C-peptide, at 12 months and 24 months, the plasma milieus of the selected CTLA4-Ig- and placebo-treated participants were more similar, with 236 and 0 probe sets being differentially induced to the aforementioned thresholds, respectively (Fig. 5b).
IPA was then used to identify candidate regulators of the 1509 probe sets differentially induced between the selected CTLA4-Ig- and placebo-treated participants (Fig. 5c). Consistent with immunomodulation anticipated by CTLA4-Ig therapy, IPA identified as being significantly activated (z score >2.0): TGF-β1; histone deacetylase co-repressor 2 (HDAC2), a transcriptional repressor that governs NFκB-regulated genes [33]; aryl hydrocarbon receptor (AHR), an important modulator of adaptive responses [34]; and the transcriptional regulator sterol regulatory element-binding protein 1 (SREBF1) [35]. Further, IPA revealed significant inhibition of CD28 and CD3, molecules important in T cell activation, and lysine demethylase 5A (KDM5A), important in natural killer cell activation [36], (z score <−2.0) in CTLA4-Ig treated individuals. Transcripts annotated under these candidate upstream regulators are known to possess roles in attenuating adaptive immune responses (SKIL, SMAD2, PTPN22, AHR), mediating cell adhesion (CEACAM5, L1CAM) and regulating proliferation/apoptosis (SDC4, BAX, NDFIP2, BTG1).
The half-life of CTLA4-Ig is 14 days. At the 3 month sampling, the TN-09 dosing schedule was such that participants had last received an infusion 4 weeks prior; as such, it was considered unlikely there would be residual CTLA4-Ig in the samples. We reasoned that if the treatment arm signature was a direct consequence of carry-over CTLA4-Ig in the plasma, it would be possible to recapitulate that signature in samples of untreated individuals by spiking the cultures with CTLA4-Ig. Plasma of untreated individuals with diabetes was therefore supplemented with 0 μg/ml, 25 μg/ml (estimated steady-state trough level) or 82 μg/ml CTLA4-Ig (estimated steady-state high level). Even at the highest level, only 880 (58.3%) of the 1509 probe sets differentially induced between the selected CTLA4-Ig- and placebo-treated individuals were directionally concordant, and only 67 (4.4%) were directionally concordant and possessed an FDR <30% (Fig. 5d). This indicates that signatures of the participants in the treatment arm were largely independent of carry-over CTLA4-Ig and that signatures of the selected CTLA4-Ig-treated participants were reflective of treatment-mediated immunomodulation.
Variation in baseline signatures defines distinct subgroups that differentially respond to CTLA4-Ig
The baseline I.I.359 was significantly related to post-onset C-peptide AUC and was useful for matching participants in analyses aimed at defining the therapeutic effects of CTLA4-Ig. However, I.I.359, which is derived from cross-sectional analyses, did not completely differentiate CTLA4-Ig treatment responders from non-responders, suggesting that I.I.359 does not entirely capture the variation among newly diagnosed individuals. Therefore, we identified the 3159 most variant transcripts common to the local and TN-09 participants at baseline (Fig. 6a). This dataset was analysed with WGCNA, a software that clusters the main patterns of variation into modules of co-expressed transcripts and correlates these to phenotypes.
Among the 3159 transcripts, WGCNA identified 12 modules. Two modules significantly correlated with the baseline C-peptide AUC among TN-09 participants (black and yellow); a third module (blue) correlated with the rate of C-peptide decline among both placebo and CTLA4-Ig treated participants (Fig. 6b, c). These modules also correlated with I.I.359, suggesting that the inflammatory/regulatory dynamic previously identified by cross-sectional analysis is also present among newly diagnosed individuals.
To investigate whether the immune activity represented by these three modules could define subgroups among newly diagnosed individuals, the baseline signatures of TN-09 participants were subjected to hierarchical clustering. This identified four major subgroups that did not significantly differ by age or baseline C-peptide AUC (Fig. 6d). For the 916 transcripts encompassed by these modules, IPA identified numerous candidate upstream regulators, including lipopolysaccharide (1.2 × 10−31), IL-1B (1.7 × 10−18), TNF (2.4 × 10−29), IFN-γ (7.7 × 10−19), IL-10 (7.7 × 10−13) and TGF-β1 (1.8 × 10−8). The yellow and black modules included numerous proinflammatory annotations, including cytokines (IL1A, IL1B, IL6) and chemokines (CCL2, CCL3, CCL4); conversely; the blue module included many transcripts related to IL-10 and TGF-β signalling (SKI, SKIL, INPP5D, SUZ12) (Fig. 7a).
The CTLA4-Ig and placebo arms of TN-09 were independently analysed from the perspective of these four subgroups. While the limited number of participants assigned to these subgroups precludes robust conclusions, significant age-independent differences were identified in the rate of C-peptide decline within each treatment arm. Among CTLA4-Ig-treated participants, those in subgroup 1 exhibited the smallest per cent change from baseline C-peptide AUC at 3, 12, 18 and 24 months and the lowest overall rate of C-peptide decline over 24 months compared with those in the other three subgroups (Fig. 7b). The CTLA4-Ig-treated participants assigned to subgroup 2 were significantly younger than their placebo-treated counterparts (Kolmogorov–Smirnov test, p < 0.04), and exhibited a faster rate of C-peptide decline and a greater per cent change from baseline C-peptide AUC at 3, 6, 12, 18 and 24 months than the other subgroups. Among the placebo-treated participants, the rate of C-peptide decline and per cent change from baseline C-peptide AUC for subgroups 3 and 4 were significantly greater than those of subgroup 2 (and greater than the single placebo-treated individual assigned to subgroup 1). Finally, within subgroup 3, CTLA4-Ig treatment significantly reduced the per cent change from baseline C-peptide AUC at 6, 18 and 24 months when compared with the placebo arm (Fig. 7b).
These data suggest that baseline plasma-induced signatures can define type 1 diabetes subgroups. Participants in subgroups 1 and 2 exhibited greater baseline regulatory bias (and lower I.I.359) and a generally slower rate of C-peptide decline. Conversely, participants in subgroups 3 and 4 exhibited greater innate inflammatory bias and a faster rate of C-peptide decline. Subgroup 3 contained most of the selected CTLA4-Ig responders and rapidly progressing individuals who received placebo. The data also suggest that participants with greater baseline inflammatory activity (subgroups 3 and 4) exhibited a better therapeutic response to CTLA4-Ig during the 24 month study period (Fig. 8).