Frequent longitudinal blood microsampling and proteome monitoring identify disease markers and enable timely intervention in a mouse model of type 1 diabetes

Abstract

Aims/hypothesis: Type 1 diabetes manifests after irreversible beta cell damage, highlighting the crucial need for markers of the presymptomatic phase to enable early and effective interventions. Current efforts to identify molecular markers of disease-triggering events lack resolution and convenience. Analysing frequently self-collected dried blood spots (DBS) could enable the detection of early disease-predictive markers and facilitate tailored interventions. Here, we present a novel strategy for monitoring transient molecular changes induced by environmental triggers that enable timely disease interception.

Methods: Whole blood (10 μl) was sampled regularly (every 1-5 days) from adult NOD mice infected with Coxsackievirus B3 (CVB3) or treated with vehicle alone. Blood samples (5 μl) were dried on filter discs. DBS samples were analysed by proximity extension assay. Generalised additive models were used to assess linear and non-linear relationships between protein levels and the number of days post infection (p.i.). A multi-layer perceptron (MLP) classifier was developed to predict infection status. CVB3-infected SOCS-1-transgenic (tg) mice were treated with immune- or non-immune sera on days 2 and 3 p.i., followed by monitoring of diabetes development.

Results: Frequent blood sampling and longitudinal measurement of the blood proteome revealed transient molecular changes in virus-infected animals that would have been missed with less frequent sampling. The MLP classifier predicted infection status after day 2 p.i. with over 90% accuracy. Treatment with immune sera on day 2 p.i. prevented diabetes development in all (100%) of CVB3-infected SOCS-1-tg NOD mice while five out of eight (62.5%) of the CVB3-infected controls treated with non-immune sera developed diabetes.

Conclusions/interpretation: Our study demonstrates the utility of frequently collected DBS samples to monitor dynamic proteome changes induced by an environmental trigger during the presymptomatic phase of type 1 diabetes. This approach enables disease interception and can be translated into human initiatives, offering a new method for early detection and intervention in type 1 diabetes.

Data and code availability: Additional data available at https://doi.org/10.17044/scilifelab.27368322 . Additional visualisations are presented in the Shiny app interface https://mouse-dbs-profiling.serve.scilifelab.se/ .

Keywords: Biomarkers; Coxsackievirus B; Disease intervention; Disease prediction; Disease trigger; Dried blood spots; Enterovirus; Immune-mediated diseases; Machine learning; Microsampling; Proteomics; Proximity extension assay; Screening; Type 1 diabetes.

Fig. 1

Longitudinal DBS protein profiling reveals dynamic proteome alterations in CVB3-infected NOD mice. (a) Schematic of study design. In two independent studies, NOD mice aged 8–9 weeks were infected with CVB3 (200 µl RPMI medium containing 10⁵ PFU CVB3, i.p.) or mock-infected (200 µl of RPMI medium). In the first study, the control and the CVB-infected groups consisted of four animals each and in the second study, the control group consisted of four animals and the CVB-infected group of five animals. A blood sample (5 µl) was collected from the tail vein at indicated time points and dispensed onto filter discs (Capitainer). Blood samples were eluted and proteins were measured using PEA (Olink), followed by data analysis. Created in BioRender. Byvald, F. (2024) https://BioRender.com/t83q649. (b) Heatmaps showing the mean protein levels (z score) per sampling day for each protein for the infected (left) and control (right) groups. The proteins are clustered based on the levels in the infected group. Red indicates a high z score and blue indicates a low z score. (c) Protein profiles for six proteins. Each dot represents a sample, and colour indicates if the included mouse belonged to the infected (yellow) or control (green) group. Smooth lines have been fitted to each group, with the 95% CI around the smooth line

Fig. 2

Protein dynamics across sampling time in CVB3-infected and control NOD mice. DBS samples were collected from CVB3- (n = 9) and mock-infected (n = 8) NOD mice for 14 days p.i. as described in Fig. 1a. Proteins were measured using PEA and the protein signals were transformed to z scores before combining the data from two independent studies. The figure shows volcano plots for each sampling day for the log₂ (fold change) between the infected and control groups plotted against the FDR-adjusted p value obtained from two-sided Student’s t test. Each dot represents a protein. The horizontal dotted line represents a p value of 0.05, and proteins above are considered significant. The vertical dotted line represents a fold change of zero, and proteins to the right of the line are found at higher levels in the infected group and vice versa. Yellow and green dots are proteins that are up- or downregulated, respectively, in the infected mice compared with the controls. Grey dots are proteins that were not considered significantly different between infected and control animals. FC, fold change

Fig. 3

Model performance for classifier to predict infection status. (a) ML pipeline for preliminary analysis of Study 1 samples aimed at identifying key features for infection prediction. Permutation feature importance was applied to quantify the contribution of covariates (proteins) to the model’s output. Here the model is a binary classifier that takes all 92 protein measurements at each time step as inputs for predicting infection status. Feature importance is assessed by randomly permuting each feature’s values and measuring the resulting drop in performance compared with the original unaltered input. Notably, permuting CCL2 and CXCL9 resulted in the largest accuracy drops, underscoring their significance. Their predictive relevance is visually confirmed in the plot depicting their temporal evolution. Note that the model used in the preliminary analysis differs from the one employed in the main experiments. In the ‘Fit neural network’ box, step 1 = day 1, step 2 = day 2, step 3 = day 3, step 4 = day 4, step 5 = day 5, step 6 = day 7 and step 7 = day 9. Created in BioRender. Byvald, F. (2024) https://BioRender.com/y07z021. (b, c) Results from the model used in the main analysis including data from both independent studies (Study 1 and Study 2). (b) Percentage accuracy of the model up to day 7 days p.i. in control (n = 8; green; circles) and infected (n = 9; yellow; triangles) mice. Individual animals are represented by individual symbols. (c) ROC curve illustrating the performance of the binary classification for infection status for the validation set (Study 2). Sensitivity (true positive rate) is plotted against 1-specificity (false positive rate). The AUROC was 0.995. NN, neural network; Std, standard

Fig. 4

Early intervention prevents virus-induced type 1 diabetes. SOCS-1-tg mice were infected with CVB3 (10⁵ PFU CVB3, i.p.). On days 2 and 3 p.i., animals were treated with either non-immune (n = 8) or immune sera (n = 6) by i.p. injection (total volume 200 µl/mouse). (a) Experimental schematic. Created in BioRender. Byvald, F. (2024) https://BioRender.com/b43t411. (b) Blood glucose values of individual animals treated with non-immune sera (n = 8; black line) or immune sera (n = 6; dotted blue line). Mice were deemed diabetic when the blood glucose level was equal to or exceeded 18 mmol/l, or when two consecutive daily measurements ranged between 13 and 18 mmol/l. Dotted black line marks 13 mmol/l glucose. (c) Diabetes incidence curves summarising results shown in (b). *p<0.05 comparing the two groups by logrank (Mantel–Cox) test. (d) Representative images of sequential pancreas sections from mice infected with CVB3 and treated with non-immune or immune sera. Pancreas was stained for insulin (panels on left) or glucagon (panels on right). Positive areas are stained brown. Scale bars, 100 µm. (e) Percentage of animals with damaged and intact tissue morphology in pancreas specimens from mice treated with non-immune sera (n = 7) or immune sera (n = 5). The infection status of two animals, one from each treatment group, could not be determined histologically due to insufficient quality of the formalin-fixed, paraffin-embedded sections, and were excluded from the assessment