Longitudinal monitoring of gene expression in ultra‐low‐volume blood samples self‐collected at home

Summary

Blood transcriptional profiles could serve as biomarkers of clinical changes in subjects at‐risk for or diagnosed with diabetes. However, transcriptional variation over time is poorly understood due to the impracticality of frequent longitudinal phlebotomy in large patient cohorts. We have developed a novel transcriptome assessment method that could be applied to fingerstick blood samples self‐collected by study volunteers. Fifteen μL of blood from a fingerstick yielded sufficient RNA to analyse > 176 transcripts by high‐throughput quantitative polymerase chain reaction (PCR). We enrolled 13 subjects with type 1 diabetes and 14 controls to perform weekly collections at home for a period of 6 months. Subjects returned an average of 24 of 26 total weekly samples, and transcript data were obtained successfully for > 99% of samples returned. A high degree of correlation between fingerstick data and data from a standard 3 mL venipuncture sample was observed. Increases in interferon‐stimulated gene expression were associated with self‐reported respiratory infections, indicating that real‐world transcriptional changes can be detected using this assay. In summary, we show that longitudinal monitoring of gene expression is feasible using ultra‐low‐volume blood samples self‐collected by study participants at home, and can be used to monitor changes in gene expression frequently over extended periods.

Introduction

Type 1 diabetes occurs in stages, beginning with the development of islet‐specific autoantibodies. This early manifestation of islet autoimmunity is followed by onset of dysglycaemia and eventually onset of symptoms, resulting in clinical diagnosis of the disease 1. Clinical consortia have been formed to follow subjects through these stages 2, 3 and intervention trials are under way for subjects in all stages of the disease. These trials and consortia collect samples from enrolled subjects as frequently as is feasible 4, balancing blood volumes that can be collected for research with burden on the subjects. Samples that can be collected at home, such as stool samples or blood spots, can be obtained relatively frequently, while samples that require clinic visits and phlebotomy are typically obtained every 3–6 months, depending on the study. Samples for whole blood or cell subset transcriptomics have typically required standard venous blood draws and thus have been collected no more than monthly. Any transcriptional perturbations associated with development of autoantibodies or onset of dysglycaemia that occur between clinic visits may be easily missed by current sampling techniques. In order to detect biomarkers that may fluctuate, and to detect them contemporaneously with any associated clinical events, more frequent sampling of subjects is required.

Systems‐scale profiling of human samples has demonstrated a potential for translation, leading to advances in the clinical management of immune‐related diseases 5, 6, 7, 8. While several transcriptome signatures of whole blood and leucocyte populations from type 1 diabetes subjects have been published 9, 10, 11, 12, 13, 14, they have not been translated into clinical use. In part, this is because transcriptional studies in type 1 diabetes have been frequently cross‐sectional in design, and thus not fully reflective of patient heterogeneity within a given disease stage. Snapshots of one or a few time‐points may not accurately capture transcriptional changes associated with pathogenesis or exposure to environmental factors. For example, diabetes subjects might have higher variability in gene expression due to highly fluctuating blood glucose levels. In such a case, larger changes in gene expression between populations and over time would be required in order to identify signal within the longitudinal noise. High‐throughput sequencing and novel quantitative PCR (qPCR) technologies have technically improved data quality in this field, addressing one obstacle to translation. However, the logistics involved in frequent blood sampling have prevented investigators from carrying out optimal studies, which would involve monitoring clinical and transcriptional changes closely over extended periods of time in large patient cohorts.

Here, we sought to solve this logistical challenge by measuring transcriptional signatures more frequently and with a lower sample volume. We developed novel sample collection and processing protocols, allowing us to implement weekly blood transcriptome monitoring in a pilot cohort of patients with type 1 diabetes and healthy control subjects. After training in a simple sample collection protocol, subjects collected and stored samples at home weekly by fingerstick for 6 months. We assessed the variance of the transcriptional signature of both long‐standing diabetes subjects and healthy controls over time. We also identified transcriptional changes contemporaneous with subject‐reported immune events.

Materials and methods

Subjects and study design

Samples for this study were obtained under informed consent from 14 control subjects and 13 participants with type 1 diabetes enrolled in the Benaroya Research Institute (BRI) Immune Mediated Disease Registry. All research protocols were approved by the BRI Institutional Review Board. Subjects visited the clinic briefly on 3 consecutive days for training in sample collection. Fingerstick samples were collected weekly at home by study participants using 15 μL Microsafe bulbed capillary tubes (Safe‐Tec, Buffalo, NY, USA), deposited into prefilled 1·5 mL tubes containing 30 μL Tempus blood RNA stabilizing solution (ThermoFisher Scientific, Waltham, MA, USA), mixed by flicking, and stored in freezer boxes at approximately −10 to −20°C in subjects' home freezers prior to transfer to BRI. The subset of diabetes subjects with detectable C‐peptide also underwent mixed‐meal tolerance testing as per standard protocols at enrolment and every 3 months thereafter 15.

Gene expression analysis

RNA was extracted using MagMax for Stabilized Blood Tubes RNA Isolation Kit (ThermoFisher Scientific) using the manufacturer's protocols, with the exception of a reduced volume of 1× phosphate‐buffered saline (PBS) (15 μL). RNA was reverse‐transcribed using the high‐capacity cDNA reverse transcription kit (ThermoFisher Scientific), followed by specific target pre‐amplification for 14 cycles in the presence of a pool of 272 primer pairs, including eight reference genes (DELTAgene Assays; Fluidigm, San Francisco, CA, USA). We also tested the Blood‐to‐Ct kit (ThermoFisher Scientific) for steps through cDNA production during the assay development phase. After pre‐amplification, samples were treated with Exonuclease I (New England Biolabs, Ipswich, MA, USA) to remove unincorporated primers. The pre‐amplified cDNAs and detection assays were loaded onto a 96.96 Dynamic Array IFC (Fluidigm). Quantitative polymerase chain reaction (PCR) was run using EvaGreen dye (Bio‐Rad, Hercules, CA, USA) for detection on a BioMark HD System (Fluidigm). Analysis was performed using the Real‐Time Analysis Software package (Fluidigm) to determine Ct values, using linear (derivative) baseline correction and autodetected, assay‐specific threshold determination. Ct detectability cut‐off was 28.

Data analysis and statistics

Ct values were first standardized to three housekeeping genes (EEF1A1, FAM105B and RPS10) that showed high correlation across samples to correct for sample input differences 16 and were then converted to fold change (FC) values compared to a commercial reference pooled cDNA sample (Biochain, Newark, CA, USA) to correct for plate batch effects. More than 65% of the genes were undetected in five of 645 samples; these samples were removed from further analysis.

FC in type 1 diabetes subjects were compared to the average FC of the healthy controls. FC of modules, comprised of four co‐ordinately expressed genes, were averaged to derive module FC values. For each individual, variance was calculated over time for each module, and then a module average variance was calculated within the healthy controls and diabetes patients to assess variance differences between these two groups; 95% confidence limits for each module were calculated based on the variance in the healthy controls during the 26 weeks. Samples that fell outside the 95% confidence intervals were classified as transcriptional outliers and were investigated for associations with contemporaneous clinical events, such as infection or blood glucose level, using logistic regression with a random effect for individual to account for the longitudinal data. Glucose values were categorized as low (< 70 mg/dL), normal (71–200 mg/dL) or high (> 200 mg/dL) for logistic regression analyses. Infection results were curated from subject‐reported notes. Infection events reported by subjects correspond to symptoms occurring on the day of sampling and/or experienced throughout the previous 6 days. As the previous sampling could be closer to the time of the event, both the week of reporting and the previous week's sample were categorized as infection events for this analysis. Multiple testing correction was performed on the regression P‐values and calculated using the Benjamini–Hochberg false discovery rate.

Results

Quality control of protocol and assay

We first identified a protocol for sample processing that yielded qPCR results for samples in the blood volume range we expected to be feasible for the clinical study (5–25 μL). A summary of methods tested in this work is presented in Table 1. We tested two methods of sample processing: a modified MagMax bead‐based method that involves RNA purification and then cDNA production, and the Blood‐to‐Ct kit, which does not require RNA clean‐up before cDNA generation. With these ultra‐low‐volume samples, we could not detect gene expression in qPCR reactions from samples generated by the Blood‐to‐Ct kit (data not shown). Samples processed using the MagMax kit showed acceptable quality qPCR data as determined by proportion of reactions yielding a Ct in the instrument's range of detection. Using the MagMax extraction protocol, we next tested minimal blood volumes that could generate reliably sufficient RNA quantity for high‐throughput qPCR. We compared 10, 15, 20 and 25 μL of blood, and found that all volumes were sufficient to generate the 100 ng of RNA needed for further processing (Fig. 1a). To determine the minimum acceptable fingerstick volume, we next compared gene expression measured using full‐sized 3 mL phlebotomy samples in Tempus solution and two potential fingerstick volumes (5 and 15 μL). More genes were undetectable by qPCR using 5 μL fingerstick samples compared to 15 μL fingersticks and standard phlebotomy samples (Supporting information, Fig. S1a), indicating that a 5 μL input sample volume was too small. To simulate what might happen under home sampling conditions, we tested two additional conditions: (1) as subjects would receive prefilled tubes with a precise quantity of Tempus solution, we looked at the impact of adding improper volumes of blood to that fixed volume of preservative, and (2) as small samples might be mixed less efficiently with home equipment, we tested the consequences of insufficient/no mixing of blood and Tempus solution. Both these conditions can yield poor RNA quality or quantity and could reasonably occur with in‐home sampling. Different types of mixing (flicking, pipetting, or vortexing the tube, or not mixing) were found to have little impact with 15 μL samples when collected with a micropipette tube. A sufficient quantity of RNA for downstream processing was generated via any mixing method. RNA quality was similar when tubes were flicked, pipetted or vortexed, but was reduced when samples were not mixed at all (Supporting information, Fig. S1b). Therefore, we instructed participants to flick their tubes for mixing, as this was performed easily in the home. Improper blood volumes, especially if too little blood was used, reduced the correlation of qPCR signatures between the fingerstick sample and a phlebotomy sample from the same time‐point (Fig. 1b). To address this, we selected capillary bulbs that had a preset volume of 15 μL and the capacity to easily eject fluid after collection.

Table 1.

Methods tested during fingerstick transcriptome protocol development

Parameter tested	Conditions	Best performer
RNA extraction kit	Magmax versus blood‐to‐Ct	MagMax
RNA quantity*	10, 15, 20, 25 μL	All volumes acceptable
Ratio of blood: Tempus solution †	1:2, 1:1·5, 1:4	1:2
Proportion of genes detected by qPCR ‡	5 μL, 15 μL	15 μL
Mixing method §	Flick, pipet, vortex, unmixed	Flick

*Data in Fig. 1a; ^†data in Fig. 1b; ^‡data in Supporting information, Fig. 1a; ^§data in Supporting information, Fig. 1b. Ct = threshold cycle; qPCR = quantitative polymerase chain reaction.

Figure 1.

Protocol development and quality control. (a) All blood volumes tested yielded sufficient RNA for cDNA preparation and subsequent quantitative polymerase chain reaction (qPCR) (> 100 ng). Duplicate samples of four different blood volumes were extracted for three subjects; samples are coloured by subject. Red, subject 1; blue, subject 2; green, subject 3. (b) Using improper blood volumes impacts qPCR results negatively. qPCR data for 176 genes from fingerstick samples with the correct ratio of blood:Tempus solution (purple) correlate more closely with data from phlebotomy Tempus tube samples (black) than from samples processed with the wrong ratio of blood:Tempus solution. Orange: too much blood and too little Tempus solution; green: too little blood per standard quantity Tempus solution; n = 4 subjects.

High subject compliance and low assay failure rates using fingerstick qPCR assay

Subjects made three visits to the clinic to learn how to perform fingersticks with these capillary tubes and how to mix and prepare their samples for freezing. At those visits, subjects received tubes prefilled with the correct quantity of Tempus solution for sample preservation, and paperwork for collection of ancillary study data. The data recorded by subjects included date and time of sample collection; reports of new medications, illnesses or vaccinations; blood glucose at the time of sampling for diabetes subjects; and reports of severe low glucose or diabetic ketoacidosis events in the previous week for diabetes subjects. Subjects performed weekly fingersticks and data collection for 6 months, and kept the samples in their home freezer for the duration of the study.

We enrolled 13 type 1 diabetes subjects of varying ages (median = 31 years) and disease durations (median = 11·6 years) as well as 14 age‐ and gender‐matched healthy subjects. The 27 subjects returned a total of 645 samples (median = 25 of 26 possible per subject, Fig. 2). Healthy and diabetes subjects had similar rates of sample return. Five of 645 (0·8%) samples failed to yield qPCR data after extraction and were excluded from further analysis.

Figure 2.

Compliance and protocol testing in type 1 diabetes (T1D) and healthy cohorts. (a) High compliance with fingerstick protocol. Samples returned by healthy (orange) and T1D (blue) subjects are depicted. Median sample count returned was 25; expected maximum was 26 samples per subject in this 6‐month study. Darker coloration denotes samples that yielded no RNA at processing (n = 5 of 645 samples received). (b) Gene expression signatures of 176 genes from 15 μL fingerstick samples are sufficiently similar to 3 mL tubes drawn by a phlebotomist. Left: correlations between systemic lupus erythematosus (SLE) and healthy or T1D samples establish minimum expected correlation of gene expression data between unrelated subjects; n = 3 SLE versus 3 healthy controls (HC) and 3 T1D for 18 total comparisons. Middle, 3 mL phlebotomy tube versus mock fingerstick: 15 μL aliquot taken from phlebotomist‐drawn tempus tube versus the same full 3 mL tube, n = 6 subjects. Shows maximum achievable correlation between small and full sample volumes. Right, phlebotomy sample versus fingerstick: correlations between fingerstick and full‐volume phlebotomy‐drawn tempus tube obtained on the same day. Shows actual correlations to gold standard tempus tube for first six subjects enrolled in this cohort; n = 6 subjects.

Samples were assessed using a custom‐designed high‐throughput qPCR assay consisting of 176 immune‐related genes. These genes were selected because they were representative of the whole transcriptome signature for many infectious and autoimmune conditions 17. For analytical purposes, the genes are divided into 44 × 4 gene units (modules), where each module represents a set of transcriptionally correlated genes; the modules are annotated when possible according to biological or cellular function of the component genes. Here, we used this assay to identify immune‐related transcriptional changes longitudinally in type 1 diabetes and healthy subjects.

As an early assessment of our subject training protocol, we measured the similarity between fingerstick samples, the standard Tempus tube (3 mL blood plus 6 mL Tempus solution) collected by a phlebotomist and a fingerstick‐sized aliquot (‘mock fingerstick’) from the Tempus tube across all 176 genes for the first six subjects to enrol. Data from these first six diabetes and healthy subject samples were compared to data from three systemic lupus erythematosus (SLE) subjects with a previously identified interferon‐associated transcript signature (Fig. 2b). This comparison established the expected concordance between biologically dissimilar signatures. Intrasubject similarities exceeded between‐group differences, as expected. The SLE subjects versus healthy or type 1 diabetes subjects showed a mean concordance of 0·817 (range = 0·754–0·862). The maximum possible gene signature concordance for this assay was established as a two‐way comparison of signatures generated from the mock fingerstick and the remaining 8·9 mL of the same Tempus tube (average: 0·980; range: 0·975–0·988). We then compared a fingerstick sample and the 9 mL Tempus tube drawn on the same day for the same six subjects. This comparison yielded concordances between these minimum and maximum values, as expected (average: 0·957; range: 0·908–0·986). Fingerstick versus tempus concordances remained similar throughout the study for the subset of subjects where these samples were collected (Supporting information, Fig. S1c). These data showed the scale of variability that can be detected with this assay in our subject types, and established the range of correlations to the gold standard that can be achieved with this collection method.

Transcriptional profiles show similarity in longitudinal profiles across subjects, with higher total variance in healthy subjects

We next sought to determine whether any transcriptional differences existed between healthy and type 1 diabetes subjects using this assay. Using two‐group comparison, no genes or modules of genes were identified to be significantly different in expression level between diabetes and healthy subjects, and their global expression profiles largely overlapped. This was not unexpected, as major transcriptional differences between healthy and long‐standing diabetes subjects have not been consistently identified in previous studies, and the gene panel was not selected to specifically differentiate these groups. We expected to see transcriptional similarity over time in samples from the same subject. Using the mean correlation of all genes in the panel, longitudinal samples from individual subjects were more strongly correlated with each other than to the mean of all samples from other subjects (Fig. 3a).

Figure 3.

Variability within any given subject is similar for type 1 diabetes (T1D) and healthy individuals, but variance across subjects is higher for healthy subjects. (a) Quantitative polymerase chain reaction (qPCR) data from longitudinal samples for all subjects correlate more strongly with each other (black dots) than with all other subjects (grey dots). Individual subject IDs listed on the x‐axis. Black dots: mean correlation of time‐points across all genes within an individual. Grey dots: mean correlation of longitudinal time‐points across all genes between individuals. For example, the grey point for subject 1001 is the correlation of data from all visits for subject 1001 to data from all visits for all other individuals; the black dot is the correlation of gene expression between all longitudinal samples from subject 1001 only. (b) As a group, transcriptional signatures from healthy subjects were more variable than those from T1D subjects. Mean variance for all subjects in a group was calculated for each module of four genes. Each point represents 1 module. In 39 of 44 modules, healthy control (HC) individuals show higher variance than T1D individuals. Hashed line indicates equivalence (equal variance in both subject types).

We next looked at longitudinal expression of the full panel of genes by subject type, and calculated a summary value for each module representing week‐to‐week variance during the course of the study. We expected that module level variance in diabetes subjects would be higher than in healthy subjects due to the groups’ natural differences in metabolic control and thus differences in glucose variability. Contrary to our initial expectations, healthy subjects showed significantly higher variance than type 1 diabetes subjects in 39 of 44 modules (Fig. 3b).

Patient‐reported infections are associated with increased interferon‐stimulated gene expression

A primary goal for assay development was to identify transcriptional changes that occurred concurrently with change in a measurable clinical variable (blood glucose at draw) or with patient‐reported data (recent infection). We first identified outlier values in our transcriptional data, and then determined whether each transcriptional outlier occurred contemporaneously with high or low blood glucose or with subject‐reported infectious symptoms.

Using logistic regression, we found no significant associations between either high or low blood glucose at draw and transcriptional outliers for any module. However, we identified a significant association between transcriptional outliers and patient‐reported infection for three modules, which all included interferon‐inducible transcripts (Table 2). The data are displayed longitudinally in Fig. 4, with each subject who experienced infection represented as one line. Transcriptional outliers are all points above or below the 95% confidence interval limits (hashed lines). All infection events are noted with filled circles. Overlaps between filled circles and outlier values show time‐points when infection and transcriptional differences were detected simultaneously (see Table 2 for numerical frequencies and statistics and Supporting information, Fig. S2 for data on all modules in the study). Outlier transcriptional status in the three interferon modules is statistically associated with infection status. These data show that our assay is capable of detecting longitudinal transcriptional fluctuations that are associated with concurrent immune events.

Table 2.

Patient‐reported infection events correspond with transcriptional changes

Module number	Adjusted P‐value*	Transcriptional outliers (n)	Patient‐reported infections (n)	Infection and outlier overlap (n)
Interferon 8·3	1·67E‐5	25	22	10
Interferon 15.127	0·028	32	22	5
Interferon 15·86	0·041	25	22	6

*P‐value after multiple testing correction.

Figure 4.

Subjects with infections tend to have concurrent transcriptional increase of interferon‐stimulated genes. Each panel depicts one transcriptional module of four interferon‐stimulated genes. Log fold change [log₂(FC)] between the subject and the mean FC of healthy subjects for that module of four genes is on the y‐axis; the x‐axis is time in weeks. Orange lines indicate healthy subjects who reported infection at any point in the study; blue lines indicate type 1 diabetes (T1D) subjects who reported infection. Closed circles denote samples from subject‐reported infection events. Hashed lines designate 95% confidence interval limits for each module; points outside hashed lines are transcriptional outliers. Association between transcriptional outliers and infections (dots) was used to assess significance. In the logistic regression model, patient was a random effect and infection was an independent variable. P‐values have been adjusted for multiple testing.

Discussion

Weekly sampling of subjects for transcriptional research in diabetes research has, historically, not been feasible because it has required subjects to visit clinics for blood draws. More frequent sampling is desirable, however, so that immune fluctuations can be detected at any time in a subject's transition through the clinical stages of type 1 diabetes. Our sample collection and processing protocol allowed subjects to collect samples at home by fingerstick. A previous publication 18 identified high correlations by microarray between fingerstick and phlebotomy samples with a larger fingerstick volume (70 μL) and slightly different RNA stabilization solution. Here, we tested lower sample volumes and used Tempus solution for RNA sample stabilization. From these samples, we collected data on a set of 640 longitudinal samples from 27 subjects. We used a transcript profiling panel implemented for generalized measurements of immune perturbations, but we note that any targeted panel of genes could be measured using this same technology. We chose to measure 176 genes, but sufficient material was generated to measure up to 272 genes. Additionally, the quantity of RNA generated with each fingerstick was sufficient for production of an RNA sequencing library suitable for full transcriptome profiling, if desired.

Our patient population included only adult type 1 diabetes subjects with relatively long‐standing disease and matched adult controls. This population was not selected to provide insight into autoantibody development or clinical onset of disease; rather, it was selected to simply show proof‐of‐principle for the collection and analysis protocol. Similarly, the genes measured in this study were not selected for association with diabetes pathogenesis. A study with a custom gene panel predicted to change longitudinally upon environmental or infectious exposure, or that contains genes expected to change in the context of metabolic differences, would be of value. Utilizing such a panel in the earlier stages of type 1 diabetes could prove fruitful in understanding clinical changes during disease. The use of custom gene panels selected to monitor specific therapeutic intervention could also have utility in clinical trial settings. Additionally, the small blood volume required would facilitate sampling of younger children via heelstick, allowing researchers to understand disease dynamics more clearly in the very young, when autoantibody development peaks 19, 20, 21, 22.

We identified increased variance in gene expression for nearly all transcriptional modules in healthy subjects compared to those with type 1 diabetes. One explanation for this finding could be that diabetes subjects are more accustomed to home medical testing in general, and fingersticks in particular, and could have been more consistent in sampling methods during the course of the study. Another possibility is that our adult diabetes subjects were somewhat more consistent in diet or other lifestyle habits compared to controls, resulting in less variability. A third distinct possibility is that relatively few of the modules in this study are impacted by metabolic control.

We note that variability was higher in control subjects for the majority of transcripts analysed, but that variability was also dramatically higher in some modules regardless of subject disease status. The most highly variable genes/modules represent examples where cross‐sectional data in diabetes or healthy subjects should be interpreted with care. When longitudinal within‐subject variability is high, one could identify cross‐sectional differences randomly between groups that are due solely to expected temporal fluctuations in the data and not due to the group difference being investigated. This is particularly true in studies with small sample sizes, as are commonly performed for relatively expensive technologies. Better understanding of longitudinal variability in this assay and others could help to appropriately interpret cross‐sectional studies.

In summary, this study demonstrates the use of a novel protocol for frequent longitudinal transcriptional profiling. It could be applied to any disease of interest, to any set of transcripts detectable in blood, and provides a new tool for investigating longitudinal transcriptome changes.

Disclosure

Authors have no conflicts of interest to disclose.

Author contributions

C. J. G., D. C. and J. M. O. designed the study. C. S. and V. H. G. performed experiments. C. S., E. W. and J. M. O. analysed the data. C. S. and E. W. wrote the manuscript. All authors approved the manuscript.

Supporting information

Additional Supporting Information may be found in the online version of this article at the publisher's website.

Fig. S1. Protocol development and quality control. (a) A larger number of the 176 genes tested were not detectable in 5 μL blood samples than in 15 μL blood samples; 3 mL is the volume of a gold standard phlebotomy‐drawn tempus tube and was included for comparison, n = subjects. (b) All mixing methods tested yielded sufficient RNA for cDNA preparation and subsequent quantitative polymerase chain reaction (qPCR) (∼100 ng). However, RNA quality (RIN) was lower for unmixed samples. Vortexing and pipetting to mix were tested in the laboratory, but are not feasible in the home setting; therefore flicking was selected for this study. Duplicate samples tested from three unique subjects; samples are coloured by subject. Red, subject 1; blue, subject 2; green, subject 3. (c) Correlations between gene signatures from fingersticks and phlebotomy‐collected tempus tubes did not change substantially during the course of the study. A subset of type 1 diabetes (T1D) subjects underwent mixed meal tolerance tests every 3 months, and a tempus tube was collected at those visits. Figure shows correlation between fingerstick samples and tempus tube samples drawn on the same day for these six subjects (subjects do not overlap with Fig. 2b). Some fingerstick samples were not returned by subjects, and are thus missing from (c).

Fig. S2. Longitudinal gene expression data for all subjects, all modules. Each panel depicts one transcriptional module of four genes. Log fold change [log₂ (FC)] between the subject and the mean FC of healthy subjects for that module of four genes is on the y‐axis; x‐axis is time, in weeks. Orange lines indicate healthy subjects, blue lines indicate type 1 diabetes (T1D) subjects. Each figure includes all subjects (Fig. 4 in the main document includes only subjects who reported infection for visual clarity). Closed circles denote samples from subject‐reported infection events. Hashed lines designate 95% confidence interval limits for each module; points outside hashed lines are transcriptional outliers for that module. Note differing axes which highlight differing variances. Association between transcriptional outliers and infections (dots) was used to assess significance. In the logistic regression model, patient was a random effect and infection was an independent variable. P‐values have been adjusted for multiple testing. Longitudinal gene expression data for all subjects, all modules.