Computational analysis identifies putative prognostic biomarkers of pathological scarring in skin wounds

We previously developed a computational kinetic model that simulated the concentration time courses for platelets, 6 cell types and 21 wound proteins (including three forms of collagen) over 40 days during an injury-initiated wound-healing response (Fig. 1) [12, 13]. A detailed list of the modeled cell types and proteins is provided in the Additional file 1. This kinetic model is system of 28 differential equations with 108 parameters. The model parameters (i.e., protein production and degradation rates, cell chemotaxis rates, etc.) were derived from in vitro cell culture data on different mammalian species including humans. In those studies, we validated the model using in vivo wound data (that had not been used in model development) from mice, pigs, rats, and humans [12, 13]. The default model parameter set corresponded to a normal wound-healing scenario (i.e., a wound that healed in a timely manner without excessive scarring). The model parameter values, parameter descriptions, model equations, and validation data are detailed in our previous work [12].

To this previously developed and validated model, we applied the newly developed computational strategies described below. In one of our analyses, we simulated a specific pathological-scarring scenario by modifying the values of two parameters in the default parameter set: we decreased by twofold the fibroblast apoptosis rate and increased by 1.5-fold the rate of collagen production by fibroblasts. We chose to modify these specific parameters because we had previously identified the biological processes represented by them as likely mechanistic factors driving pathological scarring in wounds [12].

The rates of biological processes involved in wound healing are expected to naturally vary among different mammalian species and individual animals, and also to vary due to differences in wound area, depth, and location. To reflect such natural variation in the molecular and cellular processes involved in wound healing, we simulated 120,000 distinct wound-healing scenarios, using modified model-parameter values. The corresponding 120,000 parameter sets were generated by randomly selecting the values for the 108 model parameters from intervals spanning the ranges between one-half and double the default parameter values. Our decision to sample the model parameter values from within a 0.5–2-fold range of their default values was based on the detected variability in the time courses of inflammatory/proliferative cells and cytokines in the wounds of wild-type C57BL/6J mice [12]. We carried out all computations in the software suite MATLAB R2015b (MathWorks, Natick, MA), and solved the model equations by using the MATLAB solver DDE23 with default tolerance levels.

For each of the 120,000 simulations, we calculated the fold changes in the model variables representing collagen and fibroblast concentrations with respect to their concentrations in the simulation with the default parameter values. The fold changes were calculated at the final simulation time point of 40 days after wounding, which we assumed to be representative of the time required for a HTS to develop (i.e., ~ 6 weeks [4]). From these fold changes, we classified the 120,000 simulations into three groups (Fig. 1a). Simulations for which both collagen and fibroblast fold changes did not exceed 1 were classified as “normal healing” (i.e., healing resulting in a non-pathological scar). Those for which the fold changes were simultaneously ≥ 5 and ≤ 10 were classified as “mild pathological scarring.” Finally, simulations for which the fold changes exceeded 10 were classified as “severe pathological scarring.” Our choice of the cutoff values 5 and 10 for classifying a wound-healing outcome as “mild” or “severe” pathological scarring was based on published reports of experimentally measured increases in collagen synthesis and deposition of up to fivefold in HTSs and up to 20-fold in keloids [14‐16]. If the fold changes did not satisfy the fold-change cutoffs for any of the three classification groups (i.e., fold changes > 1 and < 5), the wound-healing scenarios represented by those simulations could not be clearly classified as normal healing or as pathological scarring. The results of such simulations were excluded from the analysis of protein concentration distributions (Fig. 1a). While this step resulted in a large group of simulations being excluded from the analysis, it reduced the ambiguity in prognostic biomarker identification by considering only the simulations that were most likely to represent normal-healing or pathological-scarring wounds.

We generated protein concentration distribution histograms using the MATLAB function HIST with 50 bins partitioning the interval between the minimal and maximal concentration values for a protein in a particular group of simulations (i.e., normal healing, mild pathological scarring, or severe pathological scarring). These histograms were visualized as concentration distribution curves (Figs. 3 and 4). The percentage of the simulations for each curve was calculated by dividing the number of simulations in which a given protein’s concentration fell within the concentration range of a particular bin by the total number of simulations in that group. For each protein in our model (excluding collagen), we analysed the area of overlap between its concentration distribution generated from the normal-healing simulations and the corresponding distributions generated from the mild and moderate pathological-scarring simulations. A small overlap area indicates that the protein is consistently elevated (or decreased) in pathological-scarring simulations relative to normal-healing simulations, and was therefore more likely to be associated with a pathological outcome than proteins with larger distribution overlap areas (Fig. 1a). Concentration distributions have previously been utilized for biomarker identification, extending the prevalent practice based on fold-change analysis of gene/protein expression [17‐19]. To quantify the area of overlap between the concentration distributions, we calculated the Bhattacharyya coefficient as previously described [19]. The value of this coefficient varies between 0 and 1, which represent 0 and 100% overlap, respectively. Proteins whose concentration distributions were characterized by little overlap between the normal-healing and pathological-scarring simulations on day 40 (i.e., upon complete HTS development) were determined to be diagnostic biomarkers of pathological scarring. In contrast, proteins whose concentration distributions were characterized by little overlaps between the normal-healing and pathological-scarring simulations on days 7, 14, and 21 post-wounding were determined to be putative prognostic biomarkers. When experimental data for the model-identified diagnostic biomarkers were available in the literature [14, 20‐24], we compared the protein fold changes in pathological scars in humans (i.e., hypertrophic scars and keloids) derived from the experimental data using with the corresponding fold changes calculated from the pathological-scarring simulation described in the first subsection of “Methods” section.

We used logistic regression analysis to quantify the predictive accuracy of the prognostic biomarkers identified in our protein concentration distribution analysis (described above). For the logistic regression analysis, we used all of the 120,000 simulations and divided them into two groups: “normal healing” and “pathological scarring.” The simulations for which the fold change in the collagen level at the final simulation time point (i.e., day 40) exceeded 10 were classified as “pathological scarring,” and the remaining simulations were classified as “normal healing.” Next, the binary wound outcome (i.e., “normal” or “pathological”) of a simulation, along with the levels of the model-identified prognostic biomarkers at days 7, 14, and 21 in those simulations, were used as inputs to a logistic regression model (Fig. 2). The logistic regression model then yielded receiver operating characteristic (ROC) curves, which were used to quantify the predictive accuracy of the model-identified biomarkers (described in detail in the Additional file 1). Finally, using the logistic regression model that demonstrated the highest predictive accuracy, we performed a tenfold cross-validation analysis to estimate that model’s performance on an independent data set [25].

To identify biomarkers of HTSs, we first established criteria by which each of the wound-healing simulations could be classified as resulting in either normally healing or pathological scar formation. The pie chart in Fig. 1b shows the number of simulations falling into each group after the classification. To validate our biomarker identification methodology, we identified those proteins that were elevated (or decreased) in pathological-scarring simulations (both mild and severe) compared to normal healing simulations, and could therefore be regarded as diagnostic biomarkers for the condition. We found that six proteins—IL-10, TIMP-1, IL-6, TGF-β, chemokine CXCL8, and fibronectin—were characterized by relatively little overlap between their concentration distributions for normal-healing simulations and the corresponding distributions for mild pathological-scarring simulations and severe pathological-scarring simulations on day 40 (Fig. 3a–f).

To further assess the potential power of identifying diagnostic biomarkers using computational methods, we reviewed the existing literature to obtain the experimental measurements of TIMP-1, IL-10, fibronectin, and TGF-β in human pathological scars versus normally-healed wounds. The fold changes reported in human wounds were compared with the corresponding fold changes calculated from our simulations for normal-healing and pathological-scarring outcomes (Fig. 3g). The fold changes seen in our simulations showed agreement with the human experimental data. Differences in the experimentally obtained and model-simulated fold changes ranged from 6 to 60%), with 4 of the 6 factors having differences of less than 20% (Fig. 3g). The original data from published experimental studies using human scar tissue, including details about the protocols used for protein level measurement and the time-points at which the measurement were performed, are provided in Additional file 2.

Having identified diagnostic proteins with high levels in late stage scars, we next sought to identify proteins that may serve as prognostic biomarkers of pathological scarring. In this analysis, we looked for proteins whose concentrations at early time points in the simulations (days 7, 14, and 21 post-wounding) were highly predictive of the scarring outcome (i.e., normal or pathological scarring) at the end of those simulations (i.e., day 40). We utilized the simulation data to identify proteins whose concentration distributions showed minimal overlap between simulations that resulted in normal-healing and those resulting in mild or severe pathological-scarring (Fig. 4). The analysis demonstrated that IL-10 (Fig. 4a–c), TIMP-1 (Fig. 4d–f), and fibronectin (Fig. 4g–i) were characterized by small overlap areas between normal-healing and pathological-scarring outcomes on days 14 and 21. This indicates that, as early as days 14 and 21, levels of IL-10, TIMP-1, and fibronectin may be predictive of a pathological-scarring outcome.

At the earliest of the three examined time points post-wounding (i.e., day 7), there was considerable overlap between the concentration distributions for the normal-healing and pathological-scarring simulations for IL-10, TIMP-1, and fibronectin (Fig. 4a, d, g). Thus, none of these proteins were predictive of the scarring outcome on day 7 post-wounding. On day 14, however, the area of overlap between the concentration distributions for each of these proteins in the normal-healing and pathological-scarring simulations was ~ 14–23% smaller than on day 7 (Fig. 4b, e, h). The overlap area for each of these proteins was the smallest on day 21 (~ 30–40% smaller than the corresponding overlap areas on day 7) (Fig. 4c, f, i). This suggests that IL-10, fibronectin, or TIMP-1 levels were most likely to successfully predict a pathological scarring outcome at or later than week 3 post-wounding.

Among IL-10, TIMP-1, and fibronectin, fibronectin was characterized by the smallest concentration distribution overlap between the normal-healing and pathological-scarring simulations on both days 14 and 21. On day 14, the fibronectin distribution overlap area was 14 and 5% smaller than the corresponding areas for IL-10 and TIMP-1, respectively; on day 21, the fibronectin distribution overlap area was 8 and 1.5% smaller than the corresponding areas for IL-10 and TIMP-1, respectively (Fig. 4h, i). The corresponding overlap areas for IL-10 and TIMP-1 were comparable at both 14 and 21 days. Thus, our distribution overlap analysis showed that IL-10, TIMP-1, and fibronectin are putative prognostic biomarkers of pathological scarring in wounds.

We next sought to quantitatively assess the predictive accuracy of the IL-10, TIMP-1, and fibronectin in two instances: (1) when only one protein was used to predict a pathological-scarring outcome or (2) when two or three protein levels were used together to predict the pathological-scarring outcome. We built 14 logistic regression models that used single protein concentrations, or combinations thereof, as predictors (seven regression models for day 14 post-wounding and another seven for day 21), and derived a ROC curve for each model (Fig. 5; see Additional file 1 for further details). The regression coefficients, odds ratios, and ROC AUCs for these logistic regression models are listed in Additional file 1: Table S1. Among the six models that used the concentration of only one protein as a predictor, the model utilizing the fibronectin concentration on day 21 as a predictor demonstrated the highest predictive accuracy (ROC AUC: 0.86) (Fig. 5c). Overall, the model that used the concentrations of IL-10, TIMP-1, and fibronectin on day 21 as predictors showed the highest accuracy (ROC AUC: 0.89) (Fig. 5c).

We performed DeLong’s test to determine if any one of the ROC AUCs was significantly greater than the rest. The ROC AUCs for the seven models that used the day-14 biomarker concentrations were not significantly different from one another (Fig. 5a). This result implies that using the day-14 concentrations of any of the three model-identified biomarker proteins as predictors individually or in combination with each other could predict a pathological-scarring outcome with similar success. In contrast, among the seven models that used the day-21 biomarker concentrations, the ROC AUC for the model that used IL-10, TIMP-1, and fibronectin as predictors was significantly greater than the ROC AUCs for models that used only one protein as a predictor (Fig. 5c). However, it did not differ significantly from the ROC AUCs of the models that used two proteins as predictors (Fig. 5c). To validate our assessment of predictive accuracy by using an independent data set (i.e., data that were not used to build the logistic regression models), we performed a tenfold cross-validation for the logistic regression models that used all three biomarkers (i.e., the IL-10, TIMP-1, and fibronectin concentrations) on either day 14 or 21 as predictors. The ROC AUCs derived from this analysis for days 14 and 21 equaled 0.80 and 0.86, respectively (Fig. 5b, d). Together, the data suggest that development of highly accurate predictive biomarkers for scar formation may involve a panel of markers rather than a single factor.