Lung transplantation (LTx) is an acceptable therapy for end-stage lung disease. However, successful post-transplantation outcomes are significantly impacted by the development of severe primary graft dysfunction (PGD), which is associated with early mortality and a more rapid onset of chronic lung allograft dysfunction (CLAD) [1–4]. As such, there is an urgent need to better understand PGD pathobiology and identify those at enhanced risk. The pathophysiology of PGD is complex and involves a multifactorial inflammatory process that likely starts in the donor lung and is exacerbated by ischemia reperfusion injury (IRI). Given the contribution of inflammation in PGD, several studies have leveraged the presence of pro-inflammatory cytokines as prognostic biomarkers. These studies have largely measured the presence of soluble factors in recipient plasma, with elevated levels of plasma cytokines such as IL-6, IL-8, IP-10, MCP-1, and endothelial/epithelial injury markers shown to be associated with severe PGD [5–13].

Though plasma provides a valuable window into the biological/inflammatory processes detected on a systemic level, a handful of studies have analyzed bronchoalveolar lavage (BAL) fluid as a more representative sample of the transplanted lung-specific microenvironment. Elevated BAL levels of IL-6, IL-8, and Eotaxin were shown to be associated with PGD [14–16]. Further, ex vivo lung perfusion (EVLP) studies have shown that elevated IL-8, IL-6, and GRO-α levels in perfusion solution are associated with poorer graft outcomes [17, 18].

Together, these studies show the potential value of measuring inflammatory and injury markers as predictors of PGD. Most factors identified to date have been selected for investigation by association, in that they have been previously implicated in other acute lung injuries and then subsequently investigated in LTx. While these studies have guided our understanding of PGD, their hypothesis-driven focus may limit the discovery of other mediators or predictors of PGD. Thus, here we utilized a 71-analyte multiplex approach composed of a broad array of cytokine, chemokines, and growth factors in concert with matched donor perfusate, recipient plasma, and BAL to identify novel biomarkers and/or mechanistic factors associated with PGD. By examining multiple analytes across diverse samples, our study aimed to mitigate selection bias and uncover new correlates to PGD. We further hypothesized that a machine learning approach may serve as a valuable tool that complements classical statistics for biomarker discovery. Here we demonstrate that cytokine expression in BAL collected 2h post-LTx best predicts PGD development. Further, using a model reduction process, we identified the three strongest BAL predictors of PGD development were BCA-1 (CXCL13), Fractalkine, and IL-1Ra, factors that have not been extensively studied in the context of PGD. Together, our results provide a methodological foundation for further development of predictive modeling that could result in the discovery of analytes that can be used in a clinically relevant point-of-care test and provide novel targets for PGD mechanistic studies.

Here we sought to identify donor and recipient inflammatory risk factors for the development of PGD to characterize its pathophysiology using a “hypothesis-free” approach. Our 71-analyte multiplex panel allowed us to examine previously reported markers of PGD while achieving unbiased discovery of new potential mechanistic/therapeutic targets and/or novel biomarkers. We sampled multiple biological sites from each recipient and obtained results using the same assay and data analysis methodology, allowing us to compare the utility of each sample site against each other without bias from batch variability. It is important to distinguish the two analytical aims of this study. First, we performed associative analyses across plasma, perfusate, and BAL to identify cytokines linked to PGD status or peri-transplant temporal changes. Second, we used predictive modeling to compare the relative utility of different specimen types for discrimination of severe PGD and to prioritize a reduced set of candidate BAL analytes. Thus, the longitudinal plasma analyses were intended primarily to define peri-transplant inflammatory patterns and persistence of candidate signals over time, whereas predictive modeling focused on early specimen types most relevant to acute PGD classification. Our analysis demonstrated the presence of previously undescribed factors associated with PGD in these different biological samples and determined that T0 BAL had the strongest PGD predictive value.

To our knowledge, no previous human studies have investigated cytokine levels in static donor lung preservation solution. In keeping with a recent heart transplant multiplex study of heart preservation solutions [38], we identified a large array of detectable factors in lung perfusate, as 66 of the 71 cytokines were detected. Of the 66 identified cytokines, only IL-17a was detectably different between PGD groups. Our plasma studies provide insight into cytokine level changes across the transplant process and determine whether signatures of cytokines are associated with severe PGD. Fifteen analytes were detectably different in PGD+ vs. PGD-plasma at individual time points. Amongst these were IL-6 and MCP-1, which have been previously reported to be elevated in the plasma of PGD patients [5, 8–11]. While 15 analytes were detectably changed, only two, BCA-1 and IL-22, were detectably elevated in PGD at all timepoints, using multivariate approaches. To our knowledge, these cytokines have not been reported as plasma PGD biomarkers thus far and may indicate that PGD patients have increased early B cell signaling [39, 40]. Finally, we utilized T0 BAL as a window into the local lung microenvironment. A large array of analytes was present in BAL. Of the 71 analytes assessed, we demonstrated that eight (IL-1RA, Fractalkine, BCA-1, Eotaxin, IL-12p40, IL-6, IP-10, and M-CSF) were detectably elevated in PGD.

Using classical statistics, we identified novel analytes that have not been previously described or linked mechanistically in PGD. A secondary goal of our study was to explore whether our multiplex approach could optimize the identification of biomarkers that predict severe PGD. While most biomarker discovery studies use only statistical association tests or univariable prediction models to identify candidates, a growing number use multivariable models, often internally cross-validated [41]. For example, Wolf et al [42] and Berra et al [43] used proteins measured in BAL by mass spectrometry or Western blot, in concert with FEV, to build machine learning models that successfully stratified the risk of CLAD development. To our knowledge, no multivariable studies have utilized patient-matched biospecimens from different sites to test multiplex data and sample types. Here we show that our multivariate predictive modeling comparisons indicated the optimum sample to analyze for its capacity to predict PGD onset, in our sample types, was BAL collected 2hrs post-transplant (T0). Using this exploratory reduction strategy, we prioritized IL-1Ra, CXCL13/BCA-1, and Fractalkine as candidate BAL analytes associated with severe PGD. Using a LASSO analysis to assess the predictive power of analytes in T0 BAL specimen we identified three cytokines, IL-1Ra, BCA-1, and Fractalkine, as strong candidates that accurately predict PGD while minimizing model deviance. Except for IL-1Ra, these cytokines have not been previously correlated to PGD development. Interestingly, BCA-1 levels were also detectably elevated in plasma, both pre- and post-transplantation. Given the mechanism of action of BCA-1 [44, 45], our findings are interesting in light of recent studies implicating B cells in lung transplant IRI [46–48]. While not investigated in lung IRI or lung inflammation, Fractalkine has been shown to play important roles in the context of cardiac IRI [49–51] and inflammation [52, 53]. Taken together, our findings support the need for future functional studies investigating the roles of IL-1Ra, BCA-1, and Fractalkine to fully evaluate the significance of these mediators in PGD.

Recent studies have utilized cytokine rapid testing to identify and stratify patients at risk of hospitalization upon presenting with respiratory distress [54]. Given the early post-transplant signal of IL-1Ra, BCA-1, and Fractalkine, observed in BAL, these findings raise the possibility that a focused cytokine panel could ultimately contribute to early PGD risk stratification. Our predictive modeling passed multiple data robustness checks. However, the present study is not sufficient to establish clinical utility, and any point-of-care application will require external validation, calibration testing, and assessment in broader cohorts that include intermediate and fluctuating PGD phenotypes. As such, these markers should not yet be considered a validated predictive panel but rather a biologically plausible and clinically testable candidate signature for follow-up studies.

Previous multiplex cytokine studies in cardiac transplantation showed similar large-scale cytokine changes at different time points post-transplant and note, as we have, new target cytokines [38]. A major difference in our study, beyond its lung focus, was the inclusion of patients with different early graft outcomes. By incorporating PGD+ and PGD-patients, we were able to not only map cytokine changes across sample type and time but further identified novel analytes that could have mechanistic and predictive value for the discrimination of PGD. To promote dichotomization in our machine learning models, we included patients that were either PGD 0 or 3 at all timepoints (24, 48, and 72 h). We acknowledge the accepted PGD construct is to collapse the 0–72 h grades 1–3 into PGD3 at 48 or 72h vs. no PGD3 at 48 or 72 h¹. However, we believe our approach optimized its discovery and predictive ability given the relatively small sample size. Nevertheless, future studies are needed to validate the analytes we discovered using a larger sample cohort and accepted PGD diagnostic construct.

This study has several important limitations. Most notably, the cohort size was small relative to the dimensionality of the multiplex and modeling analyses, increasing the risk of model instability and overfitting. Accordingly, the very high predictive performance observed in some combined-feature models, including near-perfect AUROC values, should not be interpreted as evidence of a clinically ready classifier but rather as support for this framework as an exploratory approach to prioritize candidate biomarkers and informative specimen types. We focused our studies on the extremes of PGD, with our two groups split between PGD0 and PGD3 at all timepoints. We recognize that this approach does not capture the full clinical spectrum of PGD nor establish how identified markers would perform in intermediate phenotypes. Evaluation of PGD grades 1–2 and temporally dynamic trajectories will require substantially larger cohorts and will be the focus of future validation studies.

Baseline differences between PGD groups may also have influenced the observed biological signals. Although available clinical predictors were evaluated individually and in combination with biological data, and no single measured clinical variable performed comparably to the BAL cytokine signal, residual confounding cannot be excluded. In addition, several relevant donor and preservation variables, including ischemic time, donor PaO2/FiO2, ventilator settings, and flush volume, were unavailable in the de-identified tissue bank dataset, limiting our ability to assess their contribution to group differences.

Taken together, these considerations emphasize that the identified BAL analytes, including IL-1Ra, CXCL13/BCA-1, and Fractalkine, should be regarded as hypothesis-generating candidates. Validation in larger, independent, prospectively collected cohorts with more complete donor and perioperative annotation will be required.

In conclusion, we used matched perfusate, plasma, and BAL biospecimens to define inflammatory signatures associated with severe PGD after LTx. Across specimen types, early BAL demonstrated the clearest relationship with PGD status and identified BCA-1, Fractalkine, and IL-1Ra as candidate analytes for further study, supporting their potential relevance to PGD pathophysiology and the need for future mechanistic investigation. These findings should be interpreted cautiously given the limited cohort size, incomplete availability of certain donor and preservation variables, and the exploratory nature of the analyses. Nonetheless, this study provides a framework for integrating multiplex profiling with multivariable machine learning approaches to nominate biologically relevant, testable biomarkers and mechanistic targets for future translational PGD studies.