Multi-Omics and Machine Learning Predict the Risk of Acute Graft Versus Host Disease

Background

Acute graft versus host disease (aGVHD) after allogeneic hematologic stem cell transplantation (allo-HSCT) remains the prevailing cause of non-relapse mortality. In addition to improving aGVHD prophylaxis strategies, studies have identified clinical factors related to the patient and donor, as well as biomarkers indicative of aGVHD development prior to the onset of clinical symptoms. Multi-omics methodologies offer insights into proteins or metabolites that reflect current disease states. In this study, we employ a machine learning approach to develop a multi-omics-based predictive model for aGVHD.

Methods

This study involved 38 patients first receiving allo-HSCT from April 2022 to November 2022, with 20 patients clinically diagnosed as aGVHD within 100 days post-transplantation. Peripheral blood samples were collected at several time points: 7, 14, 21, and 28 days after transplantation (day 7, day 14, day 21, and day 28). Plasma separated by centrifugation was used for proteomics and metabolomics analyses. Patients with confirmed aGVHD were excluded from the aGVHD group at each time point. Proteomics data were acquired by liquid chromatography-tandem mass spectrometry with tandem mass tags-based quantification, while metabolomics was analyzed via untargeted metabolomics techniques. Proteins or metabolites with less than 50% missing values were retained and imputed using the SeqKNN method. The support vector machine (SVM) model was utilized for aGVHD prediction with fuzzy clustering approved for the temporal metabolite analysis.

Results

Patients with and without aGVHD shared similar baseline characteristics. A total of 4290 plasma proteins were identified, with 2100 proteins retained for subsequent analysis. 14 and 11 proteins showed differences between patients with aGVHD and non-aGVHD on day 7 and day 14, respectively. 36 proteins exhibited differential expression between day 7 and day 14. Analysis of 2422 plasma metabolites revealed 99 and 93 metabolites differentially expressed on day 7 and day 14, with 40 metabolites showing disparities between day 7 and day 14, respectively. The previously reported biomarker elafin displayed a statistically significant difference between aGVHD and non-aGVHD before symptom onset. The clinical test results showed substantial differences in IL-5, IL-1β, IFN-γ， chloride ion, and total protein concentration between groups on day 14. Differential proteins, metabolites, elafin, and clinical test results were integrated into SVM model for aGVHD prediction. The SVM recursive feature elimination method selected 9 variables for prediction with a relatively low root mean square error, including zinc alpha-2 glycoprotein (ZA2G), N90-VRC38.03 heavy chain variable region, phosphatidylcholine (PC) 15:0_20:5, D-ribose, diacylglycerol (DG) 10:0_18:1, lysopa 18:0, and IL-1β. Notably, ZA2G exhibited differential expression not only at days 7 and 14 but also across the transition from day 7 to day 14. 27 (70%) and 11 (30%) patients were separated into training and test datasets. In the test set, the SVM algorithm showed an area under the curve of 0.914 with 83% accuracy, 80% sensitivity, and 86% specificity. Additionally, the longitudinal changes in proteins and metabolites varied between the aGVHD and non-aGVHD groups. Elevated ZA2G levels were reported in ulcerative colitis (UC) patients, while PC was reduced in UC. Studies have demonstrated that D-ribose and DG had an anti-colitis effect in mice. In our study, ZA2G was increased in patients with aGVHD and decreased in patients with non-aGVHD from day 7 to day 14. Patients of aGVHD group exhibited a distinguishable decline of PC 15:0_20:5, D-ribose, DG 10:0_18:1, and lysopa 18:0 from day 7 to day 14, while patients without aGVHD did not observe differences. The variation between day 14, 21, and 28 was not significant for patients in aGVHD group.

Conclusion

The machine learning model integrating multi-omics with clinical test results can successfully identify risk biomarkers for aGVHD, advancing the prediction window to day 7 and day 14 post-transplantation. Concomitantly, different biomarkers may present at divergent time points during aGVHD development. Large-sample studies are currently underway to validate this observation. In vitro and ex vivo assays are required to further investigate the specific mechanism of proteins and metabolites for aGVHD development.

No relevant conflicts of interest to declare.