Background We previously demonstrated comparable immune reconstitution after HLA-matched sibling (MSD), HLA-matched unrelated (MUD), and HLA-haploidentical (haplo) bone marrow transplantation (BMT) with posttransplant cyclophosphamide (PTCy). Here, we use machine learning algorithms to identify predictive immunologic variables, providing insight for future translational interventions. Methods We assessed 11 pre-transplant factors, 22 immune subsets, and 7 serum markers at day 28 and day 56 post-BMT in 70 patients who engrafted after myeloablative conditioned (MAC) haplo BMT with PTCy (50mg/kg on days +3 and +4), mycophenolate mofetil (days +5-35), and tacrolimus (days +5-180) and 74 patients who engrafted after MAC HLA-matched BMT with PTCy (50mg/kg on days +3 and +4). Classification and regression tree machine based learning with nodal selection was applied to detect risk groups for four outcomes: overall survival (OS), progression-free survival (PFS), acute graft-versus-host disease (aGVHD), and chronic graft-versus-host disease (cGVHD). Results When compared with cytomegalovirus (CMV) seronegative recipients, CMV seropositive recipients had higher CD3+ T cells, effector memory (EM) CD4+ T cells, T effector memory expressing RA (TEMRA) CD4+ T cells, CD8+ T cells, EM CD8+ T cells, and TEMRA CD8+ T cells at early time points after both BMT platforms (not shown). Pre-BMT disease status and natural killer (NK) cell count at day 28 emerged as predictive variables for survival. Two-year OS and PFS for patients in complete remission (CR) with NK cells >54 cells/µL were 98% and 89%, in CR with NK cells 55 cells/µL, the CuI of aGVHD was 18% in those with CXCL-9 Conclusion Machine learning using immune cellular and soluble markers can be successfully applied to identify risk factors for BMT outcomes. High CD4+ counts and CXCL-9 levels at day 28 predicted aGVHD and high reg3α at day 56 predicted cGVHD. Both CXCL-9 and reg3α represent potential therapeutic targets. Finally, disease status and NK cell counts at day 28 predicted OS and PFS. These data support research to determine what influences NK cell recovery as well as to investigate NK cell therapy for recipients with active disease or for those with poor early NK cell reconstitution.
Read full abstract