Eliminating Donor Availability Disparities in Hematopoietic Stem Cell Transplantation with AI-based Prediction of Permissive HLA Mismatches in the Context of Posttransplant Cyclophosphamide

PROJECT SUMMARY Allogeneic hematopoietic stem cell transplantation (HCT) is a potentially curative treatment for various hematologic malignancies and severe blood disorders. However, patients from ethnic minority groups encounter significant health disparities in HCT utilization and outcomes. These patients experience lower donor availability and suboptimal transplant outcomes due to their underrepresentation in donor registries and the polymorphic nature of human leukocyte antigen (HLA) genes. Our goal is to eliminate this critical health disparity by improving the safety and usability of HLA-mismatched transplants. We are developing the first machine learning model designed to predict permissive HLA mismatches in the context of posttransplant cyclophosphamide (PTCy)-based graft-versus-host disease (GVHD) prophylaxis, the current standard of care in HLA-mismatched transplants. Permissive HLA mismatches are those donor-recipient HLA mismatches that do not significantly increase the risk of transplant complications. The need for a PTCy-specific permissive HLA mismatch model arises because existing models fail to accurately predict outcomes in the PTCy setting. By predicting permissive HLA mismatches, transplant centers can optimize mismatched donor selection for ethnic minority patients, thereby improving HCT utilization and outcomes. To achieve this, we have designed a Cox proportional hazards neural network model that incorporates domain-specific knowledge and employs a streamlined parameter set focused on learning the importance of polymorphic amino acid positions to alloreactivity. Our approach reduces the data required for training and enhances the model’s generalizability to unseen data. Successful completion of this project has the potential to eliminate donor availability disparities in HCT.