AUTHORS
Arda Durmaz, PhD, Aashray Mandala, BS2, Zachary Brady, BS3, Atsushi Marumo, MD4, Serhan Unlu, MD5, Mark David Orland, MD, PhD6, Nandini Rajaram Siva, MD7, Kartik Lakhotiya, MS8, Jacob Scott, MD, DPhil, DABR9, Simona Pagliuca, MD,PhD10, Carlos Bravo-Perez, MD11, Carmelo Gurnari, MD,PhD12, Valeria Visconte, PhD13, Jaroslaw P. Maciejewski, MD, PhD, FACP14
AFFILIATIONS
1 Cleveland Clinic Research, CLEVELAND, Ohio, United States
2 Taussig Cancer Institute, CLEVELAND, Ohio, United States
3 Taussig Cancer Institute, CLEVELAND, Ohio, United States
4 Taussig Cancer Institute, CLEVELAND, Ohio, United States
5 Taussig Cancer Institute, CLEVELAND, Ohio, United States
6 Taussig Cancer Institute, CLEVELAND, Ohio, United States
7 Taussig Cancer Institute, CLEVELAND, Ohio, United States
8 Taussig Cancer Institute, CLEVELAND, Ohio, United States
9 Cleveland Clinic Research, CLEVELAND, Ohio, United States
10 Centre Hospitalier Régional Universitaire de Nancy, Nancy, France
11 Hospital Universitario Morales Meseguer, University of Murcia, Murcia, Spain
12 University of Rome Tor Vergata, Rome, Italy
13 Taussig Cancer Institute, CLEVELAND, Ohio, United States
14 Taussig Cancer Institute, CLEVELAND, Ohio, United States
ABSTRACT
The etiopathogenesis of idiopathic aplastic anemia (AA) involves an interaction of complex genetic/immunogenetic configurations and unique triggers. Such constellation of circumstances may explain the rarity of this disease. To study this intricate landscape of AA, we curated a cohort of 300 AA cases with corresponding HLA/KIR/TAP/MIC haplotypes and functional immunogenetic polymorphisms among a 600 gene set restricted to immunity. Additionally, to inform the machine learning (ML) modeling, we included data from myeloid neoplasms and controls totaling 6000 cases. Mutational profiles were filtered to exclude variants common in the general population (>1% level). Readout includes clinical phenotypes, somatic mutations and outcomes such responsiveness to IST, progression or the presence of genetic immunodeficiency or bone marrow failure.
To overcome the inherent complexity of genetic interactions, we trained a latent-fusion model via specialized tokens for different data modalities. To increase the representational capacity of the model, we included sparse mixture of experts applied both at token and sequence level with distinct routing gates. Furthermore, to mitigate potential batch effects we simultaneously trained a classifier in an adversarial fashion to the latent-fusion encoder. This allowed us to extract batch agnostic embeddings characterizing prominent input configurations. The encoder is also coupled with an additional routing mechanism to select/cluster observations into 1 of 64 experts. Given the cluster assignments, to elucidate the immunogenetic heterogeneity of AA, we analyzed the distribution of HLA, KIR, TAP, and MIC and cytokine/receptor affinity and pro-secretory genotypes allotypes across distinct patient clusters (CL).
The highest concentration of AA cases among those with complete information (n=160) localized to 3 primary clusters: CL1 (n=70), CL2 (n=40), and CL3 (n=35). Within this stratification, CL1 emerged having a highly specific immunogenetic signature. CL1 was characterized by a high prevalence of HLA-A*30:01 (78%), HLA-B*39:06 (77%), and HLA-DRB1*16:02 (75%), together with almost exclusively of the killer cell immunoglobulin-like receptors KIR3DL1*004 (90.5%) and KIR2DS4*006 (90.2%) and other immunogenetic features. CL1 was further distinguished by increased allotype frequencies in genes involved in antigen processing/presentation, specifically TAP1*01:05 (72%), and distinct MIC alleles such as MICB*028:01 (75%) and MICA*068:01 (67%). Of note is that these features were virtually absent in the other major clusters.
Conversely, CL2 exhibited mutually exclusive immunogenetic profile, highlighting a completely divergent allele concentration. This secondary group was defined by the concurrent presence of HLA-C*02:10, HLA-A*34:02, and HLA-B*15:03, each observed at a frequency of 67%. These HLA features were coupled with distinct KIR allotypes, notably KIR3DL2*142 (66%), which was not found in the surrounding patient subsets. Furthermore, the genetic uniqueness of CL2 extended to its MHC class I chain-related genes, characterized by entirely exclusive MIC allotypes including MICA*072:01 and MICB*021:01, both presenting at 50%. None of these defining alleles were observed in CL1 or CL3.
Finally, CL3 was distinct by its own robust and independent haplotype, featuring a high concentration of HLA-B*15:17 (75%) and KIR3DL1*033 (66%). Alongside these primary markers, this cohort revealed a distinct secondary profile marked by the presence of TAP1*01:04 (60.0%) and highly specific MIC variants like MICB*005:08 (75.0%) and MICA*029:01 (66.7%), all of which were completely unobserved in the first two CLs. ML-derived immunogenetic CL unique to AA were compared to other related diseases and controls. We have also investigated clinical outcomes in the immunogenetic context. In fact, the presence of inborn errors of immunity or responsiveness to IST was notable in CL1/2 with frequency of 85% and 90% CR/PR respectively. In terms of disease severity, we did not observe a significant difference across the groups.
This study underlines the highly complex landscape of immunogenomic features in AA and the non-trivial nature of consolidating multi-modal data to accurately define distinct pathogenic subgroups.
DISCLOSURES
Serhan Unlu, MD: No relevant conflicts of interest to declare.