Introduction At the end of 2020, a novel entity dubbed VEXAS (for vacuoles, E1 enzyme, X-linked, auto-inflammatory, somatic) was described by Beck et al. (NEJM, 2020). This syndrome is defined molecularly by a mutation of the UBA1 gene that codes for the E1 enzyme, a major actor of ubiquitinylation. From a pathophysiological point of view, UBA1 mutations induce the transcription of a deficient enzyme. Affected cells display an altered ubiquitinylation of misfolded proteins and dysregulated autophagy. This leads in fine to an impairment of several pathways of innate immunity, overproduction of a range of pro-inflammatory cytokines and, more generally, systemic inflammation. Clinical manifestations are pleomorphic and patient-dependent, with reported cases of fever, vasculitis, ophthalmic lesions, chondritis, Sweet syndrome or pulmonary interstitial infiltration. From a hematological point of view, thromboses have also been described as well as several cytologic alterations. The latter include cytopenias of any lineage, myeloid hyperplasia with erythroblastic hypoplasia (M/E) and, notably, the presence of vacuoles essentially in granulocytes and erythroblasts precursors. Some cases are associated to myelodysplastic syndrome (MDS). Here, an in-depth morphological examination of VEXAS and non-VEXAS patients was performed, to delineate and quantify several anomalies. These data were then fed to a deep learning algorithm to evaluate the additional help provided by this artificial intelligence (AI) tool in predicting VEXAS syndrome and orienting further analyses towards the identification of the molecular anomaly. Patients, material and methods To identify diagnostic tools for VEXAS syndrome on blood and bone marrow smears, 4 patient cohorts were analyzed :i) patients with a diagnosis of VEXAS syndrome confirmed by genetic analysis (n=6), ii) MDS patients (n=4), iii) patients with an initial clinical suspicion of VEXAS invalidated by molecular analysis (UBA1 WT: n=3) and iv) 4 healthy controls. For each patient, morphological examination of blood and marrow smears was performed independently by two experts in a blind fashion. In order to guide the diagnosis of the syndrome on the blood smears, a deep learning method was trained in a multilabel classification setting to automatically evaluate the percentages of abnormal polymorphonuclears (PMN). The final dataset was therefore composed of PMN from healthy controls (n=5, 519 images) and from VEXAS syndrome (n=4, 938 images) (DI microscope, Cellavision ®). Results Among the 6 patients with VEXAS syndrome, the main clinical manifestations were: asthenia 100%; fever 83.3%; skin involvement 66.7%; thrombosis 66.7%, chondritis 33.3% and interstitial lung disease 33.3%. The most frequent cytopenia was macrocytic anemia in 66.7% of the patients. Upon morphological examination of blood smears, the following abnormalities were statistically associated with the diagnosis of VEXAS : percentage of pseudo-Pelger PMNs (p=0.003), percentage of PMNs with more than 4 nuclear projections (Barr bodies; p=0.0009), percentage of vacuolated PMNs (p=0.01). On bone marrow smears, compared to the MDS and UBA1 WT cohort, patients with VEXAS syndrome had significantly more PMNs with a pseudo-Pelger aspect (p<0.05), more vacuolization of granular (p<0.05) and erythroblastic (p<0.05) precursors. Different deep learning architectures were tested on blood smear images, with the objectives of distinguishing abnormal PMN from others. The trainings were conducted by cross-validation, and then the algorithm was evaluated on new patients. The metrics used to evaluate algorithm performance were precision, recall, F1-score and Hamming loss. These data allowed to establish a standardized computer aided diagnosis tool. Discussion/conclusion The diagnosis of VEXAS syndrome is complex. This analysis identified cytological features on blood smears leading to the development of a deep learning-based diagnosis algorithm. Promising leads are currently investigated to improve the efficiency and robustness of this deep learning algorithm such as the inclusion of other pathologies in order to train a disease agnostic anomaly detection algorithm, or leveraging data from other modalities as prior knowledge.