Defining the molecular mechanisms of novel emerging diseases like COVID-19 is crucial to identify treatable traits to improve patient care. To circumvent a priori bias and the lack of in-depth knowledge of a new disease, we opted for an unsupervised approach, using the detailed circulating proteome, as measured by 4985 aptamers (SOMAmers), of 731 SARS-CoV-2 PCR-positive hospitalized participants to Biobanque québécoise de la COVID-19 (BQC19). The consensus clustering identified six endophenotypes (EPs) present in this cohort, with varying degrees of disease severity. One endophenotype, EP6, was associated with a greater proportion of ICU admission, mechanical ventilation, acute respiratory distress syndrome (ARDS) and death. Clinical features of this endophenotype, showed increased levels of C-reactive protein, D-dimers, elevated neutrophils, and depleted lymphocytes. Moreover, metabolomic analysis supported a role for immunothrombosis in severe COVID-19 ARDS. Furthermore, the approach enabled the identification of Fibroblast Growth Factor Receptor (FGFR) and SH2-containing transforming protein 4 (SHC4) signaling as features of the molecular pathways associated with severe COVID-19. Finally, this information was sufficient to train an accurate predictive model solely based on clinical laboratory measurements, suggesting the use of blood markers as surrogates for generalizing these EPs to new patients and automating identification of high-risk groups in the clinic.