PT - JOURNAL ARTICLE AU - Myshkin, Eugene AU - Leonardo, Steven AU - Stevens, Anne AU - Seridi, Loqmane AU - Loza, Matthew J AU - Waterworth, Dawn TI - 305 Using ACR-component based unsupervised clustering with Olink proteomics to resolve SLE heterogeneity AID - 10.1136/lupus-2023-lupus21century.18 DP - 2024 May 01 TA - Lupus Science & Medicine PG - A21--A22 VI - 11 IP - Suppl 2 4099 - http://lupus.bmj.com/content/11/Suppl_2/A21.short 4100 - http://lupus.bmj.com/content/11/Suppl_2/A21.full SO - Lupus Sci Med2024 May 01; 11 AB - Background The high heterogeneity in systemic lupus erythematosus (SLE) implies that differential treatment may be required for different patient strata, hence there is an increased need of accurate patient stratification to increase treatment success.Methods Patient serum samples from a Genuity Science cross-sectional cohort of SLE patients ranging from minimal to high disease activity (total SLEDAI range 0 - 32) on current standard-of- care treatment, sampled at a single timepoint, were profiled on the Olink Explore proteomic panel. The output of 1428 protein analyte relative concentrations was normalized and corrected for differences in site, sex and age. Low variance proteins were filtered out with median absolute deviation above 1. Unsupervised k-means clustering was performed on ACR components (‘ever’ criteria). The association of proteins with these clusters was evaluated with Kruskal-Wallis tests, random forest and limma differential expression analysis.Results Unsupervised k-means clustering yielded four patient clusters as the optimal solution, characterized in figure 1A and in the table 1 below. Approximately two thirds of the patients had reported anti-dsDNA antibodies and arthritis. Half had malar rash and lymphopenia, and one third had renal manifestations. The clusters in the table 1 are arranged left to right in the order of increasing disease severity. The disease severity was determined based on clusters association with SLEDAI and number of renal manifestations. Patients in Cluster3 had the least severe disease, displaying the lowest symptom burden with only joint/skin manifestation. Patients in Cluster2 had minimal renal disease, whereas cluster4 has no malar rash but increased numbers of renal cases. Patients in Cluster1 had the most severe disease, with all four organ manifestations. The differences between clusters were statistically significant according to chi-squared test for all parameters except arthritis.Kruskal-Wallis tests of the Olink proteomic dataset identified 10 proteins that were significantly (FDR < 0.05) associated with these four clusters. The top three proteins with FDR less than 0.01 were: Kidney injury molecule 1 (HAVCR1), interferon-γ (IFNG) and matrix metalloproteinase 3 (MMP3) (figure 1B). HAVCR1 was significantly upregulated in cluster1 compared to all others. IFNG was significantly upregulated in cluster2 and cluster1. MMP3 was also increased in severe clusters2,4,1. The other seven significant identified proteins were: CXCL13, FABP1, SSC4D, DDX58, BST2, MEP1B, GH2. Top significant proteins were also confirmed by random forest and differential expression analysis.Conclusions Using ACR component criteria in an unsupervised fashion, we were able to stratify the patients into four clusters with different levels of severity based on their historical symptomatology. These clusters have significant associations with serum protein markers derived from the Olink platform. The top three identified proteins (HAVCR1, MMP3, IFNG) were associated with SLE disease manifestation and could potentially be used to classify patients into different severity clusters. Further study is needed to validate the findings in additional cohorts and test for prognostic value.View this table:Abstract 305 Table 1 Abstract 305 Figure 1 A) Results of unsupervised k-means clustering; B) Boxplots showing expression of top 3 discovered proteins with respect to clusters