Outlier Detection in Single-Cell Transcriptomics Reveals Disease-Enriched Cytotoxic Immune Populations

Single-cell RNA sequencing (scRNA-seq) enables detailed analysis of cellular heterogeneity, but standard clustering approaches can miss rare cell populations with important disease relevance. We present an outlier-aware framework that complements clustering by identifying statistically unusual cells. Applied to COVID-19 bronchoalveolar lavage fluid (BALF) samples, this approach revealed strong disease-associated enrichment of outliers: 9.62% in moderate disease and 6.16% in severe disease, compared with 1.14% in healthy controls. These cells exhibited elevated cytotoxic marker expression (NKG7, GZMB, PRF1, GNLY; 2.0-2.9x enrichment) consistent with activated effector immune states, with replication across independent datasets. To verify that out-liers were not low-quality cells, we confirmed superior QC metrics: outliers showed 68% more genes and 78% higher UMI counts than normal cells. Application to rheumatoid arthritis PBMCs showed modest enrichment (5.40% in RA vs. 4.66% in healthy), consistent with chronic inflammation. Comparison of four outlier detection algorithms identified Isolation Forest as the best-performing method. Subclustering revealed diverse functional states within outliers, including NK cytotoxic cells, activated T cells, and proliferating cells. Overall, outlier detection reveals rare, functionally distinct immune populations overlooked by conventional clustering.