The research, published in Frontiers in Lupus and conducted with the support of the Italian Society of Rheumatology, addressed a deceptively simple question: when rheumatologists prescribe belimumab, an antibody that blocks the B-cell activating factor BLyS, are they actually treating one population or several hidden ones? International guidelines leave considerable discretion to clinicians, defining refractory lupus loosely as an insufficient response to standard therapy with glucocorticoids and hydroxychloroquine, with or without failure of a conventional immunosuppressant. That flexibility, combined with the notorious heterogeneity of lupus itself, has long suggested that the belimumab-treated population conceals internal structure that average treatment statistics cannot capture.

To find that structure without bias, the team turned to unsupervised learning, a branch of machine learning that discovers patterns in data without any investigator-driven labels or outcome variables. The researchers fed 150 baseline variables, spanning demographics, organ-domain manifestations, serological markers, treatment history, comorbidities and damage items, into a Factor Analysis of Mixed Data, or FAMD, a generalization of classical principal component analysis designed to handle both continuous measurements and categorical features simultaneously. FAMD internally standardizes the variables, one-hot encodes categorical features and divides them by the square root of their probability, producing principal components that explain the maximum variance in the dataset. Five components were retained, together explaining 68.37 percent of the total variance, a number chosen using the scree and elbow criteria applied to the proportion of variance explained.

On top of those components, the team applied Hierarchical Clustering on Principal Components using Ward’s criterion and Euclidean distance, a method that successively merges the patients whose profiles are most similar while minimizing the within-cluster inertia. Inspection of the dendrogram and of the elbow in the inter-cluster inertia gain curve pointed unambiguously to a four-cluster solution. The analysis was implemented in Python using the Prince, scipy, scikit-learn and numpy libraries, and drew on prospectively collected records from the multicentric LIRE registry. Of 143 patients who had started a biologic agent, 128 belimumab initiators had complete baseline data, and 125 were assigned to the four main clusters; three patients formed small isolated terminal branches on the dendrogram and were excluded from cluster-level analyses.

The four phenotypes were immediately recognizable to any lupus specialist, yet their emergence from raw data was anything but guaranteed. Cluster 1, the largest with 55 patients, exhibited the lowest baseline damage scores and no prominent organ-specific signature. Cluster 2, with 28 patients, was dominated by renal involvement, with 89.3 percent carrying biopsy-proven lupus nephritis, frequent proteinuria and the strongest complement consumption, with 92.9 percent showing low complement levels. Cluster 3, the smallest with 14 patients, displayed severe neuropsychiatric and cardiovascular activity, including psychosis in 35.7 percent, cerebral vasculitis and seizures, together with the highest baseline SLICC Damage Index, enriched neurological and cardiovascular damage, more antiphospholipid antibodies and the lowest hydroxychloroquine use at 78.6 percent versus 95.2 percent in the other clusters. Cluster 4, with 28 patients, gathered patients with predominant hematological abnormalities, notably hemolytic anemia, which was significantly more prevalent in this group.

Crucially, the clusters were comparable in age, sex, ethnicity and even overall disease activity measured by the SLEDAI-2K, meaning the algorithm had separated patients not by how sick they were in aggregate but by the specific pattern of their illness. Articular and mucocutaneous manifestations, the domains where belimumab’s effectiveness is best established, were distributed evenly across all clusters, which the authors attribute to widely accepted data on the drug’s performance in those domains. The segregation instead tracked established damage, former treatments such as prior cyclophosphamide exposure, concomitant comorbidities and serological features, revealing that the belimumab population is heterogeneous in ways that go far beyond the overt manifestations that triggered the prescription.

When the researchers followed the patients for 12 months, disease activity fell significantly in the whole cohort and in every cluster, with median SLEDAI-2K dropping from 9 at baseline to 4 by month 12. Clinical response, defined as a clinical SLEDAI-2K of zero at 6 and 12 months, was statistically comparable across clusters, and in logistic regression models cluster membership did not significantly predict the probability of achieving response. Yet a suggestive pattern emerged: Cluster 2, the renal and serologically active phenotype, showed a numerically higher response rate at one year, 70.8 percent versus 52.3 percent in the other clusters, a difference that narrowly missed statistical significance at p = 0.082. The observation aligns with previous real-world evidence supporting belimumab as add-on therapy in renal lupus, and the authors note that patients in this cluster more frequently received mycophenolate mofetil, which may have acted synergistically with the biologic.

The damage trajectories told a more sobering story. Final SLICC Damage Index was highest in Cluster 3 throughout follow-up, consistent with the principle that established damage begets further damage, a burden compounded by higher glucocorticoid dosages and suboptimal hydroxychloroquine intake, both factors previously linked to damage accrual. More unexpectedly, Cluster 1, despite its milder initial phenotype and optimized hydroxychloroquine use, accrued damage faster than the other clusters over the 12 months, with a statistically significant though clinically modest increase in SDI items and no repetitive pattern among them. The authors caution that the median disease duration of around 10 years at belimumab initiation may have prevented early treatment and hampered damage prevention, and they speculate that this cluster may have gathered patients vulnerable to progression for reasons not yet identified, possibly related to interferon activity suggested by a relatively high proportion of anti-ENA positivity.

The study has clear limitations that the authors confront directly. The cohort was relatively small, data came from a real-world registry requiring imputation of residual missing values below 8 percent per variable, and clustering was performed only on patients with sufficiently complete baseline records, which may have introduced selection bias. No independent external cohort was available to validate the four clusters, so external validation will be required to confirm their robustness and generalizability. The authors also acknowledge that Italian reimbursement criteria, which require positive serology or a baseline SLEDAI-2K of at least 10, may have flattened subtle differences in patient profiling, and that differences in the temporal dynamics of response between clusters were not specifically addressed and warrant dedicated investigation.

From a translational perspective, the findings should not be read as definitive therapeutic subgroups, but they generate concrete hypotheses for future research: whether patients with renal and serologically active profiles benefit from earlier or more intensive biologic intervention, and whether molecular phenotyping, such as the enrichment of transitional B cells recently associated with better belimumab response regardless of clinical expression, could sharpen the stratification that clinical variables alone cannot achieve. The authors conclude that diverse immunological pathways can produce similar clinical manifestations, meaning that clinical and laboratory phenotyping cannot fully account for the variability threading through a lupus population. Unsupervised learning, they argue, may complement clinical judgment, leading to a deeper understanding of lupus phenotypes and, ultimately, to individualized drug allocation in a disease that has always resisted one-size-fits-all medicine.

Subject of Research: Unsupervised machine-learning algorithm identifies clinically distinct phenotypes across patients initiating belimumab in the Lupus Italian REgistry (LIRE)

Article Title: Unsupervised machine-learning algorithm identifies clinically distinct phenotypes across patients initiating belimumab in the Lupus Italian REgistry (LIRE)

Article References: Gatto, M., Venerito, V., Rozza, D., Landolfi, G., Scirè, C. A., Carrara, G., Di Nicola, C., Iagnocco, A., Sebastiani, G. D., Conti, F., Doria, A., & in collaboration with the Lupus Italian REgistry (LIRE) (2026). Unsupervised machine-learning algorithm identifies clinically distinct phenotypes across patients initiating belimumab in the Lupus Italian REgistry (LIRE). Frontiers in Lupus, 4, Article 1757086. https://doi.org/10.3389/flupu.2026.1757086

Image Credits: AI Generated

DOI: 10.3389/flupu.2026.1757086

Keywords: belimumab, systemic lupus erythematosus, machine learning, unsupervised clustering, LIRE registry, lupus nephritis, patient phenotypes, hydroxychloroquine, organ damage, precision medicine, rheumatology, biologic therapy