Anastasia Tsyplakova 1,2, Aleksandra Catic-Djorđevic 3, Nikola Stefanović 3, Vangelis Karalis 1,2
1 National And Kapodistrian University Of Athens (Athens, Greece), 2 Institute of Applied and Computational Mathematics (Heraklion, Greece), 3 Department of Pharmacy, Faculty of Medicine, University of Nis (, Serbia)
Introduction
Secukinumab, an interleukin-17A inhibitor, is an established therapy for moderate-to-severe plaque psoriasis, but wide interindividual variability in response is reported [1,2].Early identification of patients who are recovering rapidly or experiencing delayed or inadequate responses is difficult in routine clinical practice, complicating effective treatment, optimisation, and timely therapeutic decisions [3]. Patient response in psoriasis is commonly measured using disease severity indices such as the Psoriasis Area and Severity Index (PASI), which combines the severity and extent of lesions, and body surface area (BSA), a surrogate measure of the skin surface area affected by psoriatic plaques. While clinically meaningful improvement has been defined as endpoints such as PASI75 (75% reduction in PASI) and BSA < 3%, response kinetics are highly variable and not always interpreted solely based on baseline disease severity, suggesting a potential influence of latent patient characteristics and early treatment dynamics [1, 2]. Conventional methods, such as Cox regression with fixed endpoints, may not capture response kinetics in small, censored datasets. Time-to-event modeling, supplemented by exploratory and predictive machine learning, can better characterize treatment responses and support personalized management. [4–6].
Objectives
The purpose of this study was threefold:
1. To describe the time course of clinically important secukinumab response, as demonstrated by PASI75 and BSA <3%, using parametric time-to-event models, and to explain interindividual variation by determining statistically significant covariates affecting response kinetics.
2. To investigate possible latent patient phenotypes with disease severity and response dynamics, utilizing exploratory multivariate strategies suitable for mixed data structures.
3. To develop early predictive models of treatment response, integrating baseline patient characteristics and early response indicators to predict achievement of PASI75 and BSA <3% at week 4 of therapy.
Methods
Longitudinal real-world data were collected from patients with moderate–to–severe plaque psoriasis treated subcutaneously with 300 mg secukinumab in the clinical setting. PASI and BSA were measured weekly, along with baseline demographic, anthropometric, comorbidity, and laboratory data. A clinically meaningful response was defined as time to first achievement of PASI75 or BSA <3%; patients who did not reach the endpoint were right-censored at the last follow-up.
Time-to-event analyses were conducted in Monolix 2024R1 using nonlinear mixed-effects parametric survival models. Exponential, Weibull, log-logistic, and Gompertz hazard functions were evaluated using goodness-of-fit plots, parameter precision, and physiological feasibility. Covariate effects were also estimated at the characteristic time parameter to assess their influence on response timing, while accounting for censoring.
Additional nonparametric statistical analyses (Mann–Whitney U tests) were performed to explore whether baseline disease severity measures (PASI, BSA) and body mass index (BMI), known to be associated with lower secukinumab exposure [7], were associated with differences in response timing.
Latent phenotypic structure and multivariate associations were examined using multiple correspondence analysis (MCA), principal component analysis (PCA), and factorial analysis of mixed data (FAMD) in Python with the prince, scikit-learn, pandas, and numpy libraries. These unsupervised approaches facilitated simultaneous analysis of categorical and continuous variables, showed patient clustering, and separated baseline severity dimensions from response dynamics.
Supervised machine learning models were trained in Python on scikit-learn to predict PASI75 and BSA <3% achievement at week 4 by regularized logistic regression (LASSO, ridge, elastic net). Predictors included baseline patient measures and initial responses (e.g., the week-2 PASI and BSA reduction). The models’ predictive performance was assessed by five-fold cross-validated balanced accuracy, and predictor importance was quantified via permutation importance.
Results
Time-to-event data for PASI75 and BSA < 3% were described using a Weibull hazard model, which allowed for delayed treatment effects, consistent with indirect response–type pharmacodynamic mechanisms described for cytokine-targeting biologics [1,8]. In this context, higher BMI significantly prolonged the characteristic response time for both PASI75 (log Te_BMI = 3.28, RSE = 22.5%) and BSA < 3% (Te = 4.93, RSE = 25%, p < 0.001) achievement. However, baseline PASI and BSA were not significantly correlated with clinically meaningful differences in time to response. Mann–Whitney analyses demonstrated that the impact of baseline disease severity on response timing differed by clinical endpoint. For PASI75, baseline PASI and BSA caused minimal differences in median time to response (≈28–30 days), whereas BMI was significantly related to a marked delay in response (median 43 vs 27.5 days, p≪0.001). Conversely, time to BSA <3% was markedly prolonged in individuals with higher baseline PASI and BSA (median 32 vs 21 days and 56 vs 21 days; p < 0.001), indicating greater dependence of the absolute endpoint on baseline disease burden. Exploratory multivariate analyses consistently demonstrated a clear separation between baseline disease severity and treatment-response dynamics. For MCA, PCA, and FAMD, the first two dimensions accounted for about 70–75% of the total variance. Baseline PASI, BSA, and cumulative burden loaded strongly on a severity axis (loadings ≥0.80), while response kinetics loaded on an orthogonal dimension (≥0.77) aligned with BMI, age, and cardiometabolic comorbidities, demonstrating that response kinetics represents an independent phenotypic dimension. Regularized logistic regression models performed similarly and converged on similar predictive criteria. Balanced accuracies of 80–85% were achieved for models predicting PASI75 and BSA < 3% at week 4. Early PASI or BSA reduction at week 2 was the most significant predictor for outcomes across endpoints, while higher BMI consistently reduced the probability of early response. Baseline severity also affected the prediction of BSA < 3%, but treatment dynamics at earlier stages remained the main driver.
Conclusions
• The Weibull time-to-event model included delayed secukinumab response kinetics, reflecting indirect pharmacodynamic effects.
• Higher BMI consistently prolonged time to PASI75 and BSA <3% and lowered early response probability across all analytical approaches.
• Increased baseline PASI and BSA didn't affect time to PASI75 but delayed reaching BSA <3%, indicating differences between relative and absolute response endpoints.
• Multivariate studies demonstrated that response kinetics is a distinct phenotypic component, separate from baseline severity.
• Early PASI and BSA reductions at week 2 strongly predicted achievement of PASI75 and BSA <3% at week 4 (~80% accuracy), with higher BMI linked to lower early response.
• Comprehensive response kinetics and early-dynamics modeling enable early recognition of delayed responders and personalised treatment optimisation.
References:
References
1. Rodriguez-Fernandez, K.; Zarzoso-Foj, J.; Saez-Bello, M.; Mateu-Puchades, A.; Martorell-Calatayud, A.; Merino-Sanjuan, M.; Gras-Colomer, E.; Climente-Marti, M.; Mangas-Sanjuan, V. Personalized Secukinumab Treatment in Patients with Plaque Psoriasis Using Model-Informed Precision Dosing. Pharmaceutics 2024, 16, 1576, doi:10.3390/pharmaceutics16121576.
2. Ortiz-Salvador, J.-M.; Saneleuterio-Temporal, M.; Magdaleno-Tapial, J.; Velasco-Pastor, M.; Pujol-Marco, C.; Sahuquillo-Torralba, A.; Mateu-Puchades, A.; Pitarch-Bort, G.; MarÃ-Ruiz, J.-I.; Mataix-DÃaz, J.; et al. A Prospective Multicenter Study Assessing Effectiveness and Safety of Secukinumab in a Real-Life Setting in 158 Patients. Journal of the American Academy of Dermatology 2019, 81, 427–432, doi:10.1016/j.jaad.2019.02.062.
3. Erol, S.N.; Kartal, S.P. Can Week 4 PASI Response Serve as an Early Predictor of Apremilast Efficacy in Psoriasis? A Single-Centre Preliminary Study. Acta Derm Venereol 2025, 105, adv44624, doi:10.2340/actadv.v105.44624.
4. Venerito, V.; Lopalco, G.; Abbruzzese, A.; Colella, S.; Morrone, M.; Tangaro, S.; Iannone, F. A Machine Learning Approach to Predict Remission in Patients With Psoriatic Arthritis on Treatment With Secukinumab. Front. Immunol. 2022, 13, 917939, doi:10.3389/fimmu.2022.917939.
5. Gottlieb, A.B.; Mease, P.J.; Kirkham, B.; Nash, P.; Balsa, A.C.; Combe, B.; Rech, J.; Zhu, X.; James, D.; Martin, R.; et al. Secukinumab Efficacy in Psoriatic Arthritis: Machine Learning and Meta-Analysis of Four Phase 3 Trials. J Clin Rheumatol 2021, 27, 239–247, doi:10.1097/RHU.0000000000001302.
6. Damiani, G.; Conic, R.; Pigatto, P.; Carrera, C.; Franchi, C.; Cattaneo, A.; Malagoli, P.; Uppala, R.; Linder, D.; Bragazzi, N.; et al. Predicting Secukinumab Fast-Responder Profile in Psoriatic Patients: Advanced Application of Artificial-Neural-Networks (ANNs). JDD 2020, 19, 1241–1246, doi:10.36849/JDD.2020.5006.
7. UpToDate Available Online: Https://Www.Uptodate.Com/Contents/Secukinumab-Drug-Information?search=cosentyx%20adult&source=panel_search_result&selectedTitle=1~53&usage_type=panel&kp_tab=drug_general&display_rank=1#F27121747.
8. RodrÃguez-Fernández, K.; Mangas-Sanjuán, V.; Merino-Sanjuán, M.; Martorell-Calatayud, A.; Mateu-Puchades, A.; Climente-MartÃ, M.; Gras-Colomer, E. Impact of Pharmacokinetic and Pharmacodynamic Properties of Monoclonal Antibodies in the Management of Psoriasis. Pharmaceutics 2022, 14, 654, doi:10.3390/pharmaceutics14030654.
Reference: PAGE 34 (2026) Abstr 12069 [www.page-meeting.org/?abstract=12069]
Poster: Drug/Disease Modelling - Other Topics