A computational approach to explore the link between skin biomarkers and clinical outcomes in psoriasis

Aakankschit Nandkeolyar

doi:10.17918/00000505

Psoriasis is a chronic autoimmune skin disease that affects nearly 2% of the world's population. It is characterized by the formation of red plaques on the skin. Such plaques can cause irritation and pain to patients, not to mention the plaques are a major source of anxiety related to appearance for patients, making psoriasis a disease with high morbidity. While there are currently no cures to psoriasis, anti-inflammatory pharmaceuticals can be used to treat psoriasis by reducing the severity of the symptoms. The efficacy of treatments for psoriasis is measured by improvement in clinical outcomes, the most popular of which is Psoriasis Area Severity Index Score (PASI). These scores are provided by a dermatologist to quantify the severity of psoriasis in a patient. Quantitative Systems Pharmacology (QSP) is a new field in the area of computational disease modeling that allows for quantification and prediction of disease biomarkers in virtual patient models. Such prediction can potentially be correlated with clinical outcomes such as PASI scores to allow a scientist to predict clinical outcomes in virtual patients. The current study proposes the use of machine learning algorithms to bridge the gap between QSP simulation data for skin biomarkers of disease processes and clinical outcome measures to predict PASI scores post-treatment to get a preclinical estimate of the effect of novel drugs on PASI scores. This work specifically aims to develop predictive models for clinical outcome (PASI scores) using biomarker data (generated by a QSP model of psoriasis built by GlaxoSmithKline) via an array of machine learning algorithms which include regularized regression models, Support Vector Machines regression and ensemble methods (kth-Nearest Neighbor, Random Forest and Stochastic Gradient Boosting). To construct and test the models, a comprehensive dataset was compiled from across 8 different clinical trials that accounted for a total of 2131 patients, where these trials all together accounted for 3 different treatment regimens that target different biomarkers in patients, these included anti-IL-17, anti-IL-23 and anti-TNF-alpha. The dataset contained both PASI scores and with their respective serum biomarkers and was developed via systematic literature review. To compile the dataset, the average PASI scores across all patients in the clinical trial, recorded every week were found in literature and used as a sample PASI score. The biomarker data associated with the PASI scores was generated using an "average patient" simulation of the GlaxoSmithKline QSP model. The models were trained and tested using 6-fold cross-validation. Results indicate that machine learning methods can be utilized to predict clinical outcomes, the best models were generated using regularized regression algorithm and had a mean square error averaging 14.7 PASI score units (within PASI score range of 0-72) with an R-squared of 0.75. This is the first attempt to develop a predictive model of PASI score in Psoriasis from serum biomarker data generated using QSP simulation data to clinical state. The significance of this research is based on the novel approach for using computational methods to potentially predict PASI scores for novel psoriasis treatments and aid in accelerating clinical development.

A computational approach to explore the link between skin biomarkers and clinical outcomes in psoriasis

Files and links (1)

Abstract

Metrics

Details

A computational approach to explore the link between skin biomarkers and clinical outcomes in psoriasis

Files and links (1)

Abstract

Metrics

Details

Drexel University Social media