Authors :
First Author: Marcio Souza, PharmD, MBA – Praxis Precision Medicines
Presenting Author: Karl Hansen, PhD – Praxis Precision Medicines
Amulya Garimella, BA Candidate – Praxis Precision Medicines; Aadila Jasmin, MEng – Praxis Precision Medicines; Karl Hansen, PhD – Praxis Precision Medicines; Marcio Souza, PharmD, MBA – Praxis Precision Medicines
Rationale:
Almost a third of epilepsy patients are refractory to conventional anti-seizure medications. Alongside the need for novel agents is a need for sensitive and reliable biomarkers to facilitate and accelerate drug development. To date, detection of target engagement and therapeutic response to CNS drugs has been complicated by a lack of defined biomarkers, with quantitative EEG (qEEG) often attempted as a surrogate measure of target modulation and treatment-related brain activity.
This approach is limited in sensitivity due to high inter-subject variability and a lack of power inherent to single parameter methods, with changes typically only detected at drug concentrations at, or very close to, toxic levels. Combining complementary information from complex EEG signals has the potential to better characterize target engagement and, potentially, treatment effects. Here, we used machine learning methods to develop a composite qEEG biomarker to predict CNS presence and target engagement for novel agents.
Methods:
Standard EEG data were collected as part of Phase 1 studies of three next generation small molecules in development. Absolute band powers from the eyes-closed state (frequencies from 1-100 Hz) and utilizing two EEG electrode locations (frontal and central midline) were used for analysis. In addition to the five standard powers (delta, theta, beta, alpha, gamma), standard frequency band subsets were used for more granular information. Absolute power data were transformed according to a common average reference for denoising purposes.
Machine learning algorithms were trained to compute the optimal combination of EEG features for predicting treatment effect. A qEEG composite was constructed for each dataset using optimized coefficients generated from a logistic regression model trained to classify treated or placebo arms. Classification accuracy, precision, recall and F1 scores (balancing precision and recall) were calculated to evaluate performance of the qEEG composite within each dataset, along with areas under the receiver-operator characteristic (auROC) and precision-recall curves (auPRC). Data preprocessing and statistical analyses were conducted using Python (v3.10.6), the Scikit-Learn library (v1.1.1) and SAS (v9.4).
Results:
qEEG machine learning was able to predict and quantify the relative effect of drug in the brain for all tested small molecules. Across all datasets, qEEG analysis revealed pharmacodynamic effects that were significantly different from placebo at all treatment dose levels. Notably, constructed qEEG composites consistently discriminated treated vs placebo subjects with high accuracy and sensitivity ( >84% accuracy; f1 score >0.9; auROC >0.93; auPRC >0.98), with distinction from placebo evident as early as treatment day one.
Conclusions:
Application of machine learning methods to EEG data has the potential to accelerate drug development in epilepsy by definitively revealing pharmacodynamic effects of novel agents that are clearly distinguishable from placebo. We demonstrate applicability to three distinct next generation small molecules, with expected generalizability to any small molecule independent of class/target.
Funding: Praxis Precision Medicines.