Successfully developed a modelig approach that
Automatically captures nonlinearities
Provides predictions of complete time-course dynamics
- Provides a universal formulation, which
- Allows for easier translation among models
- Reduces the amount of model reduction required
- Eliminates the need for model selection
- Represents a more universal function that contains sigmoidal, Hill, and other complex deep learning functions as subsets
Allows for multiple inputs, with various input amounts, at various input times (multiple dosing strategies)
Automatically incorporates delay differential equations
Allows for multi-objective functions; e.g., combined PK and PD models
Builds models for any particular system within an entire domain of possible systems; e.g. population PK/PD models
- Makes accurate predictions for systems and inputs into those systems (model extensions to data that was not used to build the model – “test sets” in machhine learning terminology)
- System-Based Predictions
- Virtual Patients, in QSP terminology
- Input-Based Predictions
- Virtual Screening of Compounds
- Combined System/Input Predictions
- Population-Based Dosing Studies and Virtual Screening of Compounds
- Allows for more accurate virtual screening; i.e.,search entire system and input property space to optimize the dynamics curves for a given objective function
- Construction of objective functions from actual time curves rather than non-unique surrogates (AUC, half-life, etc.)
- Maximize area above minimum effective concentration
- Maximize distance to toxic concentration
- Makes use of virtual screening to identify and eliminate model overfitting (better information regarding overfit models can be derived from entire time curves rather than point estimates)