The management of mental illness during pregnancy and lactation represents a complex clinical situation
involving a minimum of two concomitant medical conditions (i.e., pregnancy and a psychiatric disorder) and also demanding consideration of the welfare of at least two patients (i.e., mother and child).
In this project, we aim at designing algorithms and software tools to perform automatic synthesis of personalized pharmacological therapies for treating severe mood disorders in pregnant and lactating women. Existing techniques cannot solve our problem, as (i) the number of all possible therapies is extremely large, (ii) models simulation may take months and may be given as black-box (e.g., software executables), (iii) inter-patient variability and measurement errors must be taken into account.
This project aims to use advanced Artificial Intelligence (AI) techniques to model an optimization problem, allowing us to find, in a reasonable time, the best therapy for a given patient in a conveniently modeled space of therapies. We will inspect the use of Black-Box Optimization (BBO) software for efficient exploration of the space of therapies and the synthesis of therapies that grant maximum efficacy while meeting given safety requirements.
Our project is built on the following pillars:
Pillar 1. Design and implementation of a software tool to perform automatic synthesis of personalized therapies for both black- and white-box (e.g., SBML) models.
Pillar 2. Experimental evaluation via in silico clinical trials based on models of the human physiology linked to depression and PKPD models of antipsychotic drugs.
On such pillars, we will build the four main tasks of our project, i.e., two tasks for pillar one (design of mathematical model and implementation of the software) and two tasks for pillar 2 (generating the virtual patients and performing the actual evaluation on such virtual patients).
Several existing works in the scientific literature deal with the synthesing personalized therapies in particular (i.e., not general) contexts. For instance, [23,24,25,26] present approaches to the synthesis of personalized therapies that are ad hoc for the physiological models under examination. In particular, in [23], the authors conduct an ISCT using the UVA/Padova simulator to compute the optimal insulin administration regime for patients affected by Type I diabetes. [24] proposes a Monte Carlo method to identify, via numerical simulation, the optimal dosage of a drug for the treatment of tuberculosis. Finally, [26] proposes a framework for in silico combination treatment optimization focused on a pan-cancer model, using Monte Carlo methods and evolutionary algorithms. These approaches are all based on particular methodologies for modelling the biological system, i.e., ODE and ad hoc simulation models. [25] presents a framework for personalized therapy synthesis via evolutionary algorithm designed for a model of Glioblastoma Multiforme. [22] presents a backtracking-based algorithm for the search of optimal personalized treatments in the form of sequences of clinical actions. This approach uses Modelica as the definition language for the virtual patient and the therapy models. Several available software tools, such as [27,28,29], let the user solve generic optimization problems on biological systems defined via standard languages, and in particular parameter optimization and fitting on experimental data. Recently, many works proposed quantitative models of the metabolism of dopamine and serotonin (two neurotransmitters that are strongly linked to depression) as well as of the PKPD of various drugs. In [13,14] the authors propose two PBPK models of the concentration of Escitalopram (an SSRI drug recommended for post-partum depression) in breast milk. [15] developed a mechanistic PKPD model for the prolactin release, a common adverse effect, following administration of dopamine receptor antagonists. A PBPK model of dopamine D2 receptor occupancy, which is a key element in the assessment of efficacy and safety of anti-psychotic drugs, has been proposed in [16]. Mathematical models of the metabolism of dopamine have been proposed in [17] and, more recently, in [18].
The software that we propose to design and implement (Pillar 1) in this project will innovate the tools at the state of the art from the points of view of generality and flexibility. Namely, existing approach make many assumptions over the model and the therapies that hinder their employment in the actual clinical context. Several existing works in the scientific literature deal with the problem of synthesis of personalized therapies, albeit in particular (i.e., not general) contexts. The approaches presented in [23,24,25,26], unlike the one we propose, are based on particular methodologies for modelling the biological system, i.e., ODE and ad hoc simulation models. Our software, unlike existing methods, will not make any assumption over the model formulation and will be able to synthesize the optimal therapy while treating the model as a black-box. Our tool, however, will explicitly support models defined in standard modelling languages (e.g., SBML). This, together with the fact that the tool will not require any coding from the user, will enable the personalization of treatments directly in the clinical context. Unlike the aforementioned approaches, the software that will be produced in this project will not be tied to any particular physiological model or modelling methodology. Instead, the users of our software will be able to conduct any ISCT and obtain the optimal therapy among those that can be represented by the therapy model (or template) provided as an input, on any physiological model that can be simulated (black-box). Existing approaches, including [22,24,26,25] employ many optimization methods designed ad hoc for the problem, based on the type of model. Our software will allow the user to compute personalized therapies independently of the language used to define the models, and will be able to use any off-the-shelf BBO solver to solve the optimization problem. Currently available tools for optimization of biological systems, such as [27,28,29], lack the flexibility needed to compute the optimal therapy for a population of virtual patients. Our tool, focusing on the sole problem of synthesizing personalized therapies, will provide the means to easily apply precision medicine in the clinical context. Lastly, the project will represent a step forward in the study of pharmacological treatment of perinatal depression. Namely, we plan to generate a population of physiologically admissible virtual patients and use it to conduct, for the first time, ISCT to analyze treatment efficacy and safety in a model-based approach, exploiting several models of the patho-physiology of interest for the syndrome.