Insilico Clinical Trials and Artificial Intelligence

Randomized In Vivo Clinical Trials (RVCTs) are commonly used to assess the safety and efficacy of pharmaceutical medications and therapies. RVCTs test the candidate drug/treatment on a select group of patients. Unfortunately, this strategy has many things that could be improved. To begin with, it is highly time-consuming: in many situations, it takes more than five years for a new drug to be approved by regulatory organizations. Second, it is highly costly: an RVCT typically costs hundreds of millions of Euros to complete. Third, it is only marginally useful for rare disorders due to expenses and a shortage of patients.

The foregoing situation necessitates research into methods for assessing the safety and efficacy of pharmaceuticals and therapies that can save time and money. One of the most promising approaches in such a setting is In Silico Clinical Trials (ISCTs). Shortly, ISCTs replace the physical system (in this case, a patient) with a computer model, a Virtual Patient (VP), which accounts for the necessary properties of the physical system, much like simulation-based verification of Cyber-Physical Systems (CPSs). A virtual patient (patho-) physiology is combined with the Pharmaco-Kinetics/Dynamics (PKPD) of relevant medications in a computational model. In this situation, safety and efficacy can be assessed by simulating the drug's effect on, preferably, all VPs, just as CPS verification should try to evaluate requirements under all potential operating circumstances.

ISCTs include the following processes in their most basic form. First, list VPs compatible with (patho-) physiology, PKPD, and in vivo clinical data. This action of model validation is completed once and for all. Second, simulate the safety and efficacy of a potential medicine using the aforesaid in vivo validated VPs.

The most challenging part of conducting an ISCT is assembling a cohort of VPs that is both comprehensive and minimum in that it contains no behaviors that are incompatible with the human (patho-) physiology of interest. We'll go through the state of the art regarding VP generation methods in the following section.

First, we must acknowledge that achieving completeness and minimalism simultaneously is now unattainable. As a result, completeness is frequently favored. The goal is to ensure all human patient characteristics are adequately represented in silico.

Medical knowledge from the literature and pathway databases like KEGG (Kyoto Encyclopedia of Genes and Genomes) or Reactome are typically used to construct computational models for VPs. Computational models can be exchanged between platforms and integrated into open-standard general-purpose simulation ecosystems using open standards like the Systems Biology Markup Language (SBML).

Unfortunately, knowledge of physiology is usually qualitative. Many qualitative and quantitative solutions have been suggested to tackle such a challenge.

Qualitative techniques, often known as logic-based approaches, discretize the values of interest. Boolean models have been extensively investigated in this area. In a compositional framework where quantitative models from physiology or pharmacology are available, logical models are very beneficial for qualitative analysis. Still, they are challenging to apply within a compositional framework where quantitative models from physiology or pharmacology are present. We will concentrate on quantitative models because quantitative features are critical in an ISCT.

Ordinary Differential Equations (ODEs) with stoichiometric parameters and reaction rates determined from clinical data using model identification techniques are commonly used in quantitative models to simulate the dynamics of chemical processes. ISCT is frequently supported by such models. The lack of knowledge about the values for their parameters is the most challenging barrier to overcome when employing quantitative models. Clinical data can only be used to estimate a small number of model parameter values. As a result, most parameter values must be approximated using computational approaches, often model identification techniques.

In this situation, two approaches are commonly used: optimization techniques and statistical techniques.

Global and local optimization-based approaches have been extensively researched, relying on AI and numerical methodologies. Global optimization-based techniques are computationally intensive, yet convergence to a global optimum is guaranteed. Local optimization procedures, on the other hand, are generally less computationally intensive than global optimization approaches, but convergence is only guaranteed to a local optimum.

Both global and local optimization approaches typically rely on model identifiability, which means that different values for model parameters result in distinct model behaviors. Different model identification strategies can be utilized in this situation. The proof is in the pudding. When clinical data is sparse, identification procedures based on average data from multiple patients can be used. The resulting parameter value, however, produces an inter-patient model behavior.

Because the methodologies outlined above aim to build a model parameter value that fits existing experimental data, a large amount of data per patient is required to generate a virtual population that is representative of all human phenotypes. As a result, employing model identification approaches to generate a complete set of VPs would necessitate considering a significant number of patients (preferably at least one for each critical trait) and a large amount of clinical data for each patient. Unfortunately, this is one of the challenges that ISCTs try to overcome.

Furthermore, VP models are frequently unidentifiable globally or partially, i.e., an extensive range of parameter values leads to remarkably similar model behavior. ISCTs, on the other hand, must be able to create VPs (i.e., parameter values) that result in clearly distinct model behaviors.

Statistical methods are also commonly employed. Rather than a single value, a parameter probability distribution is inferred in such methods. They're most commonly utilized in physiology-based PKPD, quantitative VP models with observable physiological quantities as parameter values (such as blood flow, organ volumes, etc.). Most VP models, however, feature difficult-to-measure characteristics like stoichiometric constants and reaction rates. Their probability distribution functions are almost always unknown.

Methods for dealing with models whose parameters are partially unknown are required due to a lack of knowledge about VP model parameters. This ensures completeness, albeit at the expense of minimalism. Put another way, we utilize an over-estimation of the set of all medically acceptable (i.e., whose behavior is physiologically reasonable) VPs.

The FDA is already utilizing simulations as part of its regulatory strategy to expedite the approval of medications, such as immunizations for children. This is currently a hotbed of regulatory science research for patient-specific computer models. Each actual individual (digital twin) or a fictional individual whose main traits (defined by the model's inputs) are sampled from the joint distribution of a representative population can be represented by computer models of illness development and treatment response (digital trials).

From the earliest identification of biomarkers and medicines through the intricacies of clinical trials, AI healthcare businesses aim to automate new medication development. It's a journey that has been measured in years and billions of dollars until lately. The pharmaceutical industry believes that these firms can offer real value and legitimate shortcuts in developing and testing novel medication ideas.

References:

[1] U. Abdulla and R. Poteau. Identification of parameters in systems biology. Math. Biosci., 305:133–145, 2018.

[2] W. Abou-Jaoudé, P. Monteiro, A. Naldi, M. Grandclaudon, V. Soumelis, C. Chaouiya, and D. Thieffry. Model-checking to assess t-helper cell plasticity. Front. Bioeng. Biotechnol, 2, 2015.

?[3] V. Alimguzhin, F. Mari, I. Melatti, I. Salvo, and E. Tronci. Linearizing discrete-time hybrid systems. IEEE TAC, 62(10), 2017.

?[4] G. Arellano, J. Argil, E. Azpeitia, M. Benítez, M. Carrillo, P. Góngora, D. Rosenblueth, and E. Alvarez-Buylla. Antelope: A hybrid-logic model checker for branching-time boolean grn analysis. BMC Bioinf., 12(1), 2011.