Artificial Intelligence for Quantitative Modeling in Drug Discovery and Development: An Innovation and Quality Consortium Perspective on Use Cases and Best Practices

Recent breakthroughs in artificial intelligence (AI) and machine learning (ML) have ushered in a new era of possibilities across various scientific domains. One area where these advancements hold significant promise is model-informed drug discovery and development (MID3). To foster a wider adoption and acceptance of these advanced algorithms, the Innovation and Quality (IQ) Consortium initiated the AI/ML working group in 2021 with the aim of promoting their acceptance among the broader scientific community as well as by regulatory agencies. By drawing insights from workshops organized by the working group and attended by key stakeholders across the biopharma industry, academia, and regulatory agencies, this white paper provides a perspective from the IQ Consortium. The range of applications covered in this white paper encompass the following thematic topics: (i) AI/ML-enabled Analytics for Pharmacometrics and Quantitative Systems Pharmacology (QSP) Workflows; (ii) Explainable Artificial Intelligence and its Applications in Disease Progression Modeling; (iii) Natural Language Processing (NLP) in Quantitative Pharmacology Modeling; and (iv) AI/ML Utilization in Drug Discovery. Additionally, the paper offers a set of best practices to ensure an effective and responsible use of AI, including considering the context of use, explainability and generalizability of models, and having human-in-the-loop. We believe that embracing the transformative power of AI in quantitative modeling while adopting a set of good practices can unlock new opportunities for innovation, increase efficiency, and ultimately bring benefits to patients.

Quantitative systems pharmacology model of GITR-mediated T cell dynamics in tumor microenvironment

T cell interaction in the tumor microenvironment is a key component of immuno-oncology therapy. Glucocorticoid-induced tumor necrosis factor receptor (TNFR)-related protein (GITR) is expressed on immune cells including regulatory T cells (Tregs) and effector T cells (Teffs). Preclinical data suggest that agonism of GITR in combination with Fc-γ receptor-mediated depletion of Tregs results in increased intratumoral Teff:Treg ratio and tumor shrinkage. A novel quantitative systems pharmacology (QSP) model was developed for the murine anti-GITR agonist antibody, DTA-1.mIgG2a, to describe the kinetics of intratumoral Tregs and Teffs in Colon26 and A20 syngeneic mouse tumor models. It adequately captured the time profiles of intratumoral Treg and Teff and serum DTA-1.mIgG2a and soluble GITR concentrations in both mouse models, and described the response differences between the two models. The QSP model provides a quantitative understanding of the trade-off between maximizing Treg depletion versus Teff agonism, and offers insights to optimize drug design and dose regimen.

Mathematical modeling of SARS-CoV-2 viral dynamics and treatment with monoclonal antibodies

The novel coronavirus (SARS-CoV-2) affects primarily the respiratory tract, and if left unchecked can cause a spectrum of pathological manifestations such as pneumonia, acute respiratory distress syndrome, myocardial injury, thromboembolism, and acute kidney injury. Medication strategies have involved minimizing the spread of the virus through antiviral medications (monoclonal antibodies or nucleotide reverse transcriptase inhibitors). Here, we develop a mathematical model that simulates viral dynamics in an untreated individual, and the evaluate the impact that a monoclonal antibody can have on slowing viral replication. Drug pharmacokinetics (PK) was informed by a typical two-compartment PK model with parameters typical of a monoclonal antibody, with a third compartment for the lung included as the drug site of action. The viral dynamics were captured using a simplified model describing uninfected target cells, infected target cells, and viral load in the body. The mechanism of action of the simulated antiviral is based on binding to the virus, thereby preventing it from infecting healthy cells. The model is used to project dosages needed to prevent severe disease under a variety of simulated conditions and subject to realistic constraints. The proposed model can capture a variety of scenarios of longitudinal viral dynamics and assess the impact of antiviral therapy on disease severity and duration. The described approach can be easily adapted to rapidly assess the dosages needed to affect duration and outcome of other viral infections and can serve as part of a fast and efficient scientific and modeling response strategy in the future as needed.

Systematic in silico analysis of clinically tested drugs for reducing amyloid-beta plaque accumulation in Alzheimer's disease

Introduction

Despite strong evidence linking amyloid beta (Aβ) to Alzheimer's disease, most clinical trials have shown no clinical efficacy for reasons that remain unclear. To understand why, we developed a quantitative systems pharmacology (QSP) model for seven therapeutics: aducanumab, crenezumab, solanezumab, bapineuzumab, elenbecestat, verubecestat, and semagacestat.

Methods

Ordinary differential equations were used to model the production, transport, and aggregation of Aβ; pharmacology of the drugs; and their impact on plaque.

Results

The calibrated model predicts that endogenous plaque turnover is slow, with an estimated half‐life of 2.75 years. This is likely why beta‐secretase inhibitors have a smaller effect on plaque reduction. Of the mechanisms tested, the model predicts binding to plaque and inducing antibody‐dependent cellular phagocytosis is the best approach for plaque reduction.

Discussion

A QSP model can provide novel insights to clinical results. Our model explains the results of clinical trials and provides guidance for future therapeutic development.

Establishment of a Disease-Drug Trial Model for Postmenopausal Osteoporosis: A Zoledronic Acid Case Study

Costly and lengthy clinical trials hinder the development of safe and effective treatments for postmenopausal osteoporosis. To reduce the expense associated with these trials, we established a mechanistic disease-drug trial model for postmenopausal osteoporosis that can predict phase 3 trial outcome based on short-term bone turnover marker data. To this end, we applied a previously developed model for tibolone to bisphosphonates using zoledronic acid as paradigm compound by (1) linking the mechanistic bone cell interaction model to bone turnover markers as well as bone mineral density in lumbar spine and total hip, (2) employing a mechanistic disease progression function, and (3) accounting for zoledronic acid's mechanism of action. Once developed, we fitted the model to clinical trial data of 581 postmenopausal women receiving (1) 5-mg zoledronic acid in year 1 and saline in year 2, (2) 5-mg zoledronic acid in year 1 and year 2, or (3) placebo (saline), calcium (500 mg), and vitamin D (400 IU). All biomarker data was fitted reasonably well, with no apparent bias or model misspecification. Age, years since menopause, and body mass index at baseline were identified as significant covariates. In the future, the model can be modified to explore the link between short-term biomarkers and fracture risk.

Population Pharmacokinetic Modeling of Acetaminophen Protein Adducts in Adults and Children

Acetaminophen protein adducts (adducts) are a well-established biomarker to diagnose acetaminophen toxicity. To date, the quantitative relationship between acetaminophen exposure, which drives adduct formation, and adduct exposure remains to be established. Our study characterized the adduct formation and disposition in adults using the approach of population parent-metabolite modeling. It demonstrated formation-limited pharmacokinetics (PK) for adducts in healthy subjects. This finding expands the existing knowledge on adduct PK that showed an apparent long elimination half-life. We then allometrically scaled the adduct PK model to children, simulated the adduct profiles, and compared these simulated profiles with those observed in an independent cohort of children. The scaled model significantly overpredicted the adduct concentrations in children early on in treatment and underpredicted concentrations following repeated acetaminophen doses. These results suggest that children demonstrate different adduct PK behavior from that of adults, most likely because of increased reactive metabolite detoxification in children. In summary, we described the first PK model linking acetaminophen and acetaminophen protein adduct concentrations, which provides a semimechanistic understanding of varying profiles of adduct exposure in adults and children.

Regulatory Perspectives in Pharmacometric Models of Osteoporosis

Osteoporosis is a disorder of the bones in which they are weakened to the extent that they become more prone to fracture. There are various forms of osteoporosis: some of them are induced by drugs, and others occur as a chronic progressive disorder as an individual gets older. As the median age of the population rises across the world, the chronic form of the bone disease is drawing attention as an important worldwide health issue. Developing new treatments for osteoporosis and comparing them with existing treatments are complicated processes due to current acceptance by regulatory authorities of bone mineral density (BMD) and fracture risk as clinical end points, which require clinical trials to be large, prolonged, and expensive to determine clinically significant impacts in BMD and fracture risk. Moreover, changes in BMD and fracture risk are not always correlated, with some clinical trials showing BMD improvement without a reduction in fractures. More recently, bone turnover markers specific to bone formation and resorption have been recognized that reflect bone physiology at a cellular level. These bone turnover markers change faster than BMD and fracture risk, and mathematically linking the biomarkers via a computational model to BMD and/or fracture risk may help in predicting BMD and fracture risk changes over time during the progression of a disease or when under treatment. Here, we discuss important concepts of bone physiology, osteoporosis, treatment options, mathematical modeling of osteoporosis, and the use of these models by the pharmaceutical industry and the Food and Drug Administration.

Markov Mixed Effects Modeling Using Electronic Adherence Monitoring Records Identifies Influential Covariates to HIV Preexposure Prophylaxis

Adherence is a major factor in the effectiveness of preexposure prophylaxis (PrEP) for HIV prevention. Modeling patterns of adherence helps to identify influential covariates of different types of adherence as well as to enable clinical trial simulation so that appropriate interventions can be developed. We developed a Markov mixed-effects model to understand the covariates influencing adherence patterns to daily oral PrEP. Electronic adherence records (date and time of medication bottle cap opening) from the Partners PrEP ancillary adherence study with a total of 1147 subjects were used. This study included once-daily dosing regimens of placebo, oral tenofovir disoproxil fumarate (TDF), and TDF in combination with emtricitabine (FTC), administered to HIV-uninfected members of serodiscordant couples. One-coin and first- to third-order Markov models were fit to the data using NONMEM^® 7.2. Model selection criteria included objective function value (OFV), Akaike information criterion (AIC), visual predictive checks, and posterior predictive checks. Covariates were included based on forward addition (α = 0.05) and backward elimination (α = 0.001). Markov models better described the data than 1-coin models. A third-order Markov model gave the lowest OFV and AIC, but the simpler first-order model was used for covariate model building because no additional benefit on prediction of target measures was observed for higher-order models. Female sex and older age had a positive impact on adherence, whereas Sundays, sexual abstinence, and sex with a partner other than the study partner had a negative impact on adherence. Our findings suggest adherence interventions should consider the role of these factors.

Investigation of the Dermal Absorption and Irritation Potential of Sertaconazole Nitrate Anhydrous Gel

Effective topical therapy of cutaneous fungal diseases requires the delivery of the active agent to the target site in adequate concentrations to produce a pharmacological effect and inhibit the growth of the pathogen. In addition, it is important to determine the concentration of the drug in the skin in order to evaluate the subsequent efficacy and potential toxicity for topical formulations. For this purpose, an anhydrous gel containing sertaconazole nitrate as a model drug was formulated and the amount of the drug in the skin was determined by in vitro tape stripping. The apparent diffusivity and partition coefficients were then calculated by a mathematical model describing the dermal absorption as passive diffusion through a pseudo-homogenous membrane. The skin irritation potential of the formulation was also assessed by using the in vitro Epiderm™ model. An estimation of the dermal absorption parameters allowed us to evaluate drug transport across the stratum corneum following topical application. The estimated concentration for the formulation was found to be higher than the MIC100 at the target site which suggested its potential efficacy for treating fungal infections. The skin irritation test showed the formulation to be non-irritating in nature. Thus, in vitro techniques can be used for laying the groundwork in developing efficient and non-toxic topical products.

Linking the population pharmacokinetics of tenofovir and its metabolites with its cellular uptake and metabolism

Empirical pharmacokinetic models are used to explain the pharmacokinetics of the antiviral drug tenofovir (TFV) and its metabolite TFV diphosphate (TFV-DP) in peripheral blood mononuclear cells. These empirical models lack the ability to explain differences between the disposition of TFV-DP in HIV-infected patients vs. healthy individuals. Such differences may lie in the mechanisms of TFV transport and phosphorylation. Therefore, we developed an exploratory model based on mechanistic mass transport principles and enzyme kinetics to examine the uptake and phosphorylation kinetics of TFV. TFV-DP median C_max from the model was 38.5 fmol/10⁶ cells, which is bracketed by two reported healthy volunteer studies (38 and 51 fmol/10⁶ cells). The model presented provides a foundation for exploration of TFV uptake and phosphorylation kinetics for various routes of TFV administration and can be updated as more is known on actual mechanisms of cellular transport of TFV.

Agmatine induced NO dependent rat mesenteric artery relaxation and its impairment in salt-sensitive hypertension

l-Arginine and its decarboxylated product, agmatine are important mediators of NO production and vascular relaxation. However, the underlying mechanisms of their action are not understood. We have investigated the role of arginine and agmatine in resistance vessel relaxation of Sprague-Dawley (SD) and Dahl salt-sensitive hypertensive rats. Second or 3rd-order mesenteric arterioles were cannulated in an organ chamber, pressurized and equilibrated before perfusing intraluminally with agonists. The vessel diameters were measured after mounting on the stage of a microscope fitted with a video camera. The gene expression in Dahl rat vessel homogenates was ascertained by real-time PCR. l-Arginine initiated relaxations (EC50, 5.8±0.7mM; n=9) were inhibited by arginine decarboxylase (ADC) inhibitor, difluoromethylarginine (DFMA) (EC50, 18.3±1.3mM; n=5) suggesting that arginine-induced vessel relaxation was mediated by agmatine formation. Agmatine relaxed the SD rat vessels at significantly lower concentrations (EC50, 138.7±12.1μM; n=22), which was compromised by l-NAME (l-N(G)-nitroarginine methyl ester, an eNOS inhibitor), RX821002 (α-2 AR antagonist) and pertussis toxin (G-protein inhibitor). The agmatine-mediated vessel relaxation from high salt Dahl rats was abolished as compared to that from normal salt rats (EC50, 143.9±23.4μM; n=5). The α-2A AR, α-2B AR and eNOS mRNA expression was downregulated in mesenteric arterioles of high-salt treated Dahl hypertensive rats. These findings demonstrate that agmatine facilitated the relaxation via activation of α-2 adrenergic G-protein coupled receptor and NO synthesis, and this pathway is compromised in salt-sensitive hypertension.

Glutathiyl radical as an intermediate in glutathione nitrosation

Nitrosation of thiols is thought to be mediated by dinitrogen trioxide (N(2)O(3)) or by nitrogen dioxide radical (()NO(2)). A kinetic study of glutathione (GSH) nitrosation by NO donors in aerated buffered solutions was undertaken. S-nitrosoglutathione (GSNO) formation was assessed spectrophotometrically and by chemiluminescence. The results suggest an increase in the rate of GSNO formation with an increase in GSH with a half-maximum constant EC(50) that depends on NO concentration. Our observed increase in EC(50) with NO concentration suggests a significant contribution of ()NO(2)-mediated nitrosation with the glutathiyl radical as an intermediate in the production of GSNO.