Integrating Model‐Informed Drug Development With AI: A Synergistic Approach to Accelerating Pharmaceutical Innovation

ABSTRACT

The pharmaceutical industry constantly strives to improve drug development processes to reduce costs, increase efficiencies, and enhance therapeutic outcomes for patients. Model‐Informed Drug Development (MIDD) uses mathematical models to simulate intricate processes involved in drug absorption, distribution, metabolism, and excretion, as well as pharmacokinetics and pharmacodynamics. Artificial intelligence (AI), encompassing techniques such as machine learning, deep learning, and Generative AI, offers powerful tools and algorithms to efficiently identify meaningful patterns, correlations, and drug–target interactions from big data, enabling more accurate predictions and novel hypothesis generation. The union of MIDD with AI enables pharmaceutical researchers to optimize drug candidate selection, dosage regimens, and treatment strategies through virtual trials to help derisk drug candidates. However, several challenges, including the availability of relevant, labeled, high‐quality datasets, data privacy concerns, model interpretability, and algorithmic bias, must be carefully managed. Standardization of model architectures, data formats, and validation processes is imperative to ensure reliable and reproducible results. Moreover, regulatory agencies have recognized the need to adapt their guidelines to evaluate recommendations from AI‐enhanced MIDD methods. In conclusion, integrating model‐driven drug development with AI offers a transformative paradigm for pharmaceutical innovation. By integrating the predictive power of computational models and the data‐driven insights of AI, the synergy between these approaches has the potential to accelerate drug discovery, optimize treatment strategies, and usher in a new era of personalized medicine, benefiting patients, researchers, and the pharmaceutical industry as a whole.

1. Introduction

Bringing a drug to market is a complex and lengthy process involving several years of preclinical and clinical research and regulatory interactions. The pharmaceutical industry invests billions of dollars for each drug in development, with a historically high failure rate [1]. Notably, advances in paradigms such as Model‐Informed Drug Development (MIDD), and, more recently, artificial intelligence (AI), have become potential game‐changers in this uphill task of drug development. Traditional drug development had a ligand focus, with the emphasis being on discovering new small molecules with drug‐like activity. Following this, the focus shifted to identifying drug targets, the biological macromolecules in the pathogen or human host, which could be modulated using drugs. In recent decades, a system perspective has become central to advancing drug discovery and development [2], with the emphasis slowly moving toward multiple targets and polypharmacology [3]. All along, modeling has come to play a key role, with in silico inputs guiding every stage of drug development, including providing mechanistic insights into drug action, enabling more rational drug trials, and improving drug safety and efficacy [4]. With the explosion of data, development of new algorithms, and increases in computational power, AI and machine learning (ML)‐based methods have made significant strides in recent years, with increasing impact on the drug development process [5]. AI, ML, and deep learning (DL) are often used interchangeably, yet represent distinct aspects of a broad set of computational systems to tackle complex tasks. AI is the overarching field encompassing various computational techniques aimed at enabling machines to perform tasks that typically require human intelligence. AI systems, driven by both machine‐ and human‐generated inputs, can make predictions, recommendations, or decisions to influence real and virtual environments [6]. Within AI, ML refers to a set of techniques that train algorithms to improve performance on specific tasks, based on data. ML algorithms may be supervised, when labeled data are provided, or unsupervised, when unlabeled data are processed to seek out hidden patterns and structures in the data.

In drug development, ML methods are used to screen potential drug candidates, predict drug interactions, and analyze patient data for improving trial outcomes. ML enables AI systems to continuously learn and improve, as they get exposed to more real‐world data. DL is a specialized branch of ML, focused on a specific algorithm—neural networks—with multiple hidden layers that process data hierarchically [6]. Each layer transforms the input data into more abstract representations, enabling the model to learn complex patterns. DL excels in tasks involving high‐dimensional data, such as genomics, molecular modeling, or medical imaging. For instance, DL models can identify subtle molecular features associated with drug efficacy or recognize patterns in medical images to support diagnosis. Generative AI (GenAI) is a new class of AI models, which generates derived synthetic content by emulating the structure and characteristics of input data [6]. GenAI works by approximating the statistical distribution of the original data, to produce synthetic text, images, or other datasets. GenAI models can also assist in designing novel drug candidates or preparing regulatory document drafts, accelerating drug discovery and development processes. In this review, we overview key aspects of MIDD and AI in drug development. We place special emphasis on the unique synergy between the methodologies, illustrating their potential to accelerate innovation in the drug development process, via real‐world examples.

2. MIDD in Drug Development

MIDD encompasses various methodologies, from simple non‐compartmental analysis and in vitro in vivo correlation (IVIVC) to more complex population pharmacokinetics and quantitative systems pharmacology (QSP). These approaches range from extracting basic drug properties to capturing intricate physiological and drug interactions for strategic decision‐making. Well‐established modeling approaches in drug development, including estimating dose from IVIVC, allometric scaling, and identifying variability of drug exposure in a specific population, often are empirical in nature and focused on interpolating or extrapolating from individual studies and optimizing subsequent studies. These approaches are fit‐for‐purpose and well accepted by regulatory agencies, and with best practices that often direct model choice [6], they are reproducible within reason. However, these methods may be limited to expanding from specific data to address a specific question and typically do not explicitly include multiple, disparate datasets or prior knowledge that may be available under different measurement formats, perturbations, or settings.

More recently, modeling approaches that integrate multiple data sources and at multiple types and scales (e.g., multi‐omics, clinical biomarkers, and in vitro and animal experiments), are increasingly being applied in drug discovery and development [7, 8]. These include physiology‐based pharmacokinetic (PBPK) models, semi‐mechanistic PK‐PD models, and QSP models of varying levels of biological and mathematical complexity. These models typically include a compartmental framework of pharmacological and biological entities (molecules, cells, proteins, biomarkers) interacting in a single or multiple compartments (e.g., cells, organs, and/or whole organism). These models often will include generally accepted biological processes, available data on flow rates across the various compartments, life cycle, and clearance parameters, and interactions between the various entities modeled using mathematical formalisms. This framework is quantified using alternate model parameterizations (“virtual patients,” in QSP) to capture observed variability in clinical response assumed to be due to underlying physiological variability. Such models, which may be applicable to multiple targets, conditions, and interventions, may be highly subjective in design; inclusion and choice of data require multiple development iterations, constant testing of assumptions and limitations, significant resource and time, and customized, one‐off approaches to address model complexity. These issues may limit broader adoption and acceptance of mechanistic approaches despite their potential utility.

MIDD has a major and increasing role in strategic and regulatory decision‐making: in particular, in recommending and justifying clinical trial design (e.g., dose and dose regimen) for clinical stages of drug development that integrate available data from prior stages of drug development, other trials (of competitors or standard‐of‐care), and mechanistic knowledge. PBPK models have also been used to adjust these trial designs for special populations (e.g., pediatric, chronic kidney disease), predict potential drug–drug interactions in small molecules, and predict drug distribution in distal tissue (e.g., brain, skin, etc.). Identifying combination dosing strategies, responder/nonresponder characteristics, and optimizing complex modalities (e.g., varying infection load, strategies that factor in immunogenicity) are often based on mechanistic models of physiology.

Overall, MIDD applications are often isolated, lacking a holistic or strategic consistency, leading to an opportunity for AI to standardize and integrate technical methods, allowing scientists to concentrate on strategic decision‐making [9]. Additionally, for techniques like PBPK and QSP, GenAI can efficiently summarize evidence from various sources such as PubMed, real‐life data, and competitor trials, enabling scientists to focus on deriving actionable insights.

Going forward, MIDD will continue to play a pivotal role in the drug development process. In fact, in a recent definition, the FDA considers one of the three key elements of MIDD as “quantifying information by developing mathematical models based on full use of all available data, from sources such as in vitro (e.g., outside the body), preclinical and clinical studies” [10] (emphasis ours). This is a call to action to improve processes of quantitative knowledge integration using AI approaches to meet the challenge of drug development with increasingly complex modalities and leading the path to personalized medicine.

3. AI in Drug Development

Although AI has always played a major role in early‐stage drug discovery and development, such as biomarker discovery, its role in later stages, such as in “virtual trials,” has become even more pronounced in recent years [5]. Notably, submissions to the FDA/Center for Drug Evaluation and Research (CDER) with a significant AI component have increased manyfold, with over 100 submissions in 2021 [7]. As illustrated in Table S1, AI has an increasingly important role to play in every stage of drug development.

In the discovery phase, ML accelerates target identification and also plays roles in high‐throughput screening and predictive modeling to rapidly shortlist promising drug candidates. Moving to the preclinical development phase, ML enables toxicity prediction and lead optimization; GenAI and DL are yielding interesting molecules, combing through large databases. In the clinical trial phase, ML enhances patient recruitment and stratification by analyzing patient data and health records to identify suitable candidates for trials, ensuring diversity, and optimizing trial outcomes. ML models also find use in optimizing phase 2 clinical trial design and also predict likely responders.

AI methods complement MIDD in drug development and can enhance predictive accuracy in clinical development, aiding in potential personalized treatment. For example, AI‐enabled de‐escalation strategies in patients with breast cancer on the basis of tumor type, stage, biomarker status (Her2 or hormone receptor), and patient preferences [11] can inform new trial designs that are more patient‐centric. Although this is still actively being researched, ML can be used to construct concentration and exposure prediction models for therapeutic drug monitoring and precision dosing. ML frameworks can also assist in screening covariates for inclusion in population PK models [12]. AI tools can assist researchers in making trial design decisions by selecting appropriate criteria, modeling trial performance more accurately, assessing the burden on patients and investigators, and prioritizing potential indications.

Overall, AI contributes to every stage of drug development, enhancing efficiency and reducing costs. Notably, AI techniques have also become potent tools in further enabling mechanistic modeling/QSP and MIDD in general, by supporting various stages of current QSP workflows, with a growing role for large language models (LLMs), tailored for integrating domain‐specific knowledge. In the following section, we further illustrate the synergies in integrating AI and MIDD, discussing multiple illuminating real‐world examples.

4. The Synergy: Integrating MIDD and AI

Although MIDD thrives on building simple (and complex) models stemming from an understanding of the underlying biological mechanisms, AI enables digesting very large and disparate multimodal data to synthesize useful insights [13]. Increasingly, a lot of synergy between the methodologies is emerging, with real‐life impact in various aspects of the drug development process. The following case studies illustrate (as yet unpublished) the real‐life impact of combining AI and MIDD in drug development decision‐making in the clinical stage. Such impact is possible in all stages of drug development, and innovation is ongoing in MIDD and AI/ML independently, as well as the opportunity for synergies. We have captured exemplar publications illustrating the utility of MIDD and AI in various stages of drug development, pointing to specific synergies using examples in Table S1. Figure 1 shows the broad aspects of the possible synergies between MIDD and AI.

FIGURE 1.

Synergy between MIDD and AI. MIDD and AI rely on a variety of data and employ unique methodologies to leverage the data and build models. The curved arrows are labeled with synergies between AI and MIDD, which together improve various aspects of the drug development process as outlined at the bottom.

4.1. Patient Selection

ML algorithms, with their capacity to analyze vast datasets, can identify complex patterns and biomarkers that may not be apparent to human researchers. In trials testing anti‐infective treatments for respiratory diseases, this capability allows for the stratification of patient populations into more homogenous groups; for instance, high‐risk patients in COVID‐19 trials with immune deficiencies or comorbidities may warrant a different dosing strategy or follow‐up, potentially increasing the efficacy and safety of investigational drugs. Meanwhile, MIDD utilizes mathematical models to simulate drug behavior in different populations, providing insights into how individuals with varying genetic, environmental, and lifestyle factors might respond to a treatment. By integrating AI/ML's predictive analytics with MIDD's simulation capabilities, researchers can more accurately predict outcomes for specific patient groups, thereby enhancing the precision of patient selection for clinical trials. This not only improves the chances of clinical trial success but also accelerates the development of personalized medicine, ensuring that patients receive treatments that are most likely to be effective for their specific conditions (e.g., pregnant women diagnosed with COVID‐19).

4.2. Drug Response Prediction and Selection of Optimal Endpoints

The integration of AI with MIDD offers a powerful approach for selecting primary and secondary endpoints in oncology clinical trials. ML/DL algorithms can process and analyze large‐scale oncology datasets, including genomic, proteomic, and clinical data, to uncover biomarkers and disease progression patterns that are critical in determining the most relevant endpoints. For example, in trials of new cancer therapies, ML can identify molecular signatures that predict treatment response, offering predictive markers. A specific application might involve using ML to uncover the bifurcation of ER‐independent and ER‐dependent mechanisms of drug resistance to CDK4/6 inhibitors in the treatment of metastatic breast cancer. MIDD can also be used to simulate disease progression and treatment response across different patient populations, helping to identify secondary endpoints that capture long‐term outcomes and quality‐of‐life improvements. Together, AI and MIDD enable a more nuanced and evidence‐based selection of endpoints that are directly relevant to patient outcomes. For example, in breast cancer trials, this synergy could lead to the adoption of progression‐free survival as a primary endpoint and patient‐reported outcomes as secondary endpoints, ensuring that the clinical trial design is both scientifically robust and aligned with patient‐centric goals.

4.3. Prioritizing Drug Combinations

The synergy between the two modalities can be instrumental in identifying promising drug combinations for clinical trials, particularly in complex diseases like ulcerative colitis. ML/DL can analyze vast datasets from preclinical and clinical studies, patient records, and genomic information to uncover potential synergies between drugs. For example, it might identify that a certain anti‐inflammatory drug, when combined with a specific immunomodulator, shows enhanced efficacy in reducing inflammation in ulcerative colitis models. Meanwhile, MIDD can simulate the pharmacokinetics and pharmacodynamics (PK/PD) of these drug combinations in virtual patient populations, predicting their safety and efficacy profiles before actual clinical trials. This approach allows researchers to prioritize the most promising combinations for clinical evaluation. An illustrative case could involve combining a novel JAK inhibitor with a well‐established TNF inhibitor [14], where ML analyses suggest a complementary mechanism of action that could offer superior control of inflammation without increasing adverse effects, a hypothesis that MIDD simulations support by showing favorable outcomes in simulated patient cohorts. This synergy not only streamlines the drug development process but also opens new avenues for personalized treatment strategies in ulcerative colitis.

5. Challenges

The FDA recognizes the pivotal role of MIDD in quantitative knowledge integration to steer clear of isolated decision‐making paradigms. Indeed, the agency has underscored this as a cornerstone of MIDD, emphasizing the need to integrate information through the development of mathematical models that harness the entirety of available data. Such data, spanning from in vitro and preclinical studies to clinical trials, lead to a diversity of insights crucial for informed decision‐making processes [10]. However, the current practice of MIDD often grapples with subjectivity and incompleteness in data integration. Its efficacy is tethered to the expertise of the team involved and the pragmatic constraints in delivering timely outcomes. ML is poised to address these challenges by enhancing the speed and objectivity of data integration [15]. Its capacity to process vast amounts of diverse data types enables a more comprehensive approach to information synthesis. Through GenAI, disparate data from sources like publications, news articles, and blogs can be harmonized into queryable databases, opening new avenues for exploration and analysis. In the near future, we anticipate AI to generate regulatory document drafts (e.g., protocols, statistical analysis plans, and clinical study reports), integrating dynamic links to source tables, biomarkers, PK/PD data, patient narratives, and target product profiles. Prompt engineering, retrieval‐augmented generation, and fine‐tuning of open‐source or proprietary LLMs can be strategically deployed to enhance the efficiency of drug development by improving the accuracy and relevance of AI‐generated predictions, thereby speeding up discovery processes and optimizing clinical trials. Yet, amidst these promises, ensuring the factual accuracy of output and addressing bias from training data pose critical concerns. These are critical issues in the drug development domain, where precision is paramount.

Without evaluating the pertinence and provenance of data to the specific query at hand, there is a risk of integrating irrelevant or misleading information. Moreover, ML models are susceptible to overfitting, necessitating measures to ensure their generalizability and the judicious selection of pertinent data sources. The “black‐box” nature inherent in many ML models further compounds these challenges, raising questions about the potential introduction of biases, particularly in sensitive domains like health care. Given the inherent biases in primary healthcare data collection and evaluation processes, this opacity exacerbates concerns regarding equity and fairness in decision‐making.

Although admirable efforts are being invested into making AI models more explainable (explainable AI, XAI), by leveraging neural network attention mechanisms [16] and counterfactual explanations [17], they may still lack the required rationale for decision‐making in critical scenarios. XAI is increasingly important in this context, especially given regulatory agencies such as the US FDA emphasize transparency in AI‐driven decision‐making. AI models such as deep neural networks have been routinely criticized for their “black‐box” nature. XAI may clarify how specific biomarkers or parameters influenced a patient's response to a drug, thereby enabling more precise patient stratification or precision dosing. XAI can increase confidence in AI models by providing interpretable justifications for models used in trial design or patient selection. Some promising advances are already being seen in explainability techniques, such as decision path visualizations, LIME (Local Interpretable Model‐agnostic Explanations) or SHAP (SHapley Additive exPlanations), which are laying the foundations for providing clearer rationales for AI‐based decisions in healthcare contexts [18]. Key challenges in XAI for drug development stem from a few key reasons. First, the AI models used are typically way more complex than the mechanistic models used in QSP/MIDD. Second, the input data that go into training the models are themselves high‐dimensional and noisy, making it challenging to interpret model predictions. Finally, XAI methods are still in development and are vulnerable to manipulation and adversarial attacks; addressing these vulnerabilities will be crucial to ensure their reliability for delivering impactful real‐world results [19]. To date, there are also no clear standards/paradigms for explaining AI models in drug development. Future advances in XAI will be critical to building trust in AI‐driven drug development and accelerating personalized medicine. Nonetheless, it is worth noting that transparency [20] may sometimes be more valuable to regulators than complete explainability, especially when dealing with high‐dimensional data such as genomics.

The integration of these approaches remains a complex and evolving endeavor, contingent upon the intricacies of each research question. Recent frameworks, such as theory‐driven ML, offer promising avenues by prioritizing causality through the integration of prior knowledge with big data. The practical implementation of such frameworks demands customization and iterative refinement to navigate the nuances of diverse research contexts effectively [21]. As the landscape of knowledge integration continues to evolve, interdisciplinary collaborations and innovative methodologies will be pivotal in harnessing the full potential of MIDD and ML to drive transformative advances in regulatory decision‐making and beyond.

6. Outlook

Historic advances in computational methods and technologies, especially in AI, are profoundly reshaping drug development processes. DL, using models such as AlphaFold3 [22] is ushering in an era of even more accurate protein–ligand binding predictions and promises to impact drug development enormously. LLMs are already beginning to have an impact on the study of protein structure and function [23], a fundamental step in drug discovery, or even in the summarization of clinical reports [24]. Nevertheless, key challenges inherent in such large complex models, from interpretability and explainability to unique issues such as hallucinations, will need to be addressed; responsible AI has increasingly become a guiding principle in the development and deployment of modern solutions in such critical areas.

Although modeling is not new to the regulatory landscape—QSP has been employed extensively [25] by regulatory authorities such as the FDA for evaluating new drugs and their dosing regimens—the integration of ML models will definitely bring new concerns about interpretability, validity, and reliability. Regulatory agencies are already fast embracing these technologies, and this is expected to be an area of significant development in the coming years.

With the expanding possibilities of synergy between MIDD and AI, as outlined in Figure 1 and Table S1, the coming years will require training a new generation of interdisciplinary scientists. These scientists should be proficient in specific domains and also possess a working knowledge beyond their domain: for example, MIDD, with an understanding of pharmacometrics and physiology, and also AI, imagining and understanding complex high‐dimensional spaces [26] where data may be represented, feature engineering, and inherent biases in ML algorithms. Most importantly, scientists should be trained to thrive in collaborative scientific environments, with others possessing complementary expertise and speaking different scientific languages, working toward common drug development goals.

Leveraging MIDD for mechanistic understanding and data integration alongside AI/ML for data analysis and prediction can create a powerful synergy for accelerating drug development. Overall, the potential of MIDD and AI in drug development is huge and transformative; by effectively harnessing the complementary strengths of MIDD and AI, we are poised to fulfill the immense promise of modern drug development, producing efficient, personalized therapies that can greatly improve patient outcomes and ultimately advance human health.

Conflicts of Interest

K.R. is a director of qBiome Research, Chennai. R.K. is a director of Vantage Research, USA, and its subsidiary Vantage Research Pvt. Ltd., Chennai. C.J.M. and S.M. are employees and shareholders of Pfizer Inc.

Supporting information

Appendix S1