Artificial Intelligence in <i>Clinical and Translational Science</i> : From Bench Insights to Bedside Impact

Abstract

The healthcare and life sciences sectors are experiencing a transformative moment that cannot be overlooked. Our biological understanding, technology, and data are coalescing to leverage unprecedented opportunities for innovation [1, 2]. At the center of this transformation lies Artificial Intelligence (AI) and Machine Learning (ML), which have advanced from speculation to working technologies that can make actual differences in patient care and drug development [3, 4]. Only a few years ago, AI was framed in the context of its potential in clinical pharmacology, drug discovery, and development. Following the 2024 Nobel Prize in Chemistry, which was awarded for the AI-based prediction of protein structure, it became increasingly difficult to ignore the scientific merit of the technology [5]. We are now experiencing implementations that are changing how we approach these disciplines [6-8]. We have transcended previous discussions about whether AI will help and are asking more nuanced questions about how we deploy these technologies in a responsible manner, such that they deliver reliable and reproducible results, and produce meaningful value in clinical and translational research. The AI-themed issue in Clinical and Translational Science addresses exactly these questions by collating different viewpoints from around the field and providing substantive evidence to demonstrate the breadth of the current AI applications that are transforming the practice of clinical and translational sciences (Table 1). From the earliest phases of drug discovery through design and optimization of clinical trials, from developing personalized treatment approaches to monitoring drug safety postmarket approval, and to collecting real-world evidence—these contributions illustrate the current state of the art in utilizing AI in drug discovery and development and also characterize our current capabilities, while providing a vision for future innovation within clinical and translational sciences. In this piece, we aggregate contributions from over 30 manuscripts in this special issue, organized into three general categories (Figure 1, Table S1): Discovery and Preclinical Innovation, Clinical Development and Precision Medicine, and Postmarketing, Safety, and Real-World Implementation. We also discuss the cross-cutting aspects of regulation, ethics, and operations that may ensure AI achieves its potential. As readers progress through this editorial, they will see that the contributions in this issue do more than showcase technical innovation. They reflect a growing maturity in how the clinical pharmacology and translational science communities approach AI, which is not being expressed as a ‘one-size-fits-all’ solution but rather as a toolkit, where the impact of AI will depend on thoughtful integration, appropriate validation, and a deeper understanding of both its strengths and limitations. At the beginning of the drug development continuum—where drug developers identify targets, discover biomarkers, and create preclinical models—AI's intuitive ability to assess and map the meaning within massive, multidimensional datasets is, more often than not, beginning to be leveraged. At the earliest stages of discovery science, AI's transformative property is not necessarily its computational speed; rather, it is AI's ability to spot deep and subtle patterns that statistical approaches have most often failed to recognize, especially nonlinear relationships. This special issue includes many contributions that demonstrate how AI can now enable broader hypothesis generation, more rapid compound optimization, and more biologically founded predictions. One area that is rapidly progressing is in silico pharmacokinetic (PK) and pharmacodynamic (PD) modeling. Walter et al., make a systematic examination of ML-based approaches to empirical, compartmental, and physiologically based pharmacokinetic (PBPK) models for predicting plasma PK profiles in rats [9]. In a study with over 1000 small molecules, the authors conclude that mechanistic/compartmental PK models with the ML model, and direct ML-based profile prediction both yield similar performance, while they all perform better than simple noncompartmental approaches (e.g., NCA-based). Importantly, ML-based approaches allow compounds to be prioritized and triaged earlier in the discovery pipeline based on predicted PK properties prior to synthesis or animal study, which speeds lead selection and reduces superfluous in vivo studies. Likewise, Pillai et al., introduce a pipeline for predicting plasma PK profiles in preclinical species (rats) from chemical structure with a chain of ML models [12]. Initially, their method uses molecular descriptors and ML to approximately predict in vivo clearance and volume of distribution, and uses the results to then train a second ML model to estimate the full plasma concentration-time curve. Their platform is fairly accurate, especially for similar compounds to the training dataset (Tanimoto value > 0.5), with mean absolute percentage errors of approximately 150% most often, and it allows the investigator to have PK data even before any experimental data at all, which will further extend the role of AI in virtual screening and compound selection. These examples imply that AI can accelerate knowledge accumulation early in the input stage of discovery, so when questions arise during drug screening, there is much less to lose in terms of financial failure or opportunity to pivot. In addition to these interesting research examples, the article by Zhang et al. represents an example of AI-driven target identification by using a pipeline to develop a drug response prediction model of neuroendocrine prostate cancer [30]. This pipeline includes transcriptomics, high-throughput drug screen (HTDS), and causal feature selection methods to derive new actionable targets—in this example identifying nicotinamide phosphoribosyltransferase (NAMPT) as a new target. Their research also presents experimental data with enhanced efficacy of NAMPT inhibitors, in NEPC models, and identifies potential predictive biomarker signatures for clinical translation. This study illustrates nicely how AI via omics data integration and in vitro validation can streamline target identification and further precision oncology approaches even as AI comes into the clinic. AI's ability to empower phenotype-driven discovery is exemplified by Basile et al. [36], who developed and tested a rules-based, machine-driven approach for expedited and high-fidelity identification and phenotyping of patients with nonalcoholic fatty liver disease (NAFLD) across multiple health systems. The application of large-scale EHR data in their work yielded high positive predictive value, and high levels of sensitivity toward NAFLD case-finding and risk stratification, illustrating real-world translational viability of ML for discovery directly in patient populations. Furthermore, the algorithm also permits enrollment into focus interventions while sustaining risk stratification relevant toward experimental or therapeutic development, reflecting the emergence of insight from population-level EHR mining as much as laboratory molecular assays. From a generative modeling standpoint, Titar and Ramanathan present variational autoencoder (VAE)–based generative modeling in order to simulate populations with renal, hepatic, metabolic, and cardiac disease with respect to multidimensional physiological biomarkers that are directly applicable for dose decision-making [11]. Their work shows that tabular VAE models can sufficiently approximate distributions of continuous physiological determinants of drug dosing; however, performance for categorical, and particularly rare disease states, is constrained. This application of generative AI fosters simulation-based hypothesis generation and clinical trial design focused on real-world interindividual variability, without reliance on either costly or hard-to-obtain data. In connecting mechanistic and data-driven paradigms, Elmokadem et al. propose, and demonstrate, a research workflow for hierarchical deep compartment modeling (HDCM) [28]. In the proposed framework, neural networks are embedded in compartmental PK model structure and uncertainty and variability are handled with Bayesian inference. The HDCM framework allows for scalable modeling of population-level trends and individual-level heterogeneity; consequently, the method includes utility (like SHAP analysis) for making predictions based on model input (covariates visible in the model). Notably, this hybrid modeling permits modeling beyond previous limitations, where ML generally provides no estimates of uncertainty nor mechanistic extrapolation capability, so the method should allow broader applicability for both simulation and interpretation in preclinical pharmacology. Collectively, all these studies indicate that AI in the early stages of discovery has progressed well past retrospective pattern mining and is focused on predictive and generative modeling, which contribute directly to experimental methods and aid resource allocation, and ultimately embedding enhanced efficiency, elasticity, and patient pertinence for downstream clinical development. Once you have a candidate molecule, the center of gravity shifts from bench to clinic, and uncertainty is driven by human heterogeneity, not chemistry. Clinical development, which can represent the most resource-demanding phase of drug development, has historically been characterized by long timelines, rigid protocols, and minimal flexibility once the trial is initiated. Artificial intelligence is questioning this framework by shifting to a more spontaneous and data-driven approach to the design of trials, patient selection, and personalization of treatment. In each of the studies in this special issue, AI was employed in three interrelated ways: optimizing trial design and execution, refining patient stratification and predictive modeling, and personalizing treatment and dose optimization. The fast-paced growth of AI in trial design and operational execution was a salient topic across the discussions during the ASCPT 2024 AI Preconference, and evidenced in the summary by Shahin et al. [8] Their summary illustrates how AI is providing opportunities for innovations like generative models for generating study protocols, digital health technologies used for endpoint surveillance, and the use of explainable AI in dose-finding and covariate analysis. In contrast to the summary focusing on adaptive randomization and specifications of protocol simulation, the summary paints a broad view of AI-enabled innovations that are now underway in modern drug development. Shahin et al. highlight the importance of multidisciplinary partnerships as AI becomes implemented in regulatory submissions and operational decision-making. They emphasize that engaging specialists from statisticians and social scientists to regulators is critical, and they insist on consideration for scientific rigor and methodological transparency when embracing new computational tools. Continuing from the operational utility of AI, Huh et al. investigated using Meta Llama 3, an open-source large language model, for automating the systematic identification and characterization of clinical trials with decentralized components from registry databases [18]. Instead of offering a mechanism for regulatory review and approval of the protocol, this study demonstrated how large language models could systematically sort and organize trials by decentralized components based on the free text description of the trials. They conclude that models like these have considerable value for trend analytics and extracting structured data from unstructured data sources, but they also warn that it is sensitive to the models the user gets transparency and accountability outputs. Their work is best understood as a proof-of-concept for the use of AI to support the operational review of decentralized clinical trials, and shortcomings remain because of changing registries' terminology and scope. Regarding data operations transformation, Podichetty et al. outline the emerging end-to-end analytic pipeline work at Critical Path Institute (C-Path) within the Amazon Web Services (AWS) cloud [34]. Their example is less a case study by itself, but, more of an overview for ongoing transformation where traditional, static data pipelines are replaced by AI-augmented, cloud-native infrastructures. They characterize these future platforms as scalable, interoperable infrastructures that enable rapid assembly, validation and surveillance of real-world data for regulatory-grade evidence generation. Near-real-time data movement and data synthesis are the current state, as they note, and this is part of the iterative and collaborative nature of trial data management in modern R&D. Stepping into areas of patient stratification and predictive modeling, the papers of this special issue highlight the power of AI to work with high-dimensional clinical and biological data to provide risk estimates and tailor precision medicine. Zhang et al. explain the usage of ML to develop and validate a nomogram which utilized lymphocyte subtyping and significantly increased predictive precision for intra-abdominal candidiasis in sepsis patients [23]. Ambe et al. leverage interpretable ML models utilized on electronic medical records to predict acute kidney injury from cisplatin-based chemotherapy highlighting that model transparency and early predictions were critical for practical use, starting before chemotherapy [24]. Lan et al. introduce MoLPre (Machine Learning Model for Lung Cancer Metastasis Prediction), a ML-based platform to predict metastasis in early-stage lung cancer creating new possibilities in personalized therapy approaches during the surgical decision [16]. Amato et al. utilize explainable AI to analyze data from tepotinib trials, leading to the idea that safety signals—like edema—could be included for precision treatments not just in postmarketing but during clinical developments [31]. In agreement with those advancements, Khozin's “From Organs to Algorithms” proposes transforming oncology classification frameworks from anatomical overclassification to an AI-enabled, data-driven, molecular-based stratification [32]. He offers a vision where the future of oncology will be increasingly shaped by multiomics inputs and computational disease are and patients are for As a these papers a from static risk and to and that leverage They that do the most and be the most when both the and data are and to and This to regulatory and evidence across In the AI personalizing treatment is not about identifying the but also about the In this issue, et al. the role of that is, a hybrid of methods and pharmacology models, in precision They the opportunities that clinical decision support utilizing a predictive model, digital or both framed on both AI and models an that can be the changing patient The authors present digital disease management platforms and discuss how hybrid models can decision-making and provide for dose optimization with an consideration for data and which are in clinical care AI et al. this on drug development and AI to development Their the possibilities when mechanistic models and are focusing on how these approaches could streamline patient selection, trial and predict drug They demonstrate how a approach can be implemented through examples early discovery to clinical trials, with on model regulatory and a training will to leverage personalization in medicine. approach to is using digital virtual patients that and data. a computational that or as to the care can provide knowledge in the management of at the clinical et al. and patient of AI-driven digital in clinical trial They digital as AI models of reflecting treatment and their of use in trial optimization are (e.g., virtual covariate The authors to utilize clinical trial to in digital and actionable validation and at the and regulatory as an for particularly through AI in clinical decision et al., a of of enhanced by AI in an electronic health The but are in actionable and those should be of The the that for AI to be there is a of technical rigor and design to clinical these scientific are methodological that provide for AI transparency and et al., deliver a practical to focused on the of in drug development The examples work through examples of how SHAP be used to feature contributions across ML model and a to in et al., provide a on for clinical and translational the and optimization of the modeling algorithm using clinical datasets in The and practical and actionable for particularly statisticians and in to real-world data. et al., a analytics pipeline based on Amazon Web Services Their in this issue that the of digital health scalable data transformation, and relevant examples through the with the generative model of in the issue, this of with platform how to technical with the of advanced analytics and AI for trial dose optimization, and dataset AI has and for drug development, drug discovery and optimizing clinical trial is its not at the approval stage of development. The postmarketing as we from in a to understanding the and of new therapeutic interventions in the and of clinical in a that the of human decision making and the impact of drug use on patient AI-enabled and real-world evidence also allow identification of safety in a and when new are by the healthcare while providing the ability to rare and less that may not have been prior to approval through clinical trial and identification of nuanced patterns of and risk in different patient populations. This issue a of studies that illustrate the ability of AI to postmarket monitoring from a and retrospective to a and The contributions in this issue how ML electronic health records relevant and at can be AI methods to patients with nonalcoholic fatty liver disease (NAFLD) were demonstrated to rapidly with the disease across healthcare by medical laboratory and for a scalable surveillance for identifying in a scalable are not but rather our ability to in study with safety and target therapeutic interventions to who critical of postmarket AI is cloud-native pipelines that allow for the and of a of data over One of the in the issue a method to develop a pipeline based on to and and analyze large-scale clinical datasets for are to enable beyond while opportunities for where and performance are especially rare and subtle a focus of AI in The studies in this issue demonstrate how explainable AI models that used and treatment to predict for patients with provide that the could and [31]. predicting acute kidney injury with data demonstrated not just high but also to the use of transparency of the model to clinical decision-making [24]. The digital and virtual in this issue as a the simulation of the and the personalization of the postmarket The increased of that are for clinical trial applications as in this issue, possibilities for their into real-world monitoring with the ability to simulate treatment to safety and efficacy or at the of for patients with healthcare these EHR cloud-native explainable models, and digital the postmarket phase can into a continuous AI not rather, it the ability to and new in a that is both and more than The use of AI in drug development will and as it through the drug development AI has been to work in it will be to validate and these in a that from all and regulatory are changing The and have on AI being used in medical and drug development, which AI in terms of and in this special issue with these the included discussions on the with data in highlight the importance of validation across of patients to any in performance in in which AI models new data inputs was a of about and being for clinical the AI model may be a so long as it achieves a it is not to regulatory and clinical because we do not it a are on AI like SHAP which are included in this issue, and we this is a to into model development pipelines This will help during the regulatory but also work in a that dose selection, prediction and and remain from an AI that are on datasets can or health care in this special issue that of data generation with variational presents methods of datasets for and for the These approaches are to ensure AI-based do not populations. of be AI models are in an because of the embedding those models into clinical and research The of AI-based applications in clinical a data framework, be to with current health and the user at the center to ensure the AI is and actionable to the The case studies in this special into a clinical workflow and platforms for and trials that technical not but around workflow to meaningful the potential to with multiple is AI is a development in clinical pharmacology and translational science, and at the of computational science, clinical regulatory and for patient and in this special issue illustrate opportunities for partnerships and regulators and technology not just to methodological development but also to to validation, and to The manuscripts in this special issue an from molecular hypothesis to real-world AI-enabled and hybrid biological and chemical modeling are a to digital trial at the and of safety at a all by an AI that scientific rigor with operational these manuscripts AI not as a to the of clinical pharmacology, but rather as the of a pharmacology. in order to full validation, and be in and This the of the authors and should not be to represent the or The authors no of The is not responsible for the or of any by the than should be to the for the

Ähnliche Arbeiten