Artificial intelligence (AI) is moving from promise to routine practice in clinical and translational research, and large language models (LLMs) are now used to scope analyses, draft and debug code, and accelerate the quantitative work that supports clinical and translational decisions. This adoption is outpacing the operational safeguards that keep such work reproducible and auditable, and the central risk is not only erroneous code but the quieter erosion of the analytical discipline on which trustworthy bench-to-bedside decisions depend. In this position paper, we argue that AI can be integrated into regulated quantitative workflows without sacrificing reproducibility, provided it operates inside a compliance-by-design architecture built on two principles: analytical intent must be formally encoded and locked before execution, and the analysis itself must run outside the LLM as specification-verified code in a validated environment. We describe three failure modes that ordinary quality control does not detect, namely context drift, the displacement effect, and implementation drift, and we link each to documented model behavior. For open-ended investigation, these principles are operationalized as a four-step execution firewall; for recurring development studies such as assay validations and animal safety studies, they are operationalized through a governed library of documented, validated agents, supported by fit-for-purpose validation, a clear separation of human and machine roles, and post-deployment monitoring. Both approaches can be implemented with components available today. Used this way, AI enhances rather than replaces quantitative expertise, which now resides less in the code a model produces than in the constraints a scientist defines before it runs.

Elunetirom (ABX-002) is a brain-targeted, oral prodrug being developed for the adjunctive treatment of major depressive disorder and bipolar depression. The active moiety, LL-340001, is a thyroid hormone receptor agonist with modest selectivity for beta over alpha thyroid hormone receptors. Upon entering the brain, the intended site of action, elunetirom is hydrolyzed by fatty acid amide hydrolase to LL-340001. In this 3-way randomized, crossover, bioequivalence study, the pharmacokinetics (PK) and safety of an oral liquid formulation used in early clinical development were compared with those of an oral solid formulation intended for later-stage development and potential commercialization, under fasted and fed conditions. Healthy adult participants received single doses of elunetirom (45 μg) as a liquid formulation under fasted conditions and as a solid formulation under fasted conditions and 30 min after consumption of a high-fat meal. Administration of the solid formulation under fasted conditions resulted in elunetirom and LL-340001 exposures similar to those observed with the liquid formulation and is therefore considered to have comparative bioavailability. Food intake increased elunetirom total exposure (AUC) compared with the fasted state but did not affect maximum exposure (Cmax). In contrast, administration with food reduced LL-340001 Cmax compared with the fasted state but did not affect AUC. Food intake delayed the time to reach Cmax (Tmax) for both elunetirom and LL-340001. Elunetirom, administered orally as a liquid and as a solid formulation, was safe and well tolerated under fed and fasted conditions.

Fatty acid oxidation (FAO) is a metabolic vulnerability in cancer cells, making pharmacologic FAO inhibition a promising therapeutic strategy. This Phase I, open-label, 3 + 3 dose-escalation study evaluated the safety, tolerability, pharmacokinetics, and preliminary antitumor activity of two oral FAO inhibitors-omeprazole (KN510), a carnitine-acylcarnitine translocase inhibitor, and trimetazidine (KN713), an acetyl-CoA acyltransferase inhibitor-given once daily in combination to patients with advanced solid tumors. Twelve patients were enrolled across four dose cohorts (omeprazole and trimetazidine, each 60-120 mg/day). The maximum tolerated dose was not reached; the recommended Phase 2 dose was set at 120 mg once daily for each agent. Treatment was generally well tolerated, with most adverse events Grade 1-2; the most common were flu-like symptoms (25.0%), constipation (16.7%), nausea (16.7%), and dizziness (16.7%). Omeprazole showed time-dependent pharmacokinetics consistent with auto-inhibition of its own metabolism, whereas trimetazidine showed dose-proportional pharmacokinetics across the studied range. The best response was stable disease in five patients (disease control rate 41.7%). Median progression-free survival was 1.66 months (95% CI, 1.18-3.55) and median overall survival was 9.13 months (95% CI, 3.29-not reached). Dual pharmacologic FAO inhibition with oral omeprazole and trimetazidine was feasible and well tolerated, supporting the mechanistic rationale for combination dosing. Disease stabilization observed in this heavily pretreated population warrants further clinical evaluation, particularly in combination with standard chemotherapy regimens. Trial Registration: ClinicalTrials.gov, NCT06012708; registered August 25, 2023.

Digital health technologies can support remote acquisition of clinical investigation data and may facilitate more flexible clinical trial designs. In March 2023, FDA revised Forms FDA 1571 and FDA 356h to include a structured yes/no field asking whether a submission contains DHT data or a proposal to collect DHT data. This change created a mechanism for identifying submissions involving DHT-based remote data acquisition. Following implementation of these fields, the Office of Medical Policy in FDA's Center for Drug Evaluation and Research developed and implemented a systematic process to track, identify, and adjudicate DHT-related submissions. This analytical report presents a descriptive, aggregated, de-identified analysis of submissions that were flagged through this operational process from March 2023 through December 2025. Operational inclusion criteria were applied to identify submissions involving DHTs used for remote data acquisition intended to support regulatory decision-making. Of 2318 flagged submissions, 277 (12%) were confirmed for inclusion in this analysis. Confirmed submissions were predominantly associated with investigational applications (237/277, 86%). When classified by the nature of data collected, submissions reflected a range of remote data acquisition approaches, including patient reported data and sensor-based measurements, with some submissions combining more than one modality. Electronic diary and electronic patient-reported outcomes were most common, exclusively appearing in 161 submissions (58%), followed by data captured from general objective sensors in 92 submissions (33%) and continuous glucose monitoring data in 24 submissions (9%). These findings establish a quantitative baseline for DHT-based remote data acquisition in CDER submissions and the implementation of FDA's corresponding guidance.

MK-1942 is a negative allosteric modulator of the metabotropic glutamate receptor 2 that was investigated for the treatment of Alzheimer's disease and depression. Nonclinical studies with recombinant human CYPs and genotyped human hepatocytes suggested that CYP2C19 is involved in the metabolism of MK-1942. Given the high incidence of different CYP2C19 polymorphisms in the Japanese population, a Phase 1 single and multiple dose study in healthy Japanese participants provided an opportunity to determine the involvement of CYP2C19 phenotype on the clinical PK of MK-1942. Also, since a first-in-human study in non-Japanese showed a decrease in exposures after administration with high-fat meal, we further tested an effect of low-fat meal on PK in the Japan study. The study was designed with flexibility which enabled dose selection for the multiple dose panels, based on the totality of safety and PK data. Oral administration of single and multiple doses of MK-1942 in Japanese participants was generally well tolerated. AUC0-inf was comparable between fasted and fed (low-fat meal) state. Following a single dose of MK-1942, AUC0-inf was ~70% increased in CYP2C19 poor metabolizer (PM) vs. extensive metabolizer (EM). Therefore, a decision was made to test the multiple dose separately for each CYP2C19 phenotype, rather than grouping them into a single panel. As opposed to the expectation from the single dose result, steady state exposures were similar among CYP2C19 phenotypes. These data supported the subsequent global Phase 2 study to permit a low-fat meal around dosing and no dose adjustment for any CYP2C19 phenotype.

Brain metastases (BMs) represent a major clinical challenge due to their high incidence in cancers such as non-small cell lung cancer (NSCLC) and breast cancer, and the limited efficacy of systemic therapies resulting from inadequate drug penetration across the blood-brain barrier. Although several brain-penetrant tyrosine kinase inhibitors (TKIs) have been developed, the quantitative relationship between intracranial drug exposure and intracranial clinical response has not been well established, leaving an important gap in dose selection strategies. In this study, we systematically evaluated the intracranial exposure-response relationship across brain-penetrant small-molecule TKIs. Fourteen TKIs were identified based on guideline-recommended therapies for BMs. For each drug, plasma and cerebrospinal fluid (CSF) pharmacokinetic profiles were simulated using published population pharmacokinetic models, with CSF exposures extrapolated using known unbound plasma fractions and CSF-to-plasma unbound concentration ratios (Kp,uu,CSF) in humans. Key intracranial exposure metrics, including the CSF steady-state trough and average concentrations relative to IC50 (Cmin,ss/IC50 and Cave,ss/IC50) and the percentage of the steady-state dosing interval during which CSF concentrations remained above IC50 (T% > IC50), were derived and correlated with intracranial outcomes from published clinical studies. Overall, significant positive correlations were observed between intracranial exposure metrics and both intracranial objective response rate (ORR) and complete response rate (CR). Greater exposure coverage above IC50 was associated with improved intracranial efficacy. This work demonstrates a positive intracranial exposure-response relationship across brain-penetrant TKIs and provides a quantitative benchmark to inform dose selection beyond traditional plasma exposure metrics. These findings highlight the importance of achieving sufficient CNS drug concentrations to optimize the therapeutic effect in patients with BMs.

The Clinical Pharmacogenetics Implementation Consortium (CPIC) bases its drug-gene recommendations on the assignment of star alleles, which map known genotypes to defined functional categories and corresponding drug dosage guidelines. The star allele framework, first proposed in 1996 for the CYP gene family and later formalized with CPIC's establishment in 2010, remains foundational to pharmacogenomics. However, this system has notable limitations. Its dependence on a restricted set of benchmark single nucleotide polymorphisms (SNPs) excludes rare or novel pathogenic variants that can invalidate a star allele call and lead to incorrect dosing recommendations. Furthermore, nearby non-pathogenic variants can interfere with haplotype interpretation, introducing additional risk of misclassification. To explore how these gaps might be addressed at the analytical layer, we describe PHARMWATCH, a proposed multilayer pharmacogenomics framework that couples standard star-allele assignment with two additional analyses: (1) germline variant screening using the ACMG-based BIAS-2015 classifier and (2) variant interpretation in context (VIIC), which reconstructs the observed amino acid sequence at each star-allele-defining locus for comparison against the sequence predicted by the assigned star allele (3). We describe the framework's architecture, specify the flag thresholds and translation rules used by each layer, and illustrate its behavior on a documented DPYD case in which a rare pathogenic variant fell outside the tag-SNP panel used for pre-treatment genotyping. We present PHARMWATCH as an exploratory framework; cohort-scale performance characterization and extension to additional SNP-driven pharmacogenes are identified as priority future work.

Controlled ovarian hyperstimulation (COH) is a precision medicine challenge, with substantial inter-individual variability in ovarian response affecting treatment efficacy and safety during in vitro fertilization (IVF). Clinical decision support (CDS) tools incorporating statistical, machine learning, and artificial intelligence approaches have been developed to individualize gonadotropin dosing and ovulation trigger timing. Therefore, this study aims to identify and evaluate CDS tools informing prescribing decisions during COH. Following PRISMA-ScR guidance, Embase, MEDLINE, PubMed, and Scopus were searched from inception to March 2026. Studies evaluating CDS tools informing COH dosing and/or ovulation trigger timing during IVF were included, stratified by predictive domain and analyzed descriptively. Primary outcomes were oocyte yield and metaphase II (MII) oocyte yield. Secondary outcomes included gonadotropin use and pregnancy-related outcomes. Of 2879 records identified, five studies (n = 16,441) were included. One study reported precision-dosing CDS tools that were associated with increased MII oocyte yield (1.5 oocytes, p < 0.001) and reduced gonadotropin use (195 IU, p < 0.001) compared with clinician-directed dosing. A trigger-timing model improved MII oocyte yield (2.5 oocytes, p < 0.01), whereas combined dosing-and-trigger models demonstrated inconsistent effects. K-nearest neighbor (n = 3) and linear regression (n = 3) approaches predominated. No pregnancy-related outcomes were reported. No clinically evaluated mechanistic pharmacokinetic/pharmacodynamic models were identified. CDS tools show promise for individualizing COH and improving treatment efficiency through precision dosing. However, clinical translation remains limited by a small and heterogeneous evidence base, reliance on surrogate outcomes, and lack of mechanistic model-informed approaches. Prospective studies evaluating clinically meaningful outcomes are required to establish the role of CDS tools in personalized reproductive medicine.

Machine learning (ML) is increasingly being used in clinical medicine. The objective of this study was to identify factors predictive of hospitalization, with the support of ML, in outpatients with mild-to-moderate COVID-19. Unvaccinated adults with mild-to-moderate COVID-19 were randomized 1:1 to molnupiravir or placebo in MOVe-OUT (NCT04575597). In this post hoc analysis, various factors were evaluated among participants who progressed or did not progress to hospitalization. ML was used as an unbiased method to evaluate multiple factors for their association with hospitalization in placebo-treated participants. Participants who progressed to hospitalization (HP) were more likely versus those who remained nonhospitalized (nonHP) to be male; age ≥ 65 years old; have active cancer, diabetes mellitus, moderate COVID-19, and negative nucleocapsid and neutralizing antibody status at baseline (study intervention Day 1). In HP, baseline mean viral load (VL) was 7.51 vs. 6.83 log10 copies/mL among participants who remained nonHP. Baseline mean CRP was 49.93 in HP vs. 16.77 mg/L in nonHP; postbaseline values increased in HP vs. nonHP. Baseline mean ALC was 1.07 vs. 1.54 × 10[9]/L, respectively; postbaseline values decreased on Days 3 and 5 in HP. ML analysis identified baseline VL, CRP, ALC, calcium, and mean change in CRP, ALC, and LDH as predictors of hospitalization. The ML model supported the initial findings that a higher VL at baseline with an increasing CRP and decreasing ALC may help identify COVID-19 patients at high risk for hospitalization. ML could be utilized to identify biomarkers predictive of hospitalization in future pandemics.

Quantitative prediction of drug removal by hemodialysis is essential for rational dose design in patients with renal failure. We aimed to develop a machine learning-based quantitative structure-pharmacokinetic relationship (QSPKR) model to predict hemodialysis clearance (CLHD) using molecular descriptors and dialysis conditions. CLHD data were compiled from published clinical studies, yielding a dataset comprising 46 compounds; 457 measurements; 7 dialysis membrane types; and dialysis parameters, including blood flow rate (QB) and dialysate flow rate (QD). Recursive feature elimination revealed six key predictors, which were used to construct and compare five models: lasso regression, random forest, support vector machine, averaged neural network, and extreme gradient boosting (XGBoost) models. Feature contributions were interpreted using Shapley additive explanations (SHAP) analysis. Among the models, XGBoost demonstrated the highest predictive performance (root mean square error = 51.71 mL/min; R[2] = 0.614). The SHAP analysis revealed that lipophilicity was the most influential predictor. The effects of QB and QD on CLHD were nonlinear and saturable, consistent with the established dialysis clearance theory. CLHD can be predicted from molecular structure and dialysis parameters using a QSPKR-based machine learning approach, providing a practical framework for estimating drug clearance and assisting in the preliminary evaluation of drug dialyzability when empirical data are limited.

Hereditary angioedema (HAE) attacks are recurrent, debilitating, and potentially life-threatening. Across the clinical program, garadacimab long-term prophylaxis demonstrated early onset and durable, sustained protection against HAE attacks and clinically meaningful improvements in health-related quality of life (HRQoL). HAE treatment goals include total disease control (attack-free) and normalization of life. Predicting treatment effect may aid shared decision-making and choice of long-term prophylaxis. Here, we report an attack-free probability model predicting garadacimab treatment effect beyond available Phase 3 clinical data and explore the relationship between HRQoL and attack-free duration. A generalized mixed-effect Poisson model was developed and validated with data from the garadacimab pivotal Phase 3 (VANGUARD) and Phase 3 open-label extension studies. Efficacy variables, including attack-free probabilities, were predicted for simulated populations of garadacimab-treated patients and assessed for goodness of fit. Analysis of covariance was used to assess the relationship between Angioedema Quality of Life questionnaire scores and attack-free duration. Model-predicted efficacy variables adequately reflected observed clinical data with high goodness of fit (Nakagawa-Schielzeth conditional R[2]: 74.3%). Following a 1-month attack-free period after garadacimab initiation, patients had a 60% cumulative probability of remaining attack-free up to Month 25 and 49% up to 5 years. HRQoL improved with increasing duration of attack-free periods, reaching minimal to no impairment after 6 months. This attack-free probability model suggests long-term effects of garadacimab over a 5-year treatment period. Patients treated with garadacimab have the potential to live attack-free for long periods of time, with substantial HRQoL improvements.

Cabozantinib is a small-molecule tyrosine kinase inhibitor approved for the treatment of various locally advanced or metastatic solid tumors. The phase 1b COSMIC-021 study evaluated cabozantinib as a monotherapy or in combination with atezolizumab in patients with advanced solid tumors, including castration-resistant prostate cancer (CRPC). Relative to previous studies, the current analysis was conceived to improve the estimation of population pharmacokinetic (PopPK) parameters, including covariate effects, given the increased number of patients by including COSMIC-021 data, and to support exposure-response analyses in CRPC. A nonlinear mixed effects modeling approach was utilized and included evaluation of covariates, including demographics, laboratory values, hepatic function, tumor type, and combination treatment with atezolizumab. The PopPK analyses included 7513 cabozantinib PK samples from 2757 patients and healthy volunteers across 8 studies, including 2164 samples from 819 patients in COSMIC-021. Cabozantinib PK was described by a 2-compartment model with linear elimination and dual-phase absorption. Apparent clearance of cabozantinib (CL/F) was 2.06 L/h, and apparent central volume of distribution was 103 L. Elimination half-life was 104 h. Combination treatment with atezolizumab did not impact cabozantinib PK. Sex and body weight had a minor impact on cabozantinib exposures. There was no significant impact of other covariates on cabozantinib PK. In conclusion, the 2-compartment model with dual-phase absorption described cabozantinib PK well, and the parameter estimates and identified covariates were consistent with previous analyses. The impact of covariates, including sex and body weight, on cabozantinib plasma exposures was not clinically relevant to warrant dosage adjustments.

Sjögren's disease (SD) is an autoimmune condition characterized by chronic inflammation and impaired secretory function of the salivary glands (SG). The proresolving lipid mediator AT-RvD1 restores SG function in SD-like mice; however, its tissue pharmacokinetics, exposure-response relationships, and human dose predictions remain poorly understood. Therefore, this exploratory study characterizes AT-RvD1 tissue pharmacokinetics, develops an SG-focused physiologically based pharmacokinetic (PBPK) model, explores the relationship between SG exposure and secretory response, and predicts human exposures relevant to clinical development. Following intravenous administration of AT-RvD1 to SD-like mice, concentrations were quantified in serum, submandibular gland (SMG), heart, lung, liver, and kidney. These data were used to evaluate a PBPK model incorporating the SG as the target site of action. PBPK-predicted cumulative SMG exposure was subsequently compared with the saliva flow rate obtained after chronic AT-RvD1 treatment. The mouse model was then used to predict systemic and SG exposure in a virtual human population across an exploratory dose range. The PBPK model described systemic and tissue pharmacokinetics, with average fold error and absolute average fold error values within predefined acceptance criteria. Human PBPK simulations predicted dose-dependent increases in SG exposure and identified 1.5 mg/kg as a candidate dose at which predicted concentrations exceeded the preclinical pharmacological reference concentration in more than 90% of simulated subjects. Collectively, these analyses provide an experimentally informed PBPK framework for exploring AT-RvD1 tissue exposure, exposure-response relationships, and candidate doses for future clinical evaluation in SD.

Protocol deviations threaten data integrity, regulatory compliance, and participant safety in clinical trials. Despite their prevalence, few longitudinal studies have characterized deviation patterns and described quality improvement strategies within clinical research units (CRUs). This descriptive longitudinal quality improvement study analyzed protocol deviations and prevented incidents reported at the University of Iowa CRU from June 17, 2020, through December 31, 2024. Events were captured via online survey and classified using a five-grade severity scale. Pearson's chi-square tests assessed associations between event characteristics and outcomes. During the survey data-collection period, 11,581 CRU visits were recorded. Overall, 261 events were reported, including 100 incidents prevented before becoming protocol deviations and 68 confirmed protocol deviations; 93 events preceded systematic collection of the prevented/resulted classification. Grade 1 was the most frequently selected severity classification (126 of 178 severity-grade selections; 70.8%). The laboratory (60.2%) and study setup (33.7%) were the most frequent event contexts. Among events with a prevented/resulted classification, Grade 1 events were more frequently prevented than Grade 2-5 events (71.8% vs. 25.0%; p < 0.0001). Quality improvement strategies implemented during the observation period included dedicated study liaisons, standardized start-up checklists, increased staffing, and structured communication processes. Lower-severity events were more frequently prevented before resulting in protocol deviations. These findings identify operational contexts and event characteristics that may inform proactive, systems-based approaches to deviation prevention in CRUs.

Rheumatoid arthritis (RA) requires lifelong treatment with disease-modifying antirheumatic drugs (DMARDs). In patients with sustained disease control, DMARD de-escalation has been proposed to lessen treatment burden, but its efficacy and safety remain elusive. We conducted a systematic review and meta-analysis of randomized controlled trials evaluating DMARD tapering or withdrawal in controlled RA. PubMed, Embase, and the Cochrane Library were searched to January 16, 2026. The primary outcome was disease flare, and secondary outcomes included radiographic progression and adverse events (AEs). Risk ratios (RRs) were pooled using Mantel-Haenszel models. Prespecified subgroup analyses were conducted by DMARD class and follow-up duration. In total, 27 publications comprising 5262 participants were included. Any DMARD tapering increased flare risk overall (RR 1.56, 95% CI 1.26-1.98), mainly driven by the biologic DMARD (bDMARD) subgroup, with increased flare risk observed < 9 months (RR 1.61, 95% CI 1.13-2.31), but not at longer follow-up (> 9 months). In contrast, any DMARD withdrawal substantially increased flare risk (RR 2.23, 95% CI 1.83-2.94), primarily driven by the bDMARD subgroup across all follow-up durations (< 9 months, 9-18 months, and > 18 months). Both tapering (weak evidence) and withdrawal (strong evidence) promoted radiographic progression at 9-18 months and > 18 months. Neither tapering nor withdrawal reduced AEs. In conclusion, DMARD tapering, particularly of bDMARDs, should be considered cautiously in controlled RA because of the increased flare risk, possible radiographic progression, without clear reduction in AEs. DMARD withdrawal should be generally avoided whenever possible. TRIAL REGISTRATION: CRD420251066774 (PROSPERO).