Portfolio
Pharmacometrics, omics and sensor data
Population PK/PD modeling and model-informed precision dosing in pediatric clinical pharmacology, plus untargeted metabolomics and wearable sensor data. My research projects, published software and papers are below.
- 6Research projects
- 6Publications
- 10Public repositories
- 6Conference abstracts
Doctoral research
PhD dissertation · expected 2027
Advancing model-informed precision dosing via mechanistic and data-driven algorithms: an infliximab case study
University of Cincinnati College of Medicine ·
Department of Pharmacology, Physiology and Neurobiology
Advisor: Tomoyuki Mizuno, PhD
Children on infliximab for Crohn's disease get sparse, irregular monitoring: a few trough concentrations and biomarkers, taken when the clinic could. My dissertation asks how far that routine therapeutic drug monitoring can go toward choosing a better next dose, and whether a mechanistic model, a learned model or both together work best.
I build a population PK/PD model of infliximab exposure and response, then test ensemble machine learning and neural ODEs against it for predicting biochemical remission, and deep reinforcement learning for the dosing policy. Two supplementary chapters extend this to adult observational cohorts.
- Population PK/PD
- Model-informed precision dosing
- Ensemble ML
- Neural ODEs
- Deep reinforcement learning
Results stay private until the defence.
Projects
PK/PD modeling and machine learning for infliximab in pediatric Crohn's disease
Mechanistic population PK/PD models for infliximab (a monoclonal antibody) in children with Crohn's disease, combined with machine learning to improve response prediction and guide dosing. Exposure varies widely between children, and standard dosing leaves many of them under-exposed.
Approach
- Population PK/PD models in NONMEM with covariates for body size, disease status and immunogenicity.
- Indirect response models for the delayed effect on disease biomarkers and clinical endpoints.
- Exposure–response analysis linking AUC and trough concentrations to disease activity scores.
- XGBoost and random forest for disease response, benchmarked against the pharmacometric models to see what the PK/PD features add.
What it gives you
- How much infliximab exposure and response vary between children, and how much of that the model explains.
- Once between-subject variability is attributed to measurable characteristics and the patient's own monitoring data, the remaining uncertainty is small enough for a model-based dose to be useful.
- A profile of who reaches and stays in biochemical remission, from induction-phase data, while the dose can still be changed.
- Together that's what model-informed precision dosing needs: a model of the variability, an exposure target tied to a clinical endpoint, and an early read on who is on track.
- NONMEM
- mrgsolve
- R
- Python
- scikit-learn
- PyTorch
Top 10 abstract Presented at the IATDMCT Congress, 2025.
Model-informed dose optimization for prophylactic piperacillin–tazobactam in critically ill children
Prophylactic antibiotic doses for critically ill children are usually extrapolated from adults and rarely validated. This work gives that dosing a population PK evidence base for children in perioperative care.
Approach
- Population PK models in NONMEM for piperacillin in critically ill pediatric patients.
- Covariate analysis of body size, renal function and critical illness on clearance and volume of distribution.
- Monte Carlo simulation of candidate regimens, scored by probability of target attainment.
- An R-Shiny dashboard where clinical teams can change weight, renal function and regimen and see the profile and target attainment.
Results
- Body size and renal function are the covariates that matter for piperacillin exposure.
- Standard regimens are suboptimal for parts of the pediatric population.
- Dosing recommendations that account for patient-specific factors, to support antibiotic stewardship.
- NONMEM
- Pirana
- mrgsolve
- R
- R-Shiny
Published Antimicrobial Agents and Chemotherapy, 2025 (doi:10.1128/aac.01227-24). Presented at ACOP 2024 and ASCPT 2024.
Ensemble learning and deep reinforcement learning for infliximab dose selection
The data-driven half of the dissertation: what machine learning adds on top of a mechanistic population PK/PD model when all you have is the sparse, irregular data from routine care.
Approach
- Ensemble models (gradient boosting and bagging) predicting biochemical remission from treatment-course features.
- Neural ODEs for the biomarker trajectory, which handle irregular sampling in continuous time without forcing it onto a grid.
- A deep Q-network learning a dosing policy, rewarded on target attainment and remission.
- The learned policy compared with standard dosing and with maximum a posteriori Bayesian dosing to a concentration target.
Status
- Analyses are being finalised, and results are held back until the defence.
- Aggregate result tables and figures are in a private repository that I'll open as the work is published.
- XGBoost
- Random forest
- Neural ODEs
- Deep Q-networks
- Python
- PyTorch
- NONMEM
Private Results repository: Wrlog/mipd-ensemble-rl (access on request until the work is published).
Metabolite–wearable associations in healthy volunteers
Healthy volunteers wore sleep, cardiovascular and activity trackers and were sampled for untargeted plasma metabolomics. I looked at what the sensor record explains about the metabolome, and how much of a metabolite's apparent association with behaviour is just the time of day the blood was drawn.
Approach
- One pre-specified exposure per sample from the sensor streams. Window, summary and reference were fixed before looking at outcomes, and alternatives are reported as sensitivity analyses instead of searched over.
- Wear-time rules set in advance, reporting how many participant-days each rule removed.
- Linear mixed-effects models with a participant intercept, splitting within- from between-person variation so the coefficient is a change from the participant's own baseline.
- Harmonic terms on sampling clock time, because the metabolome and every wearable metric vary within the day and collection time would otherwise show up as a spurious effect.
- Benjamini–Hochberg control within each wearable metric, collapsing discoveries to compounds before counting.
How it's built
- A Python pipeline from raw device exports and metabolomics deliveries to the association tables, carrying the filtering counts into the output.
- A dashboard of the diurnal profiles, QC and association results for the study team.
- Python
- Mixed-effects models
- FDR control
- Untargeted metabolomics
- Wearable sensor data
Analysis code and data belong to Enveda Therapeutics and are not public. Background on the methods: wearable data and molecular phenotypes.
Untargeted metabolomics of inflammatory bowel disease
Plasma metabolomics on an adult IBD cohort already profiled by genomics, transcriptomics and proteomics, but not metabolomics. Can the metabolome separate ulcerative colitis from Crohn's disease, and active disease from remission? And could it predict treatment response before the first dose, which the dosing work in the dissertation has to assume away?
Approach
- A reproducible notebook sequence per delivery: load and hash the exports, filter and normalize, QC and missingness, PCA, differential analysis, candidate review.
- Annotation and prevalence filters compared with alternatives, with the cost of each choice reported.
- Patient-intercept mixed models for repeated visits, with ordinary least squares and a run-order-adjusted model as sensitivity comparisons, and Benjamini–Hochberg control within each contrast.
- Non-detects left as missing rather than imputed, plus a separate presence/absence test, since a model of detected intensity can't see a feature that's absent in one group.
- Cross-batch comparability and low-signal exclusion run as diagnostics that report and stop when they don't apply.
What came of it
- A feasibility assessment instead of a classifier. Medication is recorded without start dates, so only a handful of patients in this cohort have a plasma sample from before a documented treatment start.
- That sets the design and size a study would need to answer the pre-treatment question, which is what the supplementary chapter reports.
- Python
- Mixed-effects models
- PCA
- FDR control
- Untargeted metabolomics
Analysis code and data belong to Enveda Therapeutics and are not public. Background on the methods: untargeted metabolomics, from feature table to association list.
External validation of a published infliximab model in very early onset IBD
A population PK model published in one paediatric population doesn't automatically hold in another. This tests an existing infliximab model in a very early onset IBD cohort, children much younger than the populations it was built on, and reports how well it predicts them.
Approach
- The published model with every parameter fixed and nothing re-estimated, so it's tested as published.
- Individual clearances as empirical Bayes estimates for each child, given the published priors.
- Prediction-corrected visual predictive checks for Crohn's alone and for the pooled Crohn's, ulcerative colitis and indeterminate colitis cohort.
- Interactive HTML dashboards of the cohort, diagnostics and per-patient PK for the study team.
How it's built
- Python for data preparation and reporting, NONMEM and PsN for fitting and VPCs. Versions are pinned, since a pandas change can move a histogram bin edge.
- The repository has code only: no patient data, datasets or rendered dashboards. A synthetic-data generator mimics the real workbook's structure and awkward cases, so the pipeline runs without PHI.
- NONMEM
- PsN
- VPC
- Python
- External validation
Private Code: Wrlog/VEO_CCHMC (access on request).
Software
Public repositories at github.com/Wrlog.
dosing-simulator
An R-Shiny app that simulates intravenous dosing regimens in a simulated population, using a two-compartment model with first-order elimination, allometric scaling on weight, a power function on renal function, and log-normal between-subject variability on clearance and central volume. It reports concentration-time profiles with prediction intervals, target attainment, time above target and trough statistics. No patient data; it's for research and teaching. The PK solver is a closed-form solution of the two-compartment model, so it all runs in the browser with no server.
- R
- Shiny
- WebAssembly
mipd-bayesian-forecasting
Model-informed precision dosing for a hypothetical Drug X, whose effect is followed through a biomarker. Sparse troughs and biomarker values are fitted by MAP Bayesian estimation, PK first and then an indirect-response PD model driven by average concentration per interval, and a recommender picks the cheapest regimen predicted to keep the biomarker on target. It compares exposure drivers, tests forward prediction from early visits against simpler forecasts, and checks what happens when the PD model is wrong. A browser app does the same for one patient.
- Python
- MAP estimation
- PK/PD
- Shiny for Python
rl-dosing-demo
Deep Q-learning for dose and interval choices on a simulated PK/PD patient, with the option to stop Drug X when the biomarker can't be brought to target. A ceiling regimen defines which patients are reachable at all, and the DQN is compared with the standard regimen and a Bayesian MAP comparator on 1,500 held-out patients. The DQN gets 82% of patients to target against 50% on the standard regimen, but the MAP comparator does better again at 90%, and the README goes into why.
- Python
- PyTorch
- Deep Q-learning
- PK/PD
remission-prediction-shap
Predicting whether a patient's biomarker will be on target a year into treatment, from what's known in the first weeks. A simulated cohort from three studies with different lab panels goes through a ladder of feature groups, from baseline clinical data up to updated PK/PD estimates, with repeated nested cross-validation, leave-one-study-out validation and SHAP checked against the known mechanism. Updated PK/PD information helps most, but raw values carry nearly as much as the model-based estimates.
- Python
- XGBoost
- CatBoost
- SHAP
latent-ode
A latent neural ODE that learns Drug X pharmacokinetics from simulated data where the true model is known. Doses enter as jumps in latent space, covariates condition the dynamics, and a recurrent encoder over the first few measured levels plays the part of Bayesian updating. Scored against the exact solution and a popPK MAP comparator: it matches popPK with no levels, falls behind once levels are given, and handles a mid-treatment rise in clearance better than a constant-clearance popPK model.
- Python
- PyTorch
- Neural ODE
- Population PK
litsearch
A small tool for making a literature screen reproducible. One YAML file sets the search blocks and screening rules; it queries Europe PMC, removes duplicates, applies the rules, and writes included and excluded records with reasons, the exact search strategy, a PRISMA flow diagram and a BibTeX file. The example searches population PK models of vancomycin in neonates.
- Python
- Europe PMC
- PRISMA
metabolomics-dashboard-demo
A reporting pattern for omics analysis: simulate a study, run curation and association, and render the result as one HTML file that opens offline without a server or notebook. It has QC drift, PCA and volcano plots and a filter bar to search and sort results. The data is synthetic and seeded, and the planted associations are listed so you can check the output against ground truth.
- Python
- numpy
- pandas
- scipy
- matplotlib
nonmem-model-library
NONMEM control streams for model families that come up often in drug development: two-compartment population PK, Claret tumour growth inhibition, an indirect response PK/PD model, a Weibull time-to-event model and a binary exposure–response model. Each has a simulated dataset and the parameters that generated it, so you can check a run against the known answer. Static checks on every commit compare $INPUT with the data columns in order, parameter references with declarations, and likelihood models with their $ESTIMATION record.
- NONMEM
- Python
- matplotlib
ml-clinical-pharmacology
Three R analyses on synthetic major depressive disorder cohorts from a graduate course in clinical pharmacology. The first clusters gene-expression profiles with PCA, k-means, hierarchical clustering and Gaussian mixtures, and six of ten clusterings recover the four diagnostic groups exactly. The other two predict SSRI remission with random forest, SVMs and gradient boosting in tidymodels. Clinical factors alone reach a test AUC of 0.90 and adding biomarkers takes it to 1.00, but at the default threshold several models miss most remitters, and an external cohort with lower biomarker levels shows why a fixed cutoff needs recalibrating.
- R
- tidymodels
- Unsupervised learning
ml-practice
Four machine learning projects: fraud detection, demand forecasting, semantic similarity and an implicit-feedback recommender. Each runs on a laptop CPU in under three minutes, has a test suite, and is framed around a practical decision, like pricing the two error types to pick a fraud threshold, deriving every lag from the forecast horizon so nothing leaks, or showing that the split protocol moves the score more than the model does. All data is synthetic and generated in the repository.
- Python
- LightGBM
- scikit-learn
- matplotlib
Publications
- Irie K, Tan WR, Mizuno T. Towards reinforcement learning-enabled model-informed precision dosing: concepts, applications, and implementation. Therapeutic Drug Monitoring. 2026. In press
- Taylor ZL, Nadai T, Irie K, Tan WR, Parikh S, Hosawi A, van Hoogdalem MW, Morales Junior R, Tang Girdwood S, Leino AD, Mizuno K, Vinks AA, Mizuno T. Clinical model-informed precision dosing consult service for accelerating personalized medication in pediatric patients. Clinical Pharmacology & Therapeutics. 2026. doi:10.1002/cpt.70324
- Morales Junior R, Mizuno T, Tan WR, Irie K, Tang Girdwood S. From PICU to NICU: extrapolating meropenem exposure from pediatric to neonatal intensive care patients. The Journal of Clinical Pharmacology. 2026;66(1):e70097. doi:10.1002/jcph.70097
- Tan WR, Irie K, McIntire C, Luna Torres J, Jones R, Gibson A, Mizuno T, Tang Girdwood S. Model-informed dose optimization for prophylactic piperacillin-tazobactam in perioperative pediatric critically ill patients. Antimicrobial Agents and Chemotherapy. 2025;69(3):e01227-24. doi:10.1128/aac.01227-24
- Yang Z, Tan WR, Li Q, et al. Population pharmacokinetic study of the effect of polymorphisms in the ABCB1 and CES1 genes on the pharmacokinetics of dabigatran. Frontiers in Pharmacology. 2024;15:1454612. doi:10.3389/fphar.2024.1454612
- Tan WR, Sheikh Ghadzi SM, Hyder Ali IA, Harun SN. Systematic review of population pharmacokinetic models of isoniazid in children and adults with tuberculosis. Malaysian Journal of Pharmacy. 2022;8(2):1-15. doi:10.52494/djiq7058
In preparation
- Tan WR, Irie K, Boyle BM, Noe JD, Hyams JS, Minar P, Mizuno T. Population pharmacokinetic-pharmacodynamic modeling of infliximab and fecal calprotectin in pediatric Crohn's disease: towards biomarker-guided precision dosing.
- Tan WR, Mizuno T. Early identification of long-term biochemical remission in pediatric Crohn's disease using machine learning.
- Tan WR, Mizuno T. Advancing model-informed precision dosing: a critical appraisal of emerging computational tools and point-of-care diagnostics.
Conference presentations
- Minar PP, Tan WR, Irie K, Boyle BM, Noe JD, Hyams JS, Mizuno T. Bridging drug exposure to biochemical remission: a PK/PD framework to individualize infliximab target concentrations in pediatric Crohn's disease. Digestive Disease Week, 2026.
- Tan WR, Irie K, Minar P, Mizuno T. Pharmacokinetic-pharmacodynamic modeling of infliximab in children and young adults with Crohn's disease. IATDMCT Congress, 2025. Top 10 abstract
- Morales Junior R, Tan WR, Irie K, Mizuno T, Tang Girdwood S. From PICU to NICU: refining maturation factor for meropenem clearance. ASCPT Annual Meeting, 2025.
- Tan WR, Irie K, McIntire C, Luna Torres J, Jones R, Gibson A, Mizuno T, Tang Girdwood S. Model-informed simulations to determine optimal piperacillin/tazobactam dosing regimens in pediatric perioperative care: effect of body size and renal function. American Conference on Pharmacometrics (ACoP), 2024.
- Tan WR, Irie K, McIntire C, Luna Torres J, Jones R, Gibson A, Mizuno T, Tang Girdwood S. Model-informed dose optimization for prophylactic piperacillin-tazobactam in perioperative pediatric critically ill patients. ASCPT Annual Meeting, 2024.
- Tan WR, Tahir ARM, Daud NAA. Medical problems among Rohingya refugees and medications prescribed during IMARET clinic visits in Kelantan. International Journal of Human and Health Sciences, 2021;S26.
Awards
- 2025 Top ten abstract nomination, IATDMCT Congress
- 2022–2026 Enhancement Fund Award, Department of Pharmacology, Physiology and Neurobiology, University of Cincinnati, awarded in each of 2022, 2023, 2024, 2025 and 2026
- 2023 First prize, junior category, Graduate Student Research Forum, University of Cincinnati
Service and mentoring
- Journal review
- Reviewer, Clinical Pharmacology & Therapeutics.
- Conference review
- Abstract reviewer, ASCPT Annual Meeting 2025 and American Conference on Pharmacometrics (ACoP) 2024.
- Mentoring
- Research co-mentor to Soorya Shanmugam, October 2024 – February 2025, now an undergraduate at The Ohio State University.
Get in touch
I'm open to collaborations in PK/PD modeling, methods discussions, and internship or job opportunities.
Last updated September 2026