Zifeng MaiQuantitative Clinical Pharmacology · MIDD

Zifeng Mai · Clinical Pharmacology · Boston, MA

PICK THE
RIGHT DOSE.

I’m Zifeng Mai, Clinical Pharmacology Lead at CRISPR Therapeutics. I help drug developers pick the right dose — with pharmacokinetics, modeling & simulation, and data.

Portrait of Zifeng Mai, framed by the therapeutic window between an efficacy curve and a toxicity curve
PathFrom immunology research to regulated bioanalysis to clinical pharmacology
  1. Yale School of MedicinePostgraduate Associate · pediatric hematology & oncology
  2. Bristol Myers SquibbAssociate Research Scientist · immuno-oncology
  3. Frontage LaboratoriesScientist to Associate Director · promoted three times
  4. Novartis (NIBR)Principal Scientist · bioanalytical assays
  5. EQRxAssociate Director · Bioanalytical Sciences
  6. CRISPR TherapeuticsSenior Manager · Clinical Pharmacology Lead

The featured case study, by the numbers

7

workflow steps, from SDTM to report

4

analysis datasets: PopPK, PKPD, efficacy, safety

8

NONMEM runs reproduced in nlmixr2

42

issues found auditing the deliverables

13

decision gates recorded for review

Where I’ve worked

CRISPR TherapeuticsEQRxNovartis (NIBR)Frontage LaboratoriesBristol Myers SquibbYale School of MedicineTargeted Therapeutic SolutionsUniversity of PennsylvaniaChinese Academy of Sciences
About

What I do

Three things I spend my time on, in plain language — the same questions every drug program has to answer.

  • 01

    Picking the right dose

    I figure out how much drug a patient needs, and how often. Using pharmacokinetic data and modeling & simulation, I help teams choose doses that are safe and effective — before a trial ever starts.

  • 02

    Proving the drug works in the body

    I lead the clinical assays — PK, PD, immunogenicity — that show whether a drug is doing what it should inside patients, working across discovery scientists, clinical teams, and external vendors.

  • 03

    De-risking development

    I represent clinical pharmacology on cross-functional program teams: shaping trial protocols, supporting IND-enabling work, and building the data packages behind regulatory submissions.

Core competenciesClinical pharmacology strategyPK/PD modeling & simulationPopPK & exposure–response (NONMEM, R)Dose selection & optimizationIND · BLA · CTA · MAA · RMATFDA · EMA · MHRA interactionsICH / FDA / EMA guidanceCRO, vendor & budget governanceRegulated PK, ADA/NAb & biomarkersCross-functional matrix leadershipScientific writing & project managementTeam building & mentorship
Experience

From the bench to the dose.

Research roles in drug delivery and immunology, then regulated bioanalysis, then clinical pharmacology for in vivo gene-editing programs.

  1. Aug 2023 – Present
    Current

    Senior Manager, Clinical Pharmacology Lead, CRISPR Therapeutics

    Clinical pharmacology for in vivo gene-editing programs: clinical assays (PK/PD/immunogenicity), model-informed drug development, IND-enabling work.

    Boston, MA
  2. Dec 2021 – Aug 2023

    Associate Director, Bioanalytical Sciences, EQRx

    Bioanalytical and immunogenicity strategy for global biotherapeutic filings (MHRA, EMA, FDA).

    Cambridge, MA
  3. Dec 2020 – Dec 2021

    Principal Scientist, Novartis Institutes for BioMedical Research

    Bioanalytical assays for biologics, oligonucleotides, and cell & gene therapy.

    Cambridge, MA
  4. Apr 2016 – Dec 2020

    Associate Director · Group Leader · Scientist, Frontage Laboratories

    Promoted three times; led regulated PK/TK and immunogenicity teams in a GLP environment.

    Exton, PA
  5. Aug 2015 – Apr 2016

    Associate Research Scientist, Immuno-oncology, Bristol Myers Squibb

    PK/PD assay development for cancer immunotherapy.

    New Jersey
  6. Sep 2014 – Jul 2015

    Postgraduate Associate, Pediatric Hematology & Oncology, Yale School of Medicine

    Immune assays on patient and donor samples; flow cytometry, CyTOF, Luminex.

  7. Dec 2013 – Aug 2014

    Research Associate, Targeted Therapeutic Solutions LLC

    Anticoagulant therapy development.

    Philadelphia, PA
  8. Jul 2013 – Aug 2014

    Research Assistant, Smilow Center for Translational Research, University of Pennsylvania

    Novel drug-delivery system with thrombomodulin for cardiovascular disease.

  9. Oct 2011 – Jun 2012

    Intern, South China Sea Institute of Oceanology, Chinese Academy of Sciences

    Population genetics of the Hong Kong oyster via microsatellite markers.

    Guangzhou, China

Education

  • Master of Biotechnology, Biopharmaceutical TrackUniversity of Pennsylvania · Philadelphia, PA
  • Master of Computer Information SciencesHarrisburg University of Science and Technology · Harrisburg, PA
  • BS, BiotechnologyBeijing Normal University at Zhuhai · China

Technical expertise

  • Clinical pharmacology & MIDDNCA, population PK, PK/PD and exposure–response modeling, QSP, dose selection and optimization, FIH/SAD/MAD, translational and biomarker integration
  • Modeling & softwareNONMEM (via modeling vendors), Phoenix WinNonlin, R, Python, SQL; LLM-assisted analytics workflows
  • Bioanalytical & translationalPK, PD, ADA (binding/neutralizing), cellular immunogenicity, cytokine profiling; ELISA, MSD, Luminex, Gyrolab, Quanterix, ELLA, SPR, LC-MS/MS, qPCR/ddPCR, flow cytometry, CyTOF, cell-based assays
  • Regulatory & complianceIND, CTA, BLA, MAA, RMAT submissions; briefing packages; Investigator's Brochures; protocols; GLP, GCP, GCLP
Featured work

What dose, for whom, and why.

This case study walks through a complete model-informed drug development workflow for a first-in-human LNP therapy — from population pharmacokinetics to exposure–response analysis to dosing simulation.

It shows how modeling answers the questions trial teams and regulators actually ask: what dose, for whom, and why. The entire analysis is rebuilt end to end in open-source R, so every step is transparent and reproducible.

Plasma concentration after a single IV infusion, by dose levelSynthetic data

Median and 5th–95th percentile of 300 simulated subjects per dose from a triexponential model with between-subject variability. Every figure on this page is generated in your browser from made-up parameters and contains no study data.

Case study · first-in-human dose escalation

Population PK, exposure–response and dose simulation for a first-in-human LNP therapy

Model-informed drug development for a dose-escalation study: datasets built from SDTM, population PK and PKPD models, exposure–response for efficacy and safety, and virtual-population dosing simulations. The whole pipeline was then audited and rebuilt in free, open-source R.

Walk through the seven steps

Open-source rebuild

The whole pipeline in free R, matched to NONMEM

Every delivered NONMEM run translated, refitted in nlmixr2 and compared parameter by parameter.

See the parity check

Independent audit

42 findings, 13 decision gates

Every script, control stream and dataset read line by line, each finding classified by whether it can change a reported number.

Read the findings

AI-assisted workflow

The method, packaged as a reusable agent skill

The project’s conventions distilled into a pharmacometrics skill for an AI coding agent.

See how it works
The program

One analysis plan, from data to dose

The study escalated an intravenously infused lipid-nanoparticle (LNP) therapy through weight-based and flat-dose cohorts, with LNP lipid components measured in plasma.

As the study progressed, new subjects, redoses and manufacturing lots arrived. The datasets, models, exposures and exposure–response analyses were updated with each data cut and checked against the previous delivery.

Most of the effort went into keeping the chain from data to model to exposure to response consistent as the data changed under it.

Workstreams in the program

The analyses

From raw SDTM to a dosing recommendation

Each step below follows the order the work ran in. Every later step reads the output of the steps before it.

  1. Step 1Data

    Analysis datasets and QC

    I built NONMEM-ready PopPK, PKPD, efficacy and safety datasets from SDTM domains (PC, EX, LB, VS, DM). With each data cut, a regression QC compared the new build against the delivered dataset.

    • Infusion records with duration and rate; dose number and nominal dose per administration, so redoses are handled correctly.
    • BLQ samples flagged for M1/M3 handling; baseline covariates plus derived lean body weight and CKD-EPI 2021 eGFR.
    • Lot-specific lipid content, so a dose in mL converts to mg of each lipid using that dose's own lot.
    • Cheap cross-checks run before any code: median(TAD − NTAD) caught a 24 h nominal-time offset; amount ÷ body size showed that “mg/kg” actually meant per kg lean body weight.

    Tools

    • SDTM
    • tidyverse
    • pandas
    • CTCAE v5
    # 1a_dataset_qc.log  (format example, values synthetic)
    check                          result
    unique (ID, TIME, EVID, CMT)   PASS   0 duplicates
    rows lost in joins             PASS   0
    median(TAD − NTAD), h          FLAG   −24.3  → nominal day offset
    dose basis (AMT / body size)   PASS   matches mg/kg lean body weight
    biomarker units vs prior cut   FLAG   ratio 38.7 → mmol/L vs mg/dL
    NDOSE vs AMT                   FLAG   8 records hold volume (mL)
    regression vs delivered file   PASS   DV, TIME, AMT identical
    # 3 flags → decision gates G1–G13
  2. Step 2NCA

    Exploration, NCA and dose proportionality

    I explored concentration–time data by dose, lot and covariate tertile. I also added a non-compartmental analysis (PKNCA, IV infusion) on each subject's first dosing window. Dose proportionality used the power model, judged against the Smith critical interval, with dose expressed as mg of lipid so weight-based and flat cohorts sit on one axis.

    • Geometric-mean and dose-normalised profiles; individual spaghetti plots; BLQ summaries.
    • Interactive HTML (plotly) for every plot, plus a static PNG.

    Tools

    • PKNCA
    • ggplot2
    • plotly
    Power model: Cmax vs dose (log–log)Synthetic data

  3. Step 4PKPD

    Exposures and target-protein PKPD

    I used post-hoc estimates from the final PopPK models to derive per-subject, per-dose exposures: Cmax, Tmax, Ctrough, AUC0–720 h, and AUC to the next dose. Lot-specific content converted each dose to mg of lipid. A sequential PKPD model then linked exposure to knockdown of the circulating target protein.

    • I compared three PD structures: the delivered indirect-response model, a logit-bounded Imin model, and an irreversible-effect model in which the effect plateaus and does not recover.
    • VPC on fold change from baseline, with baseline joined per subject rather than recycled across replicates.
    Target protein, fold change from baselineSynthetic data

    Irreversible-effect model: knockdown develops over weeks and then holds. Lines show the population prediction; dots are simulated observations.

  4. Step 5Exposure–response

    Exposure–response for efficacy and safety

    For efficacy, I related biomarker and target-protein change (at single visits and time-averaged from day 28) to each exposure metric, fitting null, linear, Emax and sigmoid Emax models. I selected models by likelihood-ratio test and AIC and bootstrapped the selected fits. For safety, I modelled maximum liver-enzyme fold change (and × ULN) against exposure, and fitted CTCAE grade ≥ 1 with Firth logistic regression, one metric at a time.

    • One observation per subject per dose, so random effects were dropped in favour of nonlinear least squares (nlsLM) plus a bootstrap.
    • BLQ biomarker values imputed at LLOQ/2 instead of being silently dropped.

    Tools

    • minpack.lm
    • logistf
    • bootstrap
    Biomarker % change vs exposure: Emax fitSynthetic data

  5. Step 7Report

    Abbreviated PopPK and E–R report

    The results were written up as an abbreviated population PK and exposure–response report, which went through two drafts to a final version. It covered the methods, final-model parameter tables, diagnostics, exposure summaries, ER conclusions and dosing implications for the next cohorts. The open-source rebuild renders the same report from code with officedown, so the tables and figures cannot drift from the analysis.

    1. 1Study design and data
    2. 2PopPK methods and final models
    3. 3Model-predicted exposures
    4. 4Efficacy exposure–response
    5. 5Safety exposure–response
    6. 6Conclusions for dose selection
Open-source rebuild

The whole pipeline in free R, matched to NONMEM

I re-implemented the full workflow so it runs with free packages only: no NONMEM, PsN, Phoenix or vendor-internal libraries.

The rebuild is one dependency-ordered run_all.R with one timestamped run folder per step. Each folder holds a plan.md, the script, its log, and the tables, plots and models it produced. Paths, assay limits, lot contents and unit factors live in one config.yml.

Parity came before anything else. Every delivered NONMEM run was translated with nonmem2rx, refitted in nlmixr2 (FOCE-I) and compared parameter by parameter. Objective functions matched within 0.12 units, the final models' parameters within 3.3%, and population predictions within 0.03% of the NONMEM tables.

Parity check: R estimate vs NONMEM, % differenceLayout example
acceptance: ±5% fixed effects, ±20% variances

This chart is a layout example with synthetic values. In the real comparison every parameter fell inside these acceptance bands.

Proprietary tool → free replacement
OriginalReplacement
NONMEM FOCE-Inlmixr2(est = "focei")
ADVAN subroutines / $DESrxode2 ODEs
$OMEGA BLOCKeta.cl + eta.v ~ c(…)
.lst / .ext / $TABLEnonmem2rx validation
PsN vpcvpcSim + tidyvpc
PsN scm, bootstrapcustom scm(), bootstrap_fit()
xpose / vendor GOFggplot2 GOF and pcVPC
Phoenix ER Emaxminpack.lm::nlsLM
Word templateofficedown
# Generic 3-compartment IV model (illustrative values, not the study model)
pk_3cmt <- function() {
  ini({
    tcl <- log(10); tv <- log(5); tq <- log(3); tvp <- log(20); tq2 <- log(1); tvp2 <- log(100)
    eta.cl ~ 0.09
    eta.v  ~ 0.09
    prop.sd <- 0.2
  })
  model({
    cl  <- exp(tcl + eta.cl)
    v   <- exp(tv + eta.v)
    q   <- exp(tq); vp <- exp(tvp); q2 <- exp(tq2); vp2 <- exp(tvp2)
    cp  <- linCmt()                                # dose mg, V L → mg/L; ×1000 for ng/mL
    cp ~ prop(prop.sd)
  })
}
The economics

What the same analysis costs.

The open-source rebuild wasn't just a technical exercise — it changes the unit economics of modeling. Same scope, two prices.

Modeling vendorsDone for you

$200K–400K

End-to-end popPK, PK/PD, exposure–response and simulation for a program, outsourced.

Open-source R + Claude CodeThis site

~$500

R, nlmixr2, rxode2: $0. Claude Sonnet 5.5 at $2/$10 per million tokens — even 100M tokens of agentic coding lands around $300–500.

Up to 99.9% cheaper — and about a week of focused work once the analysis criteria are set.

Independent audit

42 findings, 13 decision gates

Rebuilding meant reading every script, control stream and dataset line by line, and profiling the data independently in Python.

Each finding is classified by whether it can change a reported number. It is then fixed, kept for reproducibility with a flag, or escalated to a decision gate: a choice the scripts never make silently, recorded in a file the next step reads.

Findings by area
Findings that could change a reported number
ER

BLQ biomarker values dropped before averaging, which biased the time-averaged effect size toward a smaller change.

Effect size
ER

Likelihood-ratio test computed from AIC with the wrong degrees of freedom, so the model-selection p-values were wrong.

p-values
SIM

ER model fitted on one AUC window but driven by another in the simulations, which inflated predicted efficacy.

Efficacy
DATA

Covariate units changed between data cuts (mmol/L to mg/dL), so the two covariate models are not interchangeable in simulations.

Simulation

Examples of decision gates

G1 · AssayWhich LLOQ?The specification and the data disagreed by an order of magnitude.
G3 · DataExclusionsPre-dose quantifiable samples, outliers and dose deviations.
G4 · ModelFinal covariate modelThe delivered model or the SCM result, weighed with bootstrap CIs.
G5 · ERBLQ handlingDrop, set to zero, or impute LLOQ/2.
G7 · ERMetric per endpointWhich exposure metric and model drive each simulated endpoint.
G8 · SimVirtual populationPopulation definitions, derived-biomarker equation and sampling method.
G10 · BasisWhich data cut drives downstream?Set once in config, read by every later step.
G11 · DataMixed dose unitsA dose column that held mL for some cohorts and mg/kg for others.
AI-assisted workflow

The method, packaged as a reusable agent skill

I distilled the conventions from this project into a pharmacometrics skill for an AI coding agent.

Given PK/PD data or a NONMEM control stream, the agent writes a plan, then ready-to-run, self-checking R or NONMEM code, then an interpretation of the outputs. It covers SAD-to-MAD pipelines, concentration–time plots, stepwise PopPK, target attainment, exposure–response, dosing simulations, and auditing or regenerating a CRO deliverables folder in free R.

  1. Inspect before planningProfile every file first and run cheap cross-checks for time offsets, dose basis and unit factors.
  2. One run folder per taskA timestamped folder holding the plan, script, log, tables, plots and models.
  3. Self-checking codeCounts derived from the data, stopifnot() on joins and keys, and regression QC against previous outputs.
  4. Stop at decision gatesPresent the evidence and a recommendation, ask, and record the answer for the next step.
  5. Units written outUnit arithmetic in comments: mg and L give mg/L, which is µg/mL.
  6. Answer firstThe key numbers lead, then evidence and caveats. Keep “shown” separate from “not rejected”.
# plan.md — written before any code runs
# Covariate analysis and final model
**Analysis Description:**
1. Refit final models to the current data cut
2. Parity vs NONMEM .ext and $TABLE
3. GOF, pcVPC, bootstrap (n = 200)
**Datasets:** poppk_<date>.csv
**Columns Used:**
- TAD: time after dose (h)
- AMT: dose (mg lipid)
- WT:  body weight (kg), allometry ref
**Expected Outputs:** parity table, GOF, VPC
### Assumptions
- Effects FIXED to 0 in the control stream are omitted
- Imprecise IIV (high RSE): check bootstrap first
- Biomarker covariates in mg/dL in this data cut
Toolkit

What the work ran on

  • EstimationNONMEM 7 (FOCE-I), PsN, nlmixr2
  • Simulationrxode2, mrgsolve
  • NCAPKNCA
  • Model translationnonmem2rx, babelmixr2
  • Diagnosticstidyvpc, ggplot2, xpose
  • Exposure–responseminpack.lm, logistf
  • Virtual populationsNHANES via nhanesA
  • DataCDISC SDTM, tidyverse, Python/pandas
  • ReportingR Markdown, officedown, flextable, plotly
  • Engineeringconfig-driven pipelines, run folders, regression QC
  • AI toolingAgent skills for pharmacometrics code generation
  • StandardsCTCAE v5, CKD-EPI 2021, allometric scaling
Publications

Selected publications and presentations

  1. 2026

    Phase 1b results of an in vivo CRISPR-Cas9 gene-editing therapy

    Coauthor · Late-breaking science presentation, 2026

  2. 2023

    Harnessing NLP to generate Python scripts for optimizing calibration-curve fitting and regression model selection in bioanalysis

    Mai Z, et al. · Poster, 17th WRIB, 2023

  3. 2020

    Bioanalytical drug tolerance limit prediction and optimization model developed through machine learning with TensorFlow

    Mai Z, et al. · Poster, 14th WRIB, 2020 Poster award · Immunogenicity

  4. 2017

    Distinct dominant T-cell receptors with a tissue-resident memory phenotype in individual melanoma metastases

    Dhodapkar KM, Boddupalli CS, Bar N, … Mai Z, et al. · J Clin Oncol 35(7_suppl), 2017

  5. 2016

    Interlesional diversity of T cell receptors in melanoma with immune checkpoints enriched in tissue-resident memory T cells

    Boddupalli CS, Bar N, … Mai Z, et al. · JCI Insight 1(21), 2016

  6. 2015

    Tissue resident memory T cells create genetically distinct immune microenvironments within melanoma metastasis

    Dhodapkar K, Mai Z, et al. · J ImmunoTherapy Cancer 3(S2), 2015

  7. 2015

    Identification of a cereblon-independent protein degradation pathway in residual myeloma cells treated with immunomodulatory drugs

    Verma R, Mai Z, et al. · Blood, 2015

Contact

Get in touch.

The best way to reach me is email. You can also find me on LinkedIn.