Research

How human brain tissue builds itself, and how it goes wrong

The lab studies how the human nervous system is built, and how medicines and pollutants change it, in human stem-cell models. Advice on medicines in pregnancy goes by the week of pregnancy. If the state of the individual cell is what matters, that advice needs a different kind of evidence. Three directions follow from that question.

A circular tree. One cell at the centre divides outward in time to 96 cells at the rim. Three dashed rings mark exposure days, and orange dots mark branches that were exposed. Simulated, not data.
Simulated lineage. Dashed rings are exposure days; orange dots are exposed branches.

The three directions

  • Core central to the lab's plan
  • Next a clear first test; needs a partner or funding
  • Later intended, not yet designed

Neurodevelopment

How is a human brain built, and when is it vulnerable?

Cells in the developing cortex do not move together. At one gestational age, neighbouring cells sit in different molecular states. If a chemical can only injure cells in a short-lived state, the week of pregnancy predicts harm only as well as it predicts how many cells are in that state. The classical view, that susceptibility follows developmental time, remains the rival explanation.

Pregnant women are left out of most drug trials, and animal studies cannot show when a human brain cell is vulnerable. Human neural cells can show both timing and dose.

Models: human stem-cell-derived neural cultures and organoids. Doses are anchored to measured human exposure.

The main project

The plan is to put a molecular recorder into human brain organoids, small three-dimensional cultures grown from stem cells. It would write each cell's history into that cell's own DNA, to be read back one cell at a time. Three aims, in order; each decides whether the next happens.

  1. Aim 1. Write and calibrate a developmental record. Not started
  2. Aim 2. Calendar time, or cell state?
  3. Aim 3. Is the susceptible state causal, or only a marker?

Read the full plan →

Related directions

Core

Evidence for pregnancy medicines

Human-cell evidence on timing and dose, placed next to clinical data, for clinicians and regulators who must decide without trials in pregnancy.

Next

Nanoplastics in the developing brain

What particles do once they reach neural tissue, and whether the cells can clear them. The models that test medicines can test particles too.

Neuropathology

Inherited disease, tumours and developmental conditions, studied in human brain organoids.

Cells differ between patients, and within a single tissue. Each line here asks whether those differences predict who falls ill and who responds to a treatment. This is the earliest-stage direction: two lines have a first test, and the third has no aims yet.

Next

Huntington disease

The same CAG repeat length can give different ages of onset. If a treatment acts on the process that lets the repeat grow, its effect may differ from patient to patient, and a patient's own cells could show it before treatment.

Needs patient-derived lines and clinical follow-up data.

Next

Brain tumours: gliomas

Gliomas differ between patients and even inside one tumour, and treatments chosen by averages fail many people. The aim is to describe, in several solid human gliomas, the molecular and biological features that matter for choosing a therapy.

Needs a clinical partner, consent, ethics approval and a data-protection plan before any sample is touched.

Later

Autism and other neurodevelopmental conditions

Valproate in pregnancy raises the risk of autism, and long-term paracetamol use has been linked to ADHD. I worked on both: as first author on the valproate study and as a contributing author on the paracetamol one (see Publications). Whether paracetamol causes ADHD, and how, is unclear. If the susceptible cell states found in Aims 2 and 3 decide how a chemical injures the developing brain, the same states may be where to look for how these conditions begin. This is a hypothesis.

No aims yet. I will add them here once they are fixed.

Recording and measurement tools

The instruments and checks the other two directions depend on.

A recorder, tested in simulation first

A molecular recorder writes events into a cell's DNA. Before building one, I asked how much it could capture in a growing human organoid. A model of cortical progenitors shows an early bias built into division-coupled recorders. The simulation is posted as a preprint, a paper shared before peer review, with its code.

Read the preprint

Checking the model against real tissue

Organoid neurons need checking against real physiology. In a paper where I am second author, recordings from single neurons in human spinal cord organoids, between days 15 and 32, showed steady maturation and the firing pattern typical of motor neurons early on.

Shared tools

Core

A benchmark for cell-state recorders

Shared simulators and tests that say what a recorder can resolve before it is built.

Core

Protocols that read several things in the same culture

Cell death, differentiation, stress and function in one dish, published in the lab handbook so others can run them.

Next

How precisely can an organoid be dated?

Papers that use organoids convert culture day into fetal stage, and rarely say how wide that answer is. A paper on this is coming.

How I work

The same habits apply to all three directions.

  1. I start from a question that, as far as I know, no current method answers. Epidemiology dates exposure to a pregnancy. Single-cell atlases give cell states without histories. Lineage recorders such as iTracer give ancestry in human cerebral organoids, and signal recorders such as ENGRAM have logged signalling activity in mouse stem-cell models. I know of none that returns ordered exposure, ancestry and cell state for the same human neural cell. Building the method is therefore part of the project (the full plan).
  2. I write the test and the stop rule, the result that ends the idea, before the data. Each aim has a test that can fail, and the design allows the rival explanation to win.
  3. I say what kind of new it is. Here the new part is a combination of three advances, none from this lab, made to survive five months of culture. When "new" means a new tool, a new explanation or a new use, I say which. Roughly, a new tool changes the working principle or the architecture, a new explanation changes the framework, and a new use is a market claim. "Disruptive" gets a definition or is dropped. The table below shows these axes and the other ways of drawing the line.
  4. I test the model before I trust it. Organoids lack blood vessels, immune cells and adult maturity, and an organoid's developmental age is an estimate, not a known quantity. Neurons are checked against real physiology, doses against measured human exposure, and each question gets more than one readout.
  5. I count repeating and maintaining as science. A protocol is version 0.1 until someone has run it from the page and a second person has repeated it.
  6. I use AI to propose and to sort, and people decide. A model can list a lot of options quickly. Choosing the question, judging what is really new and signing the claim stay with people. The rules are on the Resources page.
  7. I ask early who will use the result. I have seen preclinical evidence weighed in drug development and clinical research, and I have seen where it fails to answer the question put to it. So before the experiment I ask what a regulator, a clinician or a company would need in order to trust the result.
  8. I leave room for what I did not plan. Before the data I write down what I expect and what would surprise me. Results that are interesting but off the plan go into a log, and later I decide whether any deserve a test. It is a habit, not a proven way to be more creative.

What I mean by new

In the biosciences, "radical" is applied to new methods, new mechanisms and new treatments as if they were one thing.

They are different things, and so are "radical" and "disruptive". The same piece of work can be radical on one axis and incremental on another, so I say which axis I mean. In biomedicine there is a second gap to keep open: a result can be radical at the bench and still need years of safety, efficacy and regulatory evidence before it changes practice. I keep the bench claim and the clinical claim apart.

Eight ways of drawing the line, with examples from biomedicine and where this lab standsMost of what I read does not use "incremental" and "radical" as a pair. Each source has its own, named in the second column. The table puts them side by side, and the matching is my reading, not theirs. When you read that something is radical, first find its row.Swipe sideways to see every column.
Eight ways of drawing the line, with examples from biomedicine and where this lab stands
AxisSources, and their own pairIncremental sideRadical side, or the source's own oppositeIn biomedicine (my examples)This lab
Working principleJudged by how the thing worksArthur 2009Own terms: improvement; radically newImproves something that works on the same base principle.Radical: uses a new base principle for the same purpose.A high-fidelity Cas9 that cuts the same way with fewer off-target edits is incremental. RNA-guided Cas9, replacing protein-guided zinc-finger and TALEN editing for the same purpose, is a new principle.No new principle. The three advances it combines already exist.
FrameworkJudged by the community, usually in hindsightKuhn 1962, Dosi 1982Own pairs: normal science and revolution (Kuhn); trajectory and paradigm (Dosi)Normal progress inside the accepted problems and standards.Radical: changes what counts as a problem and as a good answer.Mapping one more variant in a known pathway is normal progress. Treating peptic ulcer as a bacterial infection, not a result of stress and acid, changed the question.The central hypothesis, if it held, would change what counts as a window of vulnerability. That is the bold part. Whether it is framework-changing is for others to judge, later.
ArchitectureJudged by whether the links between parts changedHenderson and Clark 1990Own pair: incremental and radical, with modular and architectural between themChanges a component and keeps the links between components.Radical: new components and new links. A third kind keeps the components and changes the links. It can look minor and still hurt established firms.CAR-T cells join antibody-derived binding domains to T-cell signalling domains: known parts, a new linkage.Closest fit: known parts, put together in a new way.
MarketJudged by what happens in the market over yearsChristensen 1997Own pair: sustaining and disruptiveCalled "sustaining": serves current customers better. It can be technically radical.Not a radical claim. Christensen's own term is disruptive: it starts with overlooked customers, with a simpler or cheaper offer, then moves up.A first-in-class drug for the same patients and prescribers is sustaining, however new its mechanism. A simple point-of-care test that starts below laboratory accuracy and improves is the kind of case the disruption literature describes.No market claim. There is no company.
CitationsJudged by an index over many papers, not by importanceWu 2019, Park 2023Own pair: consolidating and disruptingConsolidating: later papers cite it together with the work it builds on.Disrupting: later papers cite it and leave out the work it builds on.Park and colleagues report the decline across fields, including the life sciences. Wu and colleagues find that small teams disrupt and large teams develop.Not a claim I can make about my own work.
Who judges, and whenJudged by who is asked, and how long they waitKuhn 1962, Dosi 1982Own pairs: normal and revolutionary (Kuhn); trajectory and paradigm (Dosi)Judged inside the standards in force. Within a paradigm an agreed criterion of progress exists (Dosi).The standards are part of what is in question, so a verdict needs the community and usually time (Kuhn). Even after the fact, comparing two directions can be ambiguous (Dosi).The bacterial account of peptic ulcer was contested for years before it replaced the stress-and-acid account.I do not judge my own work radical. Others decide, and later.
Can it be plannedJudged by what plans achieved, a debated pointStanley and Lehman 2015, Vinsel and Russell 2020, Christensen 1997Own pair: modest and ambitious objectives (Stanley and Lehman)A plan with modest objectives works: an objective only a stepping stone away is sensible to set and follow.The sources disagree. Stanley and Lehman argue that fixed objectives mislead when the goal is ambitious. Vinsel and Russell say disruption cannot be worked toward or planned. Christensen advises planning to learn, on the assumption that forecasts are wrong.Drakeman and colleagues (2021) use FDA priority review as their proxy for breakthrough drugs. For 1998 to 2016 they report biotech companies originated more of them than large pharma, and credit many parallel projects stopped quickly when they fail. The authors have a stake in the industry. Book.No plan can promise a radical result. Each aim of the recorder programme has a test that can fail, and its criteria were fixed before any data were seen.
AI-enabled scienceJudged by what was actually tested: few results, mostly smallBoden 1998, Gao 2024, Ding and Li 2025Own pair: exploratory and transformational creativity (Boden)Exploring a defined space faster. Boden found exploratory creativity the most successful kind in AI.Changing the space itself (Boden's transformational creativity) is harder still for AI. Gao and colleagues describe agents that generate creative, de novo hypotheses as a level that has yet to be realized.In one simulated lac-operon task, ChatGPT-4 produced fewer hypotheses than human subjects and never noticed an anomaly. The authors ran it twice, and the human data came from an earlier study.I do not call a result radical because AI produced it. The rules for AI use in this lab are on the Resources page.

Where creativity fits. In most rows above, the creative step is the one that changes the space: a new principle, a new framework, a new link between parts. Most research does not do that. Kuhn noted that hardly any research, even by the greatest scientists, is aimed at being revolutionary. So I plan for both: a floor of careful work, and a bold part that may fail.

Does one of these meet a problem you have?

Tell me in two or three lines. A first test costs little and can start before any grant does.

Email me Collaborate