Loaded, not fired

What psychiatric genetics and heritability actually mean, for conditions and for traits. We start at the beginning, with the three billion letters in every one of your cells, and work our way up to the question people ask me most: will my kids get it?

By Reid Robison, MD. 8,511 words, with a zoom from your body to a base pair, a tape measure for your DNA, a chromosome shelf, two gardens, a twin lab, a board of four natural experiments, seventy years of twin studies in one chart, a thousand dice, a gene-finding expedition, a Manhattan plot you can read, a two-hit simulator and a few light bulbs along the way.

I started my career in neurodevelopmental genetics, on what we called gene-finding expeditions: a family with something terrible running through it, a freezer full of blood samples, and the hope that somewhere in three billion letters was the one that had changed. I was on the team that found one of those, a new disorder in a family in Ogden, Utah, and I will tell that story below, because it is the cleanest example I know of what genetics can do.¹ I then spent fifteen years watching the field discover that almost nothing in psychiatry works that way, and that the real story is stranger and, I think, more hopeful.

This is the long version of what I tell patients and families when they ask about genes. It is also for the curious. I have tried to give you several light-bulb moments, and I have marked them, because the whole subject is a series of things that sound obvious once you see them and are nearly impossible to see before. I am also the parent of identical twins, which gives me a standing interest in the twin studies you will meet in section 2, and a small daily reminder that identical genomes do not make identical people.

1Start here: the genome

Everything in this post sits on a physical object, so let us look at it. You are made of about thirty trillion cells.⁴ Most of them carry a nucleus, and inside each nucleus is the same text: about 3.1 billion letters, written in an alphabet of four (A, C, G, T), in two copies, one from each parent, divided into 46 chromosomes.⁵⁶ The double helix itself has been known since 1953.⁷ Zoom in.

How long is it, anyway?

DNA is a very thin thread: two nanometers across, with the letters stacked 0.34 nanometers apart.⁸ Six billion letters at that spacing is about two meters of thread, folded into a nucleus six millionths of a meter wide. That is the first number people never believe. The second is what happens when you multiply by the body.

Fun facts, in the house style

• 3.1 billion letters in one copy; about 6.2 billion in the two you carry. Printed in ordinary type at a thousand pages a volume, the reference genome fills 130 volumes.⁹
• Reading it aloud at ten letters a second, without sleeping, takes about ten years. Nobody has. (3.1 × 10⁹ ÷ 10 ÷ 86,400 ÷ 365.)
• About 20,000 genes code for proteins, in roughly 1.5% of the text.¹⁰ The rest is regulation, structure, old viruses, and a great deal we are still arguing about. The idea that we have 100,000 genes, which I learned in medical school, was wrong by a factor of five.
• Any two people differ at four to five million positions, which sounds like a lot and is 0.1%.¹¹ Everything in this post lives in that 0.1%.
• Each of us is born with about sixty new mutations that neither parent had, most of them from the father, and about two more for each year of his age.¹² A few of those matter a great deal in section 4.
• Sequencing a whole genome cost about $2.7 billion and thirteen years the first time, in 2003, and is a few hundred dollars and a day now. The first fully complete sequence, including the parts the 2003 draft could not read, was finished in 2022.⁶
• Identical twins are not genetically identical. On average they differ by about five mutations that arose in the first days after the egg split, and in about one pair in seven, by many more.¹³ Their epigenetic marks drift apart with age, and their fingerprints never matched.¹⁴

The chromosome shelf

The text is divided into 23 pairs of chromosomes, numbered by size (mostly) from 1 to 22, plus X and Y. Here they are to scale, with the mental-health address on each one that I would point to first. Tap a chromosome.

2What heritability means, and what it does not

Now the word that causes the most trouble. Heritability is a statistic about a population: the fraction of the variation in a trait, across people in a particular place and time, that is attributable to genetic differences between them.² It is written h², it runs from 0 to 1, and it is one of the most misread numbers in medicine. Three things it is not:

• It is not a fraction of you. "Height is 80% heritable" does not mean 80% of your 180 cm is genes. Every centimeter of you is genes and environment at once; the number is about why people differ from each other, not about how one person was built.
• It is not fixed. Heritability depends on the environment the population lives in. If everyone is fed the same, genes explain more of the differences in height, so heritability goes up. If some people starve and others do not, it goes down. The heritability of intelligence test scores rises from about 20% in early childhood to 60–80% in adulthood, as people increasingly pick environments that match their genes.¹⁵
• It says nothing about changeability. Phenylketonuria is as heritable as a trait can be, and a diet prevents the brain damage entirely.¹⁶ Poor eyesight is highly heritable; glasses fix it in a minute. A high heritability tells you treatments will not come from changing the environment that already exists. It does not tell you a new environment cannot work.

The cleanest way I know to feel all three at once is the geneticist Richard Lewontin's two gardens.¹⁷ Try it.

Where the numbers come from: twins

Most heritability estimates for psychiatric conditions come from twins. Identical (monozygotic) twins share essentially all their DNA; fraternal (dizygotic) twins share about half of what varies, like any siblings. Both kinds usually share a womb, a house, a school and a family. So if identical twins resemble each other more than fraternal twins do, the extra resemblance is a fingerprint of genes, and a simple formula, from Falconer in 1965, turns the two correlations into an estimate.¹⁸ Play with it.

The method has assumptions, and they are argued about for a living. The big one is that identical twins are not treated more alike than fraternal twins in ways that matter for the trait (the "equal environments" assumption). It holds up better than critics expected when tested directly (twins who were wrong about their own zygosity resembled each other according to their true zygosity, not their belief),¹⁹ and twin estimates for psychiatric conditions have been confirmed in rough outline by adoption studies and, more recently, by DNA-based methods that need no twins at all.² But "80% heritable" from twins is best read as "high," not as "80.0."

What a high heritability looks like in a room

Here is the number that reorganized how I think about this. Schizophrenia has a twin heritability of about 80%, among the highest in medicine.²⁰ So if one identical twin has it, what are the odds the other does?

I will say the personal part once. When you raise identical twins you get a daily seminar on this. Same genome, same house, same dinner table, and from the first week two different people, with different fears, different tempos, a different relationship to the same dog. Whatever the numbers say about populations, the two individuals in front of you were never going to be the same, and the genome was never going to be the thing that decided who they became.

How we got to those numbers: twins, adoptees, and two mentors

The twin method is older than genetics. Francis Galton proposed it in 1875, as a way to weigh "nature and nurture," a phrase he coined, by comparing twins who were alike from birth with twins who were not.²¹ It took seventy years and some bad data to turn it into the numbers above, and the story of how the schizophrenia estimate moved is a good lesson in how a heritability is built, and how it can be built wrong. Here is every major design on one board. Tap each one.

Classic twins, and the first wrong answer. The first big twin study of schizophrenia, Franz Kallmann's in 1946, reported 691 index twins from New York state hospitals and a concordance in identical twins of 86% after age correction, which is where the textbook claim "schizophrenia is almost entirely genetic" came from.²² Kallmann decided zygosity by how alike the twins looked, diagnosed the co-twins himself knowing the index diagnosis, and drew from chronic hospital wards where the sickest, most familial cases collected. Each of those inflates the number. By 1972 Gottesman and Shields, working from a consecutive register in London with blind diagnoses and blood-group zygosity, had it at 42% pairwise (58% counting each affected twin as a proband), and 9% in fraternal twins.²³ The Maudsley series in 1999 and the Danish national register in 2018 landed near 40% and 33%.²⁴²⁰ The heritability stayed high throughout, because fraternal concordance fell too. The lesson is not that the early studies were worthless; it is that ascertainment, blinding and zygosity testing each move the number, and that a figure quoted without its method is not a figure.

Twins reared apart. The cleanest natural experiment, and the rarest. If identical twins are separated in infancy and raised in different homes, whatever still correlates between them cannot be the home, so the correlation is itself an estimate of heritability, with no formula. James Shields found 44 such pairs in Britain in 1962 by appealing on television.²⁵ The Minnesota Study of Twins Reared Apart, from 1979, assembled over a hundred pairs from around the world and reported in 1990 that separated identical twins correlated about 0.70 on IQ, close to twins raised together, and around 0.50 on most personality traits.²⁶ The famous "Jim twins," reunited at 39 with the same brand of cigarettes and the same name for their dogs, were from that study, and are the reason to be careful: with enough pairs and enough traits, striking coincidences are guaranteed. The sober numbers hold up; the anecdotes are anecdotes. The design's weaknesses are that "apart" often meant raised by relatives in the same town, and that separation was rarely from birth.

Adoption, and my two mentors. Twins are one natural experiment; adoption is the other, and the one that settled the schizophrenia argument. A child conceived by one set of parents and raised by another has genes from one home and an upbringing from a different one. If the illness follows the genes, it should turn up in the biological family and not the adoptive one. Leonard Heston did the first version in 1966: 47 children born to mothers with schizophrenia in Oregon state hospitals and placed in foster care within days, against 50 matched foster children; five of the first group developed schizophrenia and none of the second.²⁷ Then came the Danish studies, which used Denmark's complete adoption and psychiatric registers and were run by Seymour Kety, David Rosenthal and Paul Wender. Starting from adoptees who had developed schizophrenia, they traced both families: the illness and its milder cousins clustered in the biological relatives the adoptees had never lived with, and not in the adoptive relatives who raised them.²⁸²⁹ Starting instead from parents with schizophrenia whose children were adopted away, the children carried the risk with them.³⁰ And in the cross-fostering study that Wender led, children of healthy parents raised by an adoptive parent who became ill did not show elevated rates, which took the "schizophrenogenic mother" of mid-century psychiatry off the table.³¹ Wender did the same for mood disorders in 1986: the biological relatives of adoptees with major affective illness had far more depression and, strikingly, far more suicide than the adoptive relatives.³²

Paul Wender spent decades at the University of Utah, and he and Fred Reimherr were the two people who taught me clinical research, early in my career there. I learned from them how a trial is actually run: the rating, the blinding, the patient in front of you. Wender had by then turned from schizophrenia to the question that made him famous, whether the "minimal brain dysfunction" of childhood persisted into adult life, which he argued for in 1971 when almost no one believed it, and which is now called adult ADHD.³³³⁴ The Wender Utah criteria and the Wender Utah Rating Scale came out of that clinic,³⁵ and so did the Wender–Reimherr Adult Attention Deficit Disorder Scale, the clinician-rated interview that still carries both their names. I am an author on its psychometric validation paper, and so, in a nice piece of family genetics, is my mother.³⁶ The adoption studies are in every textbook; the people who taught me to do the work were sitting down the hall.

And then DNA. In 2010 a method arrived that needs no twins and no adoptees: take thousands of unrelated people, measure how slightly more genetically similar any two of them happen to be across hundreds of thousands of common variants, and ask whether the more similar pairs are also more similar in the trait.³⁷ That is the "SNP heritability" column in the ladder below, and it is the reason the twin estimates can be checked at last. It comes out lower, every time, and the gap is the next section.

3The ladder, condition by condition

Here are the twin heritabilities for the conditions I see, next to a few ordinary traits for scale, and next to two newer numbers that the next sections explain: the "SNP heritability" that can be measured from DNA alone, and how much a polygenic score can currently predict. The gap between the bars is the story of the field.

Sources for the chart, condition by condition: schizophrenia,²⁰³⁸ bipolar disorder,³⁹⁴⁰ autism,⁴¹⁴² ADHD,⁴³⁴⁴ anorexia,⁴⁵⁴⁶ alcohol use disorder,⁴⁷ OCD,⁴⁸ depression,⁴⁹⁵⁰⁵¹ PTSD,⁵²⁵³ generalized anxiety;⁵⁴ and for the traits, height,⁵⁵⁵⁶ BMI,⁵⁷ IQ,¹⁵ education,⁵⁸ personality.⁵⁹ One more number belongs here: conditions share genes. The common-variant correlation between schizophrenia and bipolar disorder is about 0.7, and the eight major psychiatric conditions overlap so much that over a hundred GWAS regions are shared across several of them.⁶⁰⁶¹ The diagnoses are drawn with sharper lines than the genome draws them.

Three patterns in that chart. First, the psychiatric conditions are all substantially heritable, and the neurodevelopmental ones (schizophrenia, bipolar disorder, autism, ADHD) are up with height. Depression, anxiety and PTSD are lower, in the range of personality, with a correspondingly larger role for what happens to people. Second, the DNA-based numbers are always smaller than the twin numbers, which is called the missing heritability, and is partly rare variants that common-variant methods cannot see, partly interactions, partly twin-study assumptions, and partly that we have not sequenced enough people yet. Third, the prediction bars are smallest of all. A polygenic score for depression, built from the largest genetic study ever done in psychiatry, explains about 6% of the liability.⁵¹ That is scientifically remarkable and clinically, for now, close to useless on its own, and both halves of that sentence matter.

4One gene, or a thousand

Here is the fork in the road. There are two completely different ways that genes cause a condition, and almost every confusion about psychiatric genetics comes from applying the logic of one to the other.

Mendelian. One gene, one broken copy (or two), and the condition follows, usually in a recognizable pattern through a family tree: Huntington's disease, cystic fibrosis, the Ogden syndrome in the next section. These are rare, each one, and the gene can be found in a single family by looking for the variant that tracks with the illness.

Polygenic. Thousands of common variants, each nudging risk by a percent or two, adding up with the environment to a liability that, past some threshold, becomes a diagnosis. This is nearly every common psychiatric condition. No single family can reveal the genes, because every family has a different handful of the thousands, and no one variant does anything you could see. Roll the dice.

There is also a middle. Some structural variants (a missing or duplicated chunk of a chromosome, called a copy-number variant) sit between the two worlds: rare, found in a small fraction of cases, but with a kick far larger than any common variant. Deleting a stretch of chromosome 22 (22q11.2) raises the risk of a psychotic disorder to roughly one in four, from a baseline of about one in a hundred.⁶² A handful of such CNVs together explain a few percent of schizophrenia, and when one is found, it changes the conversation with the family entirely.⁶³ Rare damaging variants in single genes have now been found for schizophrenia (ten genes, each rare, each with a several-fold effect) and for autism (over a hundred), many of them new in the child rather than inherited.⁶⁴⁶⁵ Those are the sixty new mutations from section 1, landing somewhere important.

5A gene-finding expedition: Ogden, Utah

The Mendelian world is where I started, and this is the case I use to show what it looks like when it works. In a family from Ogden, Utah, boys were dying in infancy. Five of them, across two generations, all on the mother's side, all with the same look: an aged face, a wrinkled forehead, prominent eyes, large ears, lax skin, little muscle tone, failure to grow, and a heart that eventually stopped. None lived past two.¹ The pattern, affected boys through unaffected mothers, is the signature of a gene on the X chromosome, which a boy gets only from his mother and has only one copy of.

Twenty years earlier, that pedigree would have been the end of the story. In 2010 it was the beginning. The team, led by Gholson Lyon, sequenced the protein-coding parts of the genome (the exome) in family members and asked a simple question: which rare variant is on the X, is present in every affected boy, is carried by every mother who had an affected son, and is absent from the men who were fine? The problem is that every person carries tens of thousands of variants, so the answer is a filtering problem, and a new program developed at the University of Utah, VAAST, did the ranking.⁶⁶ It came back with one change, a single letter, in a gene called NAA10, which adds a small chemical tag to the beginning of thousands of other proteins. The same letter was then found in an unrelated family in California with three boys who had died the same way.¹ We called it Ogden syndrome, after the town.

6How a GWAS is different from a disease gene

If no single family can reveal a polygenic condition, the only way to find the thousand variants is to compare enormous numbers of unrelated people with and without the diagnosis, position by position, and look for letters that are a little more common in the cases. That is a genome-wide association study. The first widely credited one, in 2005, found a variant for macular degeneration in a few hundred people, because that variant happened to have an unusually large effect.⁶⁷ Psychiatry had no such luck. The first schizophrenia GWAS found almost nothing, and the field learned the hard way that it needed not thousands but hundreds of thousands of participants. The 2022 schizophrenia study had 77,000 cases and 244,000 controls and found 287 regions of the genome.³⁸ The 2025 depression study had 689,000 cases and 4.4 million controls and found 697 associations.⁵¹

The results are drawn as a Manhattan plot: every variant along the genome on the x-axis, and on the y-axis how unlikely its association would be by chance. The skyline of towers is the finding. Here is a schematic one you can read.

Four things make a GWAS hit different from a disease gene. The hit is a region, not a gene, and often sits in regulatory DNA between genes, so working out which gene it affects is its own project. The effect is tiny: the strongest common variants for schizophrenia change the odds by perhaps 10–20%, versus the several-fold or near-certain effects of a Mendelian variant. The hit is common: typically carried by a third of the population, most of whom are fine. And the hit is only meaningful in aggregate, which is where polygenic scores come from.

Polygenic scores, and the liability threshold

Add up a person's risk letters, each weighted by its GWAS effect, and you have a polygenic score. Imagine the whole population's liability as a bell curve; past a threshold on the right, people meet criteria. A polygenic score shifts where you sit on the curve. Drag it.

The candidate-gene graveyard

Before GWAS, the field ran for fifteen years on candidate genes: pick a gene with a plausible story, test a variant in a few hundred people, publish. The most famous was a 2003 paper reporting that a variant in the serotonin transporter gene made people more likely to become depressed after stressful life events.⁶⁸ It was cited thousands of times, it was in every textbook, and it was the basis of an entire genre of "gene × environment" findings. When the question was finally asked in samples of hundreds of thousands, there was no support for it, or for any of the eighteen most-studied depression candidate genes.⁶⁹⁷⁰ The genes were not wrong about biology; the samples were simply far too small to see anything real, so what they saw was noise. I believed those papers. I cited them. The lesson generalizes well beyond genetics: an effect that needs 200 people to find and 200,000 to lose was never there.

7Cancer, two hits, and the loaded gun

Psychiatry borrowed its favorite metaphor from oncology, so it is worth seeing the original. In 1971 Alfred Knudson looked at children with retinoblastoma, a tumor of the eye, and noticed two populations: children with a family history, who got tumors early and often in both eyes, and children without, who got one tumor, later.⁷¹ His explanation, which turned out to be exactly right, was that the tumor needs both copies of a gene (RB1, as it was later named) to be knocked out in one cell. A child who inherits one broken copy is born one hit from a tumor in every cell; any retinal cell that takes a second hit becomes one. A child born with two good copies needs two hits in the same cell, which is rare, late and usually single. Two hits. Watch the arithmetic.

That is the model behind inherited cancer syndromes. A BRCA1 variant does not cause breast cancer; it removes one of the safeguards, so that the ordinary accidents of a lifetime are far more likely to finish the job: about 72% of carriers by age 80, against about 12% in the general population.⁷² A loaded gun, not a fired one. Most of the other 28% carried the same variant their whole lives and nothing happened.

Psychiatry borrowed the frame at the end of the 1990s as the "two-hit hypothesis" of schizophrenia: an early hit, genetic or prenatal, that leaves the developing brain vulnerable, and a later hit, often around adolescence, that tips it over.⁷³ The best experimental version is a mouse study in which neither a prenatal immune challenge nor adolescent stress alone did much, while the two together produced the lasting changes.⁷⁴ It is a hypothesis, not a law, and the "hits" in psychiatry are not two discrete events in one cell but a long accumulation of small nudges. But the shape of it holds up: the heritable part of most psychiatric conditions is a vulnerability, not a sentence.

The line I use, and the warning label I attach to it

"Genetics loads the gun; environment pulls the trigger." I say it in clinic, and I say it with disclaimers, because it is useful and it is wrong in three specific ways.

8What it means for you

Will my kids get it?

This is the question, and the honest answer is a table of modest numbers. Risk rises with how closely you are related to the person who has the condition and with how heritable the condition is, and in every row the most likely outcome for a relative is to be fine. Pick a condition and a relationship. The figures are drawn from the twin register, pooled family studies and sibling cohorts.^{20773949437845}

Two refinements to that table. Risk is higher when more than one relative is affected, when the relative's illness began young, and when the condition is one of the highly heritable ones. And risk is a lifetime figure for a condition that is treatable, which is a different thing from a prognosis.

What a genetic test can tell you today

• Diagnostic testing for neurodevelopmental conditions. For a child with autism, intellectual disability or multiple congenital anomalies, a chromosomal microarray and, increasingly, exome or genome sequencing are first-tier tests, with a diagnostic yield of roughly a third in intellectual disability and perhaps 10–20% in autism without other features.⁷⁹⁸⁰ Finding a 22q11.2 deletion or a damaging variant in a known gene changes medical surveillance, recurrence counseling and sometimes treatment. This is the Mendelian world, and it is the one place genetic testing is standard of care in psychiatry.
• Pharmacogenomics. Testing the liver enzymes that clear antidepressants (CYP2D6, CYP2C19) can explain why a drug failed or caused side effects at an ordinary dose, and the professional consensus is that it is reasonable to use with care and not to over-read. I wrote the long version in One size fits few.⁸¹
• Polygenic scores. Not clinically actionable yet for any psychiatric condition, per the field's own statement, and worse than that outside the European-ancestry populations most GWAS were done in, where they predict less and could widen disparities if used.⁸¹⁸²⁸³ Direct-to-consumer "depression risk" reports are built on them. Treat them as entertainment.
• Family history. Still the best genetic test we have for common conditions, free, and underused. Three generations, both sides, with ages of onset.

What heritability does not take away

A high heritability does not mean a condition is untreatable, unchangeable or anyone's fault. ADHD is about as heritable as height and responds to medication in weeks. Schizophrenia is 80% heritable and the single strongest predictor of outcome is how long psychosis goes untreated, which is entirely a property of the environment. The genes set the terrain. Everything a clinic does is weather.

This is education, not genetic counseling. Risk figures are population averages, rounded, and vary with the study; if you are weighing a decision, a genetic counselor can work through your own family's numbers with you. The Ogden syndrome account is from the published paper; I was one of many authors and the credit belongs to the families and to the team that led the work.

9The light bulbs, collected

1. Most of your cells carry no DNA at all, and the rest carry two meters each.
2. Heritability is about differences between people, not the makeup of one person, and it changes when the environment does.
3. High heritability within groups says nothing about differences between groups. Two gardens.
4. An identical twin of someone with an 80%-heritable illness usually does not get it.
5. "The gene for depression" does not exist, and that is the finding, not the failure.
6. A single family can reveal a single gene. A thousand small genes need a million people.
7. A GWAS hit is common, tiny and usually carried by healthy people; it only means something in aggregate.
8. A polygenic score is a forecast, not a verdict; most people at the top of the curve stay well.
9. The candidate-gene era was mostly noise, and I believed it too.
10. Two hits: inherited risk is a loaded gun, and most loaded guns never fire. The line needs its warning label.
11. Most people who become ill come from the middle of the risk grid, not the corner.
12. The best genetic test for common psychiatric conditions is still a careful family history.

References