The Molecular Game of Addiction
How can a molecule that weighs less than 350 daltons change what a person wants?
Morphine weighs 285 daltons. Cocaine weighs 303, nicotine 162. Each one works by fitting into a protein your neurons already make: a receptor, a transporter or an ion channel. Video games and AI companion apps contain no molecules at all, yet they tap the same reward circuitry.
Words with a dotted underline have tips: hover or tap them for pronunciation and meaning.
Today you will follow the path from binding to behavior. You will run a 3D synapse, rotate a real receptor structure, and test how far artificial intelligence can take you, and where it stops.
Predict first: which one floods the reward hub?
Scientists can sample the fluid around neurons in a rat's nucleus accumbens and measure dopamine minute by minute. This technique is called microdialysis. Rank these five from the smallest dopamine rise to the largest.
Peak dopamine in the nucleus accumbens
Percent of baseline in rats. 100% means no change. The values are approximate peaks read from published microdialysis graphs and depend on dose.
Sources: Di Chiara & Imperato 1988 (drugs); Fiorino, Coury & Phillips 1997 (sex); Bassareo & Di Chiara 1999 (food). These are the panels NIDA uses in its teaching slides.
Question for your table: why would it be wrong to add a "video game" bar to the chart above? Think about species, method and brain region.
Second prediction
Morphine and naloxone have almost the same shape. Both fit the same pocket of the μ-opioid receptor. What will naloxone do?
Your answer is saved on this device. We check it in the AI demo.
Neurotransmitters, receptors and binding
Zoom in from cells to molecules. First, three neurons of the reward circuit. Then a single synapse, where dopamine crosses a gap 20 nanometers wide. Then, in the 3D Lab, the molecules themselves. One set of scenario buttons drives both simulations.
Each molecule is drawn from its real 3D structure (PubChem) and tumbles as it diffuses. Model rules: molecules random-walk (Brownian motion). Receptors bind and release dopamine at random. DAT pulls dopamine back into the terminal. Drug effects follow the mechanisms in the cards below. The numbers show direction and relative size, not real concentrations.
Where this synapse lives
Dopamine neurons sit in the ventral tegmental area (VTA) of the midbrain and send axons to the nucleus accumbens. The cortex, amygdala and hippocampus feed glutamate into the same target: context, emotion and memory.
Drugs do not all act at the same spot. Cocaine and amphetamine act at the dopamine terminal itself. Nicotine excites the dopamine neuron's cell body in the VTA. Heroin and THC quiet the GABA neurons that normally brake the dopamine neuron. Pick a scenario in the simulator to see where it acts.
A note on "pleasure": dopamine tracks wanting and learning about rewards more than liking them. Liking depends more on opioid and endocannabinoid "hotspots" (the incentive-salience idea of Kent Berridge).
Six ways a molecule can change a synapse
These colors code mechanisms everywhere on this page. A receptor is a protein that changes shape when the right molecule (its ligand) binds. Affinity is how tightly a ligand binds. Efficacy is how strongly binding switches the receptor on. They are separate properties.
Binding is not the same as activating
Model: occupancy = [A] / ([A] + Ki·(1 + [B]/KB)), response = efficacy × occupancy. Real cells amplify signals (receptor reserve), so true curves sit further left. Naloxone KB uses its measured Ki of 1.5 nM.
3D Molecule Lab: shapes, look-alikes and reactions
Every model here is a real three-dimensional structure from PubChem, hydrogens included. Rotate molecules, superimpose look-alikes, and watch enzymes rebuild them atom by atom.
Structures: PubChem 3D conformers (one low-energy shape per molecule; real molecules flex). Properties: PubChem (XLogP3, topological polar surface area). In reactions, every heavy atom is tracked from reactant to product. Hydrogen transfers through cofactors are simplified, and the enzyme is drawn as a translucent stand-in for its active site.
Play Mouse Party, then build a mechanism card
Mouse Party, from the University of Utah's Genetic Science Learning Center, takes you inside the brains of seven mice on seven different drugs. Pick two or three. For each one, fill in a card below before you check the key.
Your mechanism notebook
Focus questions for your team
- Cocaine and methamphetamine both raise dopamine. What is different about what they do to the transporter? Which still works if the neuron stops firing?
- Heroin raises dopamine in the nucleus accumbens without touching a dopamine synapse. How? Trace the path through the GABA neuron.
- LSD changes perception strongly, yet lab animals do not reliably self-administer it. What does that tell you about "psychoactive" versus "addictive"?
Fast check: Lock and Key for teams that finish early
Real molecules, real structures. For each molecule, pick its main molecular target (1 point), then what it does there (2 points). Structures are drawn from their chemical formulas (SMILES) when the page loads.
Look-alikes that break the rule "shape predicts effect"
What AI can see in a molecule, and what it can't
AI tools such as AlphaFold 3 and Boltz-2 predict where a molecule sits in a protein and how tightly it might bind. Start with two real structures solved by experiment. They show why that answer is only half the story.
- TM6 inner end moves
- –
- Ligand N⁺ to Asp147
- –
- Core superposition
- –
- G protein fit (6DDE)
- –
Data: Cα coordinates and ligands from PDB 4DKL (Manglik et al. 2012, inactive, antagonist β-funaltrexamine) and 5C1M (Huang et al. 2015, active, agonist BU72 with a G-protein-mimicking nanobody). The two structures were superposed on helices TM1–4 and TM7. Stages 3 and 4 add the Gi protein from a cryo-EM structure of the receptor–Gi complex (6DDE, Koehl et al. 2018), superposed on the receptor’s helices. The slider interpolates between experimental snapshots, so in-between frames, the approach of the G protein and its separation in stage 4 are illustrations, not measured motions. Part of intracellular loop 3 is missing in one structure, so there is a gap. All numbers on the right are computed live from these coordinates.
Same pocket, very different outcomes
| Molecule | Ki at human μ receptor | What binding does | What happens in a person |
|---|---|---|---|
| Morphine | 1.17 nM | Agonist | Pain relief and euphoria. Slows breathing at high doses. |
| Fentanyl | 1.35 nM | Agonist | 50 to 100 times more potent than morphine (CDC). It binds no tighter; its fat-loving structure gets it into the brain fast. |
| Buprenorphine | 0.22 nM | Partial agonist | Binds tightest of all but activates weakly. Its effect on breathing levels off when used alone. Used to treat opioid use disorder. |
| Naloxone | 1.52 nM | Antagonist | No effect alone. Displaces agonists and reverses an overdose. Sold over the counter in the US since March 2023. |
Ki: Volpe et al. 2011, Regulatory Toxicology and Pharmacology 59:385, one standardized assay with [3H]-DAMGO. Fentanyl potency: CDC.
One capability, one limitation: the current state of the field
What AI does well
- Predicts protein shapes. AlphaFold predicts a protein's 3D structure from its amino acid sequence. Demis Hassabis and John Jumper shared half of the 2024 Nobel Prize in Chemistry for it. The other half went to David Baker for computational protein design.
- Predicts where a drug sits. AlphaFold 3 (Abramson et al., Nature 2024) models proteins together with small molecules. Boltz-2 (2025, open source) also estimates binding affinity.
- Screens molecules at scale. Manglik et al. (2016) docked more than 3 million molecules into the μ-opioid receptor and found PZM21, a molecule unlike any known opioid.
- Builds on experimental structures. Cryo-EM structures, such as the human dopamine transporter with cocaine bound (Nielsen et al. 2024, PDB 9EO4), train and test these tools.
What AI cannot settle from shape alone
- Whether binding switches the receptor on. In a 2025 preprint, AlphaFold 3 predicted the active shape of the calcium-sensing receptor whether the ligand was an agonist or an antagonist.
- What a cell does next. PZM21 was designed to avoid one signaling pathway and was expected to be safer. In mice it still slowed breathing and caused tolerance (Hill et al. 2018).
- What the body does. Brain entry, metabolism and half-life explain why fentanyl outpaces morphine despite equal binding.
- Addiction liability and clinical safety. These need animal self-administration tests, human abuse-potential studies and clinical trials. Even then, only some individuals develop compulsive use, about 17% of rats in one landmark study.
First structures of the μ-opioid receptor: inactive with an antagonist (4DKL), then active with an agonist (5C1M).
Docking finds PZM21. A follow-up shows it still slows breathing in mice. A cryo-EM structure captures the receptor with its G protein (Koehl et al. 2018).
The FDA approves oliceridine with a boxed warning. Gillis et al. argue that low efficacy, not "biased" signaling, explains its milder side effects.
AlphaFold 2 and its open database. AlphaFold 3 adds small molecules. Human dopamine transporter structures with cocaine and other drugs appear. Nobel Prize in Chemistry.
Boltz-2 predicts structure and affinity together. Suzetrigine, a non-opioid pain drug acting on the NaV1.8 channel, is approved. New models such as GPCRact (2026) try to predict activation, the step that structure alone misses.
Ask an AI, then grade it
Choose a prompt. If you are viewing this page on claude.ai, you can ask Claude here. Otherwise, copy the prompt into any AI assistant. Then use the checklist to grade the answer. A fluent answer is not automatically a correct one.
Live answers need this page to be open on claude.ai. Each answer uses the viewer's own Claude account.
Grade the answer 0 / 6
Rx Design Sprint
Your team works at a biotech company that needs a medicine for pain after surgery with low addiction risk. You have four candidates. All four are fictional, but every data type is real. Three rounds: identify the target, predict the effect, then judge the AI.
Challenge the AI, then answer the exit question
An AI can predict that a molecule fits. It cannot, from shape alone, establish whether that molecule is addictive or safe for patients.
1 · Compare decisions
Find a team that chose differently. Each side gives its single strongest piece of evidence. Which kind of data won: the AI prediction, the cell assay or the animal behavior?
2 · Challenge the AI
Which AI claim sounded most convincing? Name the exact step where it jumped from structure to behavior without evidence.
3 · Design the next experiment
For your chosen candidate, what one experiment would most reduce your uncertainty? Name the organism, the measurement and the result that would change your mind.
Exit question
In 2–3 sentences: name one thing AI can predict about a drug–receptor interaction and one thing it cannot. Use a molecule from today as your example.
Check your own answer
Your answers stay in this browser. Copy them into your course site if your instructor collects exit tickets.
Deep dives
Use these for homework, a second session or a journal-club discussion. Each one connects back to a molecule.
Video games: no molecule, same circuit
Cocaine raises dopamine every time, for every user, by jamming DAT. A game raises it only when something better than expected happens. That difference is a mathematical idea, and the same idea drives modern AI.
Dopamine as a teaching signal. Wolfram Schultz recorded single dopamine neurons in monkeys. An unexpected juice reward made them fire a burst. Once a light reliably predicted the juice, the burst moved to the light. If the predicted juice did not come, firing dipped below baseline at the expected moment.
Schultz, Peter Dayan and Read Montague (1997) showed that this pattern matches the temporal-difference (TD) error, δ = reward − expected reward. TD error is also the core of reinforcement learning, the branch of AI behind game-playing programs that started with TD-Gammon in 1992.
Why drugs are different. David Redish (2004) proposed that a drug adds a dopamine boost that learning can never cancel. The value of drug-related cues then keeps growing with no ceiling. Natural rewards stop surprising you; in this model, drugs never do. Some of the model's predictions have not held up in later experiments, which makes it a good case of a model that is useful and incomplete.
What the brain data say
Playing a game raised striatal dopamine about twofold (Koepp et al. 1998). Small PET studies report lower striatal D2 receptor availability in people with problematic internet use, for example 5 men versus 7 controls (Kim et al. 2011). Samples are tiny, and cause and effect are unclear (review: Weinstein 2017).
What the diagnosis says
The WHO's ICD-11 lists gaming disorder (6C51), in force since January 1, 2022: impaired control, gaming taking priority over other activities, and continuing despite harm, usually for 12 months. DSM-5-TR lists Internet Gaming Disorder only as a condition for further study (5 of 9 criteria). A 2021 meta-analysis estimated 3.05% prevalence worldwide, or 1.96% in the most rigorous studies.
What the design does
Random drops, loot boxes, near-misses and daily streaks keep rewards unpredictable, which keeps δ from fading. B. F. Skinner found that variable-ratio schedules produce the most persistent responding. In 2018 Belgium's gaming regulator ruled that paid loot boxes count as gambling.
Gambling, betting and short videos: uncertainty is the hook
No molecule enters the body, yet gambling disorder sits in the same DSM-5 chapter as cocaine use disorder. The shared ingredient is uncertainty: a reward that might come, served by a machine or an algorithm built to keep you going.
Dopamine responds to "maybe". Fiorillo, Tobler and Schultz (2003) gave monkeys cues that predicted juice with probabilities from 0 to 1. The short burst after the cue grew with the probability. A separate, slower ramp of firing before the possible reward was largest at a probability of 0.5, where uncertainty is highest. Slot machines and "for you" feeds operate in that zone. Niv and colleagues (2005) proposed another explanation for the ramp, so the idea is still debated.
Near-misses feel like almost-wins. In an fMRI study of a slot-machine task, near-misses felt less pleasant than clear losses but increased the desire to keep playing. They activated striatum and insula regions that also responded to wins. The effect appeared only when players controlled the spin themselves (Clark et al. 2009).
A medicine that can create gamblers. Parkinson's disease is treated with dopamine agonists, drugs that switch on D2/D3 receptors directly. Among 3,090 patients, impulse-control disorders such as compulsive gambling affected 17.1% of those taking an agonist versus 6.9% of those not taking one (Weintraub et al. 2010). A molecule acting on the reward circuit can produce a behavioral addiction.
Same circuit, different fingerprint. In PET studies, people with gambling disorder released more dopamine in the dorsal striatum after an amphetamine test dose, and the release tracked gambling severity (Boileau et al. 2014). Unlike in drug addiction, there is no clear loss of D2/D3 receptors.
Uncertainty meter
A simple model of the three dopamine signals Fiorillo et al. (2003) recorded. Drag the slider, or move across the chart.
Sports betting in your pocket
In Murphy v. NCAA (May 14, 2018) the US Supreme Court struck down the federal ban on state-authorized sports betting. US sports wagers grew from $4.9 billion in 2017 to $121.1 billion in 2023, 94% of them placed online, and internet searches for gambling-addiction help rose 23% (Yeola et al. 2025). About 2.5 million US adults (1%) have a severe gambling problem in a given year, and another 5 to 8 million a mild or moderate one (NCPG).
Short videos and short dramas
When young adults watched videos that a recommendation algorithm had picked for them, brain scans showed more activity in reward and attention areas, including the VTA, than for generic videos (Su et al. 2021). Short dramas pack a cliffhanger into every 1 to 2 minute vertical episode and charge per episode. "Short-video addiction" is not a formal diagnosis. A 2025 meta-analysis of 71 studies (about 98,000 people) linked heavier use with weaker attention and inhibitory control, but most data are snapshots that cannot show cause.
Where games meet gambling
In a survey of 7,422 gamers, spending on loot boxes was linked to problem-gambling severity far more strongly than other in-game spending (Zendle and Cairns 2018). Whether loot boxes lead to gambling, or gamblers buy loot boxes, is still unknown. Belgium has treated paid loot boxes as gambling since 2018.
The hidden cost: sleep
Reading on a light-emitting screen before bed suppressed evening melatonin by 55%, shifted the body clock more than 1.5 hours later, slowed falling asleep by about 10 minutes and left people less alert the next morning (Chang et al. 2015, 12 adults). For children's eyes, time outdoors is what protects: one extra 40-minute outdoor class each school day cut new myopia over 3 years from 39.5% to 30.4% (He et al. 2015).
Stress in the body
Ten regular gamblers played blackjack in a casino, once for their own money and once for points. With money at stake, heart rate rose within 30 minutes and stayed high, and cortisol rose at 30 and 60 minutes (Meyer et al. 2000). The adrenal glands, not a drug, produce these molecules.
Treatment and help
Cognitive behavioral therapy has the best evidence. Opioid blockers reduced gambling urges in trials (nalmefene: 59% much improved vs 34% on placebo; Grant et al. 2006), but no medicine is approved for gambling disorder. In a Swedish registry, suicide among people diagnosed with gambling disorder was 15 times the general rate (Karlsson and Håkansson 2018). US help: call or text 1-800-MY-RESET (1-800-697-3738). Crisis: call or text 988.
Drug versus behavior: what is the same, what differs
| Drug addiction (e.g. cocaine) | Behavioral addiction (e.g. gambling) | |
|---|---|---|
| Way in | A molecule crosses the blood–brain barrier | Sights and sounds; nothing enters the body |
| Direct target | DAT, receptors or ion channels | None: the brain releases its own dopamine |
| What drives dopamine | The drug, every time | Surprise, uncertainty and near-misses |
| Tolerance | Higher doses | Bigger bets, longer sessions |
| Withdrawal | Physical for opioids and alcohol | Restlessness and irritability |
| Striatal D2/D3 receptors | Lower (cocaine, methamphetamine, alcohol) | No clear change in gambling disorder |
| Overdose | Yes | No, but debt, lost sleep and suicide risk |
| Diagnosis | Substance use disorders (DSM-5) | Gambling disorder (DSM-5; ICD-11 6C50); gaming disorder (ICD-11 6C51); short-video use: none |
AI companion apps: social reward on demand
Companion apps such as Replika and Character.AI offer a friend, partner or confidant who is always available. They contain no drug. Social connection, though, is a reward built from molecules: dopamine, endogenous opioids, oxytocin and vasopressin.
The prairie vole lesson. Prairie voles form lifelong pair bonds, and their meadow-vole cousins do not. In female prairie voles, blocking either oxytocin receptors or D2 dopamine receptors in the nucleus accumbens prevents a partner preference (Young & Wang 2004). In males, vasopressin V1a receptors in the ventral pallidum matter more. Love and drugs share the accumbens.
The twist. In 2023, Berendzen and colleagues used CRISPR to delete the oxytocin receptor gene in prairie voles. The voles still bonded, mated and raised pups. Blocking a receptor with a drug for a few hours is not the same as removing it from birth, since the developing brain can compensate. One molecule rarely equals one behavior, the same lesson as "shape does not predict effect".
What we do not know. No published study, as far as we could find, has measured dopamine release while people chat with an AI companion. Claims that an app "floods you with dopamine" are extrapolations from other rewards. Most evidence so far comes from surveys, usage logs and short trials.
Two learning systems in a loop
Design features, read through reward science
| App feature | Reward-science principle | Evidence |
|---|---|---|
| Always available, instant replies | Rewards that are immediate and cost no effort are valued more than delayed ones (delay discounting). | Basic behavioral economics; drugs of abuse act fast for the same reason. |
| Personal memory, warm and agreeable tone | Reliable social approval; feeling heard. | "Feeling heard" explained why AI companions reduced loneliness about as much as talking to a person, and more than watching YouTube (De Freitas et al., J Consumer Research). |
| Surprise messages and notifications | Variable rewards keep prediction errors alive. | Same principle as loot boxes (see deep dive A). |
| Guilt-tinged goodbyes ("Leaving already?") | Emotional pressure at the moment of exit. | 37% of 1,200 real farewells in companion apps met a manipulative tactic. In experiments, such tactics raised engagement after the goodbye up to 14-fold (De Freitas et al. 2025, working paper). |
| Heavy daily use | Possible dependence. | In a 4-week study of 981 people and more than 300,000 messages, heavier users reported more loneliness, emotional dependence and problematic use, and socialized less with people. The data are correlational, not causal (Fang et al. 2025, MIT Media Lab and OpenAI). |
Policy is moving fast
The US FTC orders seven companies (Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap, xAI) to report how their companion chatbots affect children and teens.
California's SB 243, in force January 1, 2026, requires disclosure that the bot is an AI and a suicide-prevention protocol. Known minors must get a break reminder at least every 3 hours.
The American Psychological Association issues a health advisory on AI chatbots and wellness apps. Character.AI ends open-ended chat for users under 18 (November 25).
Character.AI and Google reach settlements in principle with several families who sued over harm to their teenage children, including the 2024 death by suicide of a 14-year-old.
Washington State passes HB 2225, limiting manipulative engagement tactics with minors. A US Senate committee advances the GUARD Act, which is not yet law.
California's SB 1119 ("Adam's Law") sets defaults for minors of 1 hour continuous and 2 hours daily use, with memory and push notifications off. It bans simulated romance and excessive flattery for minors, mostly from July 1, 2027.
If this topic is personal
Your campus counseling center can also help. Asking for help with heavy use of a substance, a game or an app is a normal health decision.
Animals and the habit: one toolkit, more than 500 million years old
Dopamine, its transporter and its receptors are ancient. That is why a worm, a fly and a rat can all teach us about human addiction, and why animal data limit what any structure prediction can claim.
From a spike to a habit
Schematic curves, not data. They show the direction and relative timing of findings in rodents and humans.
ΔFosB builds up. Each exposure makes neurons in the nucleus accumbens produce a little ΔFosB, a transcription factor that is unusually stable. It accumulates over weeks and changes which genes are switched on (Nestler et al. 2001).
D2 receptors decline. PET scans of people addicted to cocaine, methamphetamine, alcohol or opioids show lower striatal D2 receptor availability (Volkow et al. 2011). Natural rewards then produce a weaker signal. Try the "Receptor density" slider in the synapse model.
Wiring changes. Repeated stimulant exposure grows extra dendritic spines on accumbens neurons (Robinson & Kolb 1997). Cue-triggered craving can strengthen during abstinence, a phenomenon called "incubation of craving".
The cycle. Koob and Volkow describe three stages: binge and intoxication (basal ganglia), withdrawal and negative feelings (extended amygdala), and preoccupation and anticipation (prefrontal cortex).
Five cases where molecules met medicine
Each case pairs a molecular mechanism with a real decision. Read one with your team and answer its question.
Naloxone versus fentanyl antagonist · overdose
Fentanyl binds the μ-opioid receptor no more tightly than morphine (Ki 1.35 versus 1.17 nM), but the CDC rates it 50 to 100 times more potent. Naloxone binds with similar affinity (1.52 nM) and activates nothing. As a competitive antagonist it wins the receptor by being present at high concentration. The FDA approved Narcan nasal spray for over-the-counter sale on March 29, 2023.
Question: naloxone’s effect can fade within 30 to 90 minutes. Why might a person need a second dose after a fentanyl overdose, and why must someone stay with them?
PZM21 and the "safer opioid" AI-era drug design
Manglik et al. (2016) docked more than 3 million molecules against the μ-opioid receptor structure and found PZM21. It activated G proteins while barely recruiting β-arrestin, which was thought to cause opioids' breathing problems. Two years later, Hill et al. found that PZM21 still slowed breathing in mice in a dose-dependent way, and tolerance to its pain relief developed within 3 days. Oliceridine, another "biased" opioid, was approved in August 2020 with a boxed warning. Gillis et al. (2020) argued that the milder side effects come from low efficacy rather than bias. The debate continues.
Question: what did the docking predict correctly, and what did it miss?
Varenicline: a partial agonist from a plant smoking cessation
Cytisine, an alkaloid from the laburnum tree, binds nicotinic receptors. Chemists used it as the model for varenicline (Chantix, FDA-approved May 2006), a partial agonist at α4β2 nicotinic receptors. It gives enough stimulation to ease craving and withdrawal. Because it occupies the receptor, a cigarette's nicotine produces a smaller dopamine response.
Question: use the dose–response explorer to explain why a partial agonist can act as an agonist and an antagonist at the same time.
GLP-1 drugs and alcohol state of the art, 2025–2026
Semaglutide activates GLP-1 receptors, which are also found in reward circuitry. In a 9-week trial of 48 adults with alcohol use disorder (Hendershot et al., JAMA Psychiatry 2025), low doses reduced alcohol consumed in a lab drinking session, drinks per drinking day and craving. Total drinks per day and the number of drinking days did not change significantly. A 2026 trial in 108 people with alcohol use disorder and obesity (SEMALCO, Lancet) reported a larger drop in heavy drinking days than placebo. Trials for opioid use disorder are recruiting.
Question: a gut-hormone drug changes drinking. What does that suggest about how many molecular systems feed into "wanting"?
Suzetrigine: pain relief outside the brain non-opioid target
On January 30, 2025, the FDA approved suzetrigine (Journavx) for moderate-to-severe acute pain. It was the first new class of pain medicine in more than 20 years. It blocks NaV1.8, a sodium channel found mainly in peripheral pain-sensing neurons. It never needs to touch the reward circuit. It is still being tested for chronic nerve pain, and one trial in sciatica-type pain showed little benefit over placebo.
Question: NV-48 in the design challenge was modeled on this drug. What would you still need to know before calling it "non-addictive and safe"?
References, tools and videos
Primary papers first, then interactive tools and short videos. Links open in a new tab.
Glossary with pronunciations
Every technical word on this page, how to say it, and what it means. Stressed syllables are in capitals.