How the Body Switches Off Its Own Cannabinoids, and What Happened When a Drug Company Tried to Stop It

Fatty acid amide hydrolase, FAAH, hydrolyzes anandamide into arachidonic acid and ethanolamine. Monoacylglycerol lipase, MAGL, hydrolyzes 2-arachidonoylglycerol into arachidonic acid and glycerol and accounts for roughly 85 percent of 2-AG hydrolase activity in the brain, with ABHD6 and ABHD12 handling most of the remainder.

Because these enzymes terminate signaling rather than initiate it, blocking them amplifies endocannabinoid tone only where and when endocannabinoids are already being released. That is the theoretical advantage over a direct receptor agonist, and it is why the field pursued these targets so hard.

Patients and clinicians ask why cannabis produces such variable effects. Part of the answer lives in this pathway: how quickly a person clears their own endocannabinoids sets the baseline that any exogenous cannabinoid acts on top of.

The drug development story matters for a different reason. It is a clean, documented case of a well-reasoned mechanism producing catastrophic harm because of what a specific molecule did outside its intended target, and it explains why enthusiasm about modulating this pathway should be tempered rather than assumed.

Benjamin F. Cravatt and colleagues at The Scripps Research Institute cloned the enzyme now called fatty acid amide hydrolase and reported it in Nature in 1996. They had originally characterized it as oleamide hydrolase, then showed it also converts anandamide to arachidonic acid, and renamed it in recognition of the family of fatty acid amides it accepts as substrates.

That breadth is important and often glossed over. FAAH does not exclusively handle anandamide. It also degrades other N-acylethanolamines, which means inhibiting FAAH raises several lipid mediators at once, only some of which act at cannabinoid receptors.

Monoacylglycerol lipase was identified as the principal 2-AG degrading enzyme by T. P. Dinh and colleagues in the laboratory of Daniele Piomelli at the University of California Irvine, published in the Proceedings of the National Academy of Sciences in 2002. They cloned it from a rat brain library, found its messenger RNA concentrated in regions rich in CB1 receptors including hippocampus, cortex, anterior thalamus, and cerebellum, and showed a laminar distribution in hippocampus consistent with presynaptic localization.

Jacqueline L. Blankman, Gabriel M. Simon, and Cravatt later quantified the division of labor using functional proteomics in mouse brain, publishing in Chemistry and Biology in 2007. About 85 percent of brain 2-AG hydrolase activity is attributable to MAGL. Most of the remaining 15 percent is handled by two enzymes that had not previously been characterized, ABHD6 and ABHD12, which occupy different subcellular compartments and may therefore control different pools of 2-AG.

A direct CB1 agonist acts everywhere the receptor exists, at a concentration set by the dose rather than by physiology. That produces analgesia and also intoxication, tachycardia, and cognitive effects, which is the familiar trade-off with THC.

Endocannabinoids behave differently. They are synthesized on demand in response to local signals, act briefly, and are destroyed. Blocking their destruction extends signaling only at sites already generating it, in principle producing a physiologically targeted effect without global receptor activation.

Human genetics supports the idea that this pathway carries real behavioral weight. Iva Dincheva and colleagues, publishing in Nature Communications in 2015, built a knock-in mouse carrying the human FAAH C385A variant and compared it with human carriers of the same allele. Reduced FAAH expression associated with the variant allele tracked with enhanced fronto-amygdala connectivity, better fear extinction learning, and lower anxiety-like behavior in both species.

So the premise was reasonable, and the first clinical results were reassuring rather than alarming. Which is what makes the 2016 event so instructive.

Anne Kerbrat and colleagues at Rennes University Hospital reported the case series in the New England Journal of Medicine in November 2016. BIA 10-2474 is an orally administered reversible FAAH inhibitor that had been given to 84 healthy volunteers in sequential cohorts, single doses from 0.25 to 100 mg and repeated doses of 2.5 to 20 mg for ten days, with no severe adverse events reported.

A further cohort was then assigned to placebo, two participants, or 50 mg of BIA 10-2474 per day, six participants. An acute and rapidly progressive neurologic syndrome developed in three of the four actively treated participants who consented to have their data included, beginning on the fifth day of dosing. The clinical features were headache, a cerebellar syndrome, memory impairment, and altered consciousness.

Magnetic resonance imaging showed bilateral and symmetric cerebral lesions, including microhemorrhages and hyperintensities on fluid-attenuated inversion recovery and diffusion-weighted sequences, predominantly involving the pons and hippocampi. One participant became brain dead. Two improved, one with residual memory impairment and one with a residual cerebellar syndrome. One remained asymptomatic. The authors stated plainly that the underlying mechanism of the toxic cerebral syndrome was unknown.

Regulatory safety pharmacology studies conducted before the trial, later published in full, had found no adverse effects on major organ systems in rats and dogs. The authors of that publication noted that their data shed no further light on what caused the accident.

This is the part that rescues the field’s scientific credibility and that clinicians should know. Other FAAH inhibitors tested in humans had not produced anything resembling this, which pointed toward something specific to this molecule rather than to the target.

Annelot C. M. van Esbroeck and colleagues, working with Mario van der Stelt at Leiden and with Cravatt at Scripps, applied activity-based protein profiling to map what BIA 10-2474 actually binds in human cells and tissues. Their report in Science in 2017 found the drug inhibits several lipases that are untouched by PF-04457845, a highly selective and clinically tested FAAH inhibitor. In cultured human cortical neurons, BIA 10-2474 produced substantial alterations in lipid networks while PF-04457845 did not.

A follow-up proteome-wide analysis led by Zhen Huang at Pfizer with Cravatt’s group, published in ACS Chemical Biology in 2019, extended the finding. Des-methylated metabolites of BIA 10-2474 covalently modify the conserved catalytic cysteine of aldehyde dehydrogenases, including ALDH2, an enzyme implicated in protecting the brain from oxidative stress related damage.

The conclusion reached across these studies is that promiscuous lipase inhibition, plus off-target covalent modification by metabolites, can produce metabolic dysregulation in the nervous system. The lesson is about selectivity and about metabolites, not about the merits of raising anandamide.

PF-04457845 is the selective FAAH inhibitor against which BIA 10-2474 was compared, and its clinical record is instructive in a different way. John P. Huggins and colleagues at Pfizer tested it in a randomized, placebo and active controlled trial in pain due to knee osteoarthritis, published in Pain in 2012. The drug reduced FAAH activity by more than 96 percent and substantially raised four endogenous fatty acid amides. It produced no analgesic separation from placebo, while naproxen in the same trial did. The study was stopped at interim analysis for futility.

The authors described the finding carefully: the lack of analgesic effect of FAAH inhibition in humans contrasts with data from animal models, and that disconnect between species needs further study. It remains one of the clearest translational failures in the field.

The same compound did produce a signal elsewhere. Deepak Cyril D’Souza and colleagues at Yale ran a phase 2a randomized, double-blind, placebo-controlled trial in 70 men with cannabis dependence, published in The Lancet Psychiatry in 2019. PF-04457845 reduced cannabis withdrawal symptoms during the inpatient phase and lowered self-reported cannabis use and urinary THC-COOH concentrations at four weeks. There were no serious adverse events.

On the MAGL side, Kirsten R. Müller-Vahl and colleagues tested Lu AG06466, previously ABX-1431, a selective MAGL inhibitor, in a phase 1b crossover study of 20 adults with Tourette syndrome, published in Pharmacopsychiatry in 2022. A single 40 mg dose produced statistically significant reductions on two of three tic scales at various timepoints and reduced premonitory urges. The common adverse events were headache, somnolence, and fatigue. A 20-patient single-dose exploratory study is a beginning, not a result.

The biochemistry is about as settled as biology gets. Enzyme cloning, substrate identification, regional and subcellular localization, and functional proteomic quantification of hydrolase activity have been replicated across laboratories over three decades. The description of what FAAH and MAGL do is not in dispute.

The clinical evidence for modulating this pathway is early and mixed. One randomized trial showed near-complete FAAH inhibition producing no analgesia, another showed a real effect on cannabis withdrawal, and a 20-patient single-dose crossover suggested MAGL inhibition may reduce tics. None of that constitutes an established therapy, and the safety record contains one catastrophic event.

The most seductive claim in this area is that enhancing endogenous signaling is inherently safer than receptor agonism. The knee osteoarthritis trial undercuts the efficacy half of that claim, and BIA 10-2474 undercuts any assumption that this class carries a benign safety profile by virtue of its mechanism.

The FAAH C385A findings are frequently used to argue that raising anandamide reduces anxiety in people. A knock-in mouse plus imaging and behavioral measures in human variant carriers is a strong translational design, and it still describes a lifelong genetic difference rather than the effect of a drug given to an adult.

The Tourette result rests on 20 patients receiving a single dose in a crossover design, with two of three tic scales reaching significance at various timepoints. That is an exploratory finding, and the investigators described it as such.

This literature does not show that FAAH or MAGL inhibition treats pain in humans. The one adequately designed analgesia trial was negative despite confirmed target engagement above 96 percent.

It does not show that BIA 10-2474 harmed volunteers by inhibiting FAAH. The proteomic work points to off-target lipase inhibition and to covalent modification of aldehyde dehydrogenases by metabolites, and the original clinical report stated the mechanism was unknown.

It does not show that cannabis use meaningfully alters FAAH or MAGL activity in a way that changes clinical response, and it does not support any consumer product claim about supporting or boosting endocannabinoid enzymes.

Endocannabinoid degradation sits alongside the synthesis and receptor arms of the system, and it is the arm that determines duration. Clinically, that makes it the part of the system most likely to explain between-person variation in baseline endocannabinoid tone, which is the background any exogenous cannabinoid acts against.

The drug development arc here is a recognizable one in pharmacology: a clean mechanistic rationale, promising animal data, a failed efficacy trial, and a safety disaster traced to properties of one molecule rather than to the target. Dual FAAH and MAGL inhibition continues to be pursued, which our coverage of current dual inhibitor development takes up.

I teach this pathway to anyone who wants to understand why cannabis affects two people so differently. Your receptors matter, but so does how fast you clear the cannabinoids you make yourself, and that clearance runs through two enzymes with a genuine genetic variant behind one of them.

The BIA 10-2474 trial is the part I insist people learn. Six healthy volunteers took a drug with a sensible mechanism and a clean preclinical package. Three developed a progressive brainstem and hippocampal syndrome starting on day five, and one of them did not survive it. It took a year of activity-based proteomics to establish that the molecule was hitting lipases and aldehyde dehydrogenases it was never designed to touch.

The takeaway is not that this pathway is dangerous. It is that mechanism is not safety, and preclinical toxicology is not safety either. When someone tells me a compound must be safe because it works with the body’s own system, this is the story I think of.

FAAH degrades anandamide into arachidonic acid and ethanolamine. MAGL degrades 2-AG into arachidonic acid and glycerol and accounts for about 85 percent of brain 2-AG hydrolysis, with ABHD6 and ABHD12 covering most of the rest. Inhibiting these enzymes is a reasonable strategy that has so far produced one clear efficacy failure in pain, one positive result in cannabis withdrawal, an early tic signal, and one phase 1 catastrophe traced to off-target activity rather than to the target itself.

Carry forward the biochemistry, which is reliable, and the distinction between a target and a molecule. BIA 10-2474 failed because of what that specific compound and its metabolites bound, not because raising anandamide is inherently harmful. Do not carry forward the assumption that enhancing endogenous signaling is automatically safer or more effective than direct receptor agonism, because the trial record does not support either half of that assumption.