One Man’s Garage Lab and ChatGPT Yield a New Schizophrenia Candidate

Douglas Yao posted a photo of a small vial filled with yellow powder. The former Harvard computational biology PhD said he synthesized the contents himself on folding tables in his garage. The compound, PAC-3310, came from a formula shaped by ChatGPT. Yao described it as a selective M4 muscarinic receptor agonist aimed at schizophrenia, modeled on the recent breakthrough drug Cobenfy but with sharper precision.

The announcement landed on X this week and quickly drew millions of views. Some called it inspiring. Others questioned the wisdom of self-taught chemistry producing a potential antipsychotic. Yet the claim rests on data. Yao’s company, Pace Pharmaceuticals, published functional assays showing nanomolar potency at the M4 receptor and more than 100-fold selectivity over M1, M2, M3 and M5 subtypes. In mice, the molecule reduced MK-801-induced hyperlocomotion, a standard test for antipsychotic activity. No obvious side effects appeared even at higher doses.

Yao isn’t the first to chase muscarinic pathways. Nature reported in 2024 that Cobenfy, the combination of xanomeline and trospium approved by the FDA the previous September, marked the first new schizophrenia mechanism in decades. Unlike traditional dopamine blockers, it targets M1 and M4 receptors in the brain. Researchers immediately saw potential beyond psychosis. Early signals suggested similar compounds might ease cognitive decline or agitation tied to Alzheimer’s disease. About half a dozen drugs in this class now sit in trials for psychiatric and neurological conditions.

But success remains uncertain. One highly anticipated schizophrenia candidate failed in Phase 3 trials shortly after Cobenfy’s approval. Regulators demand rigorous safety data for any central nervous system drug. Yao’s project, built outside traditional industry walls, raises fresh questions about oversight, manufacturing standards and long-term toxicology.

He earned his doctorate studying computational biology. Then he turned to artificial intelligence for molecule design. Over the past year, Yao says he used large language models to generate several thousand new small molecules. He taught himself synthetic organic chemistry. He assembled basic lab equipment at home. He produced roughly 100 compounds and began testing them in cell lines and mice. One of those efforts targets Alzheimer’s. Another became PAC-3310.

“It was designed by ChatGPT,” he wrote in the viral post. He followed with details on receptor selectivity and mouse behavioral results. The thread linked to supplementary data on GitHub, including raw plate-reader outputs, analysis code and an OT-2 liquid-handling protocol. Pace Pharmaceuticals lists both PAC-3310 for schizophrenia and a separate GalR1 antagonist called PAC-832 for Alzheimer’s on its site. The company says both have completed discovery and entered IND-enabling studies. Human testing could come within a year if funding and regulators align.

Experts have mixed reactions. Some praise the acceleration that AI and accessible robotics bring to early discovery. Others worry about corners cut when a single founder handles design, synthesis, testing and promotion. Yao discloses that he founded Pace and holds the patent on PAC-3310. His father, Dongyuan Yao, appears as a co-author on the technical report. The work received some support from BioCurious, a community lab that provided reagents and equipment access.

This isn’t pure garage improvisation. Yao integrates AI at every step. The models generate and score analogs. They help design assays. They program robots to run plates around the clock. They analyze results and suggest the next iteration. The approach compresses timelines that once required large teams and deep budgets. Similar strategies appear in established biotech. Companies such as PsychoGenics have used machine learning platforms to advance schizophrenia candidates like ulotaront into late-stage trials.

Yet PAC-3310 stands out for its origin story. A selective M4 agonist could avoid the gastrointestinal and cardiovascular side effects that plague less specific muscarinic drugs. Cobenfy requires a companion peripheral antagonist to manage those issues. A truly clean M4 molecule might offer cleaner efficacy with fewer complications. Yao’s mouse data hint at that advantage. The compound cut hyperlocomotion by nearly 67 percent at 30 milligrams per kilogram. Higher doses produced none of the typical non-selective muscarinic effects.

The overlap between schizophrenia and Alzheimer’s research grows clearer each year. A July 2026 paper in Molecular Psychiatry found shared genetic architecture between the two disorders, pointing to loci such as 16p11.2 and pathways tied to synaptic signaling and axonal growth. Psychosis appears in many Alzheimer’s patients and predicts faster decline. Antipsychotics developed for younger schizophrenia patients often carry heightened mortality risks in older adults with dementia. A safer muscarinic option could address unmet needs in both populations.

Recent work adds urgency. On the same day Yao’s post gained traction, Psychiatric Times highlighted a Northwestern University study identifying reduced soluble α2δ-1 in cerebrospinal fluid of schizophrenia patients. A synthetic peptide based on that molecule restored circuit balance and behavioral function in mouse models without sedation. The findings underscore how many biological levers beyond dopamine remain underexplored.

Yao’s project sits at the intersection of these trends. He didn’t invent the muscarinic approach. He refined it with AI, built the molecule at home and generated preclinical proof-of-concept. The yellow powder in that vial represents both promise and risk. If PAC-3310 advances, it could validate a new model of drug creation that bypasses some corporate gatekeepers. If problems emerge, it could fuel calls for tighter rules on DIY biotechnology.

Regulatory pathways remain demanding. IND-enabling studies require formal manufacturing under current good practices, detailed toxicology in multiple species and independent quality controls. Yao’s team will need capital to scale production and run human trials. Pace Pharmaceuticals describes itself as a new kind of drug company that deeply integrates AI and robotics. Whether that model sustains through clinical development will test the limits of garage-to-clinic ambition.

For now the data look intriguing. Cell assays confirm selectivity. Mouse behavior shows efficacy. The GitHub repository offers transparency that many academic papers lack. And the story captures a larger shift. Artificial intelligence no longer simply analyzes data. It proposes new chemical structures that humans then synthesize and test. One person with the right skills and equipment can now move from idea to animal proof far faster than before.

Questions linger. How reproducible are the results? Will PAC-3310 maintain its clean profile in humans? Could an even better M4 agonist emerge from traditional labs or competing AI platforms? Yao continues his work. He has posted about additional compounds and invited scrutiny of his methods. The scientific community will decide whether this garage experiment marks an outlier or the start of something broader.

Schizophrenia still disables millions. Current medicines help many but leave others with residual symptoms or harsh side effects. Any credible new option deserves examination. PAC-3310 may or may not reach patients. Its real contribution could lie in demonstrating how far accessible tools have come. A vial of yellow powder, born from code and chemistry in a suburban garage, now sits on the edge of serious drug development.