The Medical Matchmaking Machine

Episode Summary

Executive Summary: Radiolab profiles Dr. David Fagenbaum, who survived repeated near-death crises from Castleman disease after first losing his mother to brain cancer, then identified a repurposed drug, sirolimus/rapamycin, that put him into long-term remission. He now uses machine learning to search 4,000 drugs across 18,000 diseases, aiming to uncover overlooked treatments while debating the risks and ethics of making such AI predictions public.

Main Topics: David Fagenbaum’s childhood, football dream, and mother’s death (Priority: 5/5): Fagenbaum’s identity was shaped by football ambition and by the loss of his mother to glioblastoma while he was in college, which redirected his life toward medicine and treating cancer-like diseases. Personal illness, diagnostic odyssey, and repeated relapses (Priority: 5/5): As a medical student, Fagenbaum developed multi-organ failure, spent 11 weeks without a diagnosis, learned he had idiopathic multicentric Castleman disease, and experienced repeated cycles of temporary remission and relapse. Finding a repurposed drug that saved his life (Priority: 5/5): By studying his own blood and immune data, he identified mTOR overactivation and matched it to sirolimus/rapamycin, which has kept him in remission for 11.5 years. Repurposing drugs for other patients (Priority: 4/5): Fagenbaum applied the same approach to a relative with angiosarcoma, finding PD-L1 expression that led to pembrolizumab treatment and remission, showing the broader potential of drug repurposing. Building AI to search medicine at scale (Priority: 5/5): His team created the MATRIX system, which scores every drug-disease pairing across 4,000 drugs and 18,000 diseases, using machine learning to identify promising repurposed treatments humans could not evaluate manually. Risks, off-label use, and public release of the algorithm (Priority: 5/5): Latif and David debate whether publishing the tool could empower patients or cause harm, including false hope, doctor-patient conflict, or misuse; David argues it should guide research, not direct self-treatment. Ethics of hope, dignity, and individualized decision-making (Priority: 4/5): The episode contrasts aggressive treatment and dying with dignity, ending on the idea that whether to pursue a repurposed treatment must ultimately be the patient’s case-by-case decision.

Key Arguments: Fagenbaum’s mother’s death gave him a lifelong mission to find treatments for devastating diseases; his own survival later intensified that mission. Rare-disease and refractory-disease patients often have some plausible existing drug option, but the real challenge is finding it in time. His remission was not due to a new molecule but to an existing FDA-approved drug repurposed based on biological signals in his own body. The medical system is biased toward one-drug/one-disease labeling because FDA approval, marketing restrictions, and pricing structures discourage systematic repurposing. Off-label prescribing is common in the U.S., representing a substantial share of real-world medicine, but evidence standards remain uneven. Machine learning can act as an idea generator across tens of millions of drug-disease possibilities, but human experts must validate any prediction before treatment. Publicly sharing AI scores could democratize discovery and accelerate research, but it also risks misuse by desperate patients and erosion of trust in clinicians. The right framework is not to let AI tell patients what to take now, but to help researchers and doctors discover what should be studied next.

Data Points: Mother’s diagnosis date: July 2003 - Fagenbaum’s mother was diagnosed with grade 4 glioblastoma. Mother’s survival after diagnosis: 15 months - She died on October 26, 2004. Average survival for glioblastoma: around 6 months - Doctors described the prognosis after her surgery. Longest survival mentioned for glioblastoma: around 5 years - The family was told this as a rare upper bound. Time to diagnosis for Fagenbaum’s illness: about 11 weeks - It took weeks to identify idiopathic multicentric Castleman disease. Number of relapse-remission cycles: 5 times in 3.5 years - His disease repeatedly came back after temporary improvement. Years in remission on sirolimus/rapamycin: 11.5 years - He said a few days before the interview that he had reached this milestone. Size of drug list in MATRIX: 4,000 drugs - The AI system evaluates all drugs against diseases. Size of disease list in MATRIX: 18,000 diseases - The AI system evaluates all diseases against drugs. Total scores generated by MATRIX: about 75 million - All drug-disease pairs are ranked by machine-learning output. Share of U.S. prescriptions that are off-label: about 20%–30% - Fagenbaum explained how commonly medicines are used outside their labeled indication. Clinical benefit cited for angiosarcoma patient: 99% PD-L1 positivity - A low-cost test found strong evidence for pembrolizumab use in his uncle’s cancer. Angiosarcoma remission duration mentioned: 9 years - His uncle Michael had been in remission by April of this year. COVID drug effect: 35% reduction in mortality - Latif and David discussed dexamethasone’s benefit in the pandemic. EveryCure public release timeline: about 9 months / 9 to 12 months - David said the algorithms would be made publicly available after more refinement.

Pivotal Quotes: "I want revenge. I want to do whatever I can, to take this thing on." — David Fagenbaum: He describes how his mother’s death from brain cancer transformed his motivation. "The drug that saved my life." — David Fagenbaum: He is referring to rapamycin/sirolimus, the repurposed medicine that put his Castleman disease into remission. "This is not a solution engine. This is an idea generator." — David Fagenbaum: He explains how the public AI tool should be understood and used.

Implications: The episode suggests AI could unlock hidden uses for existing drugs, especially in rare disease, but only if humans keep control, evidence standards stay high, and patient desperation does not turn prediction into prescription.