270 ‒ Journal club with Andrew Huberman: metformin, power of belief, & how to read scientific papers

KEYPOINTS

• 01

  Metformin weakly blocks the cell's power plant, and its main job in diabetes is to rein in the liver's glucose output.

  In this journal-club format, Peter explains that metformin, a generic drug used for 40–50+ years, first in Europe, partially inhibits complex I of the mitochondrial electron transport chain. Cyanide, by contrast, shuts down a later step completely. Why this helps diabetics is debated. What's clear is that it lowers hepatic glucose output. A healthy person has only about 5 g of glucose in circulation. Consistently carrying one extra teaspoon of it is roughly the difference between normal metabolism and type 2 diabetes. Insulin resistance starts mostly in muscle, which stores 75–80% of glucose. Insulin signaling moves glucose transporters to the cell surface, and fat building up inside muscle cells disrupts that signal, work Peter credits largely to Gerald Shulman at Yale.

• 02

  A 2014 study made metformin look like an anti-aging drug, but Peter later found a key flaw.

  Bannister and colleagues compared UK patients with type 2 diabetes on metformin alone with matched non-diabetic controls. Crude death rates were 14.4 versus 15.2 per 1,000 person-years, and modeling showed a 15% relative reduction in all-cause mortality over a median 2.8 years. Diabetes normally shortens life by about six years, so this was striking. Around 2018 Peter noticed 'informative censoring': patients who stopped metformin, were lost to follow-up, or progressed to stronger drugs were dropped. His analogy is a smoking study that stops counting smokers who die.

• 03

  A larger 2022 Danish study using twins found the opposite.

  Keys and colleagues sampled about half a million people from Danish registries. They compared 7,842 metformin users with diabetes against matched non-diabetics, and 976 same-sex twin pairs discordant for diabetes. Crude mortality was about 24.9 versus 16.9 per 1,000 person-years in the singletons, and about 24.7 versus 12.9 in the twins. Hazard ratios ran from 1.48 unadjusted to about 1.32–1.33 after adjusting for medications, marital status and education. In the twins they ran from 2.15 to about 1.7–1.8. Censoring like Bannister's only nudged these down, to 1.39 and 1.97. It never produced a benefit.

• 04

  What the Danish result does and doesn't show.

  Peter says it shows metformin doesn't erase the extra mortality of diabetes. It doesn't prove metformin is useless: those patients might have fared even worse on nothing. The groups' medication use couldn't be matched (lipid-lowering drugs 45.6% versus 15.4%, for example). Only a randomized trial like TAME (Targeting Aging with Metformin) can settle it. In the rigorous mouse Interventions Testing Program, metformin failed while rapamycin, 17-alpha estradiol (in males), acarbose and the SGLT2 inhibitor canagliflozin extended life. Peter still prescribes metformin for insulin resistance but doesn't see it as a longevity tool for insulin-sensitive people who exercise.

• 05

  Peter quit metformin after seeing his resting lactate.

  He started at a full 2 g a day in May 2011 without titrating up, and was so nauseous at the Berkshire Hathaway meeting he couldn't eat ice cream. In 2018, finger-prick testing for zone 2 training showed a fasted resting lactate of about 1.6 mmol/L. Healthy is below 1, roughly 0.3–0.6. That fits a mild mitochondrial toxin pushing glucose toward lactate. He also suspects metformin blunts gains in muscle size and strength. The only glucose-lowering drug he takes now is an SGLT2 inhibitor, based on mechanistic data, the mouse results and human trials including heart failure in non-diabetics.

• 06

  Statistical tools you need to read papers.

  Rates per 1,000 person-years normalize for follow-up time. A 95% confidence interval that excludes no effect signals significance. Kaplan-Meier survival curves only go down or stay flat, but they can cross. A hazard ratio minus one gives the relative change: 1.48 means 48% higher risk and 0.85 means 15% lower. Power matters: a null result only rules out effects the study was sized to detect. Overlapping error bars don't automatically mean no difference. Epidemiology makes up for not randomizing with huge samples and long follow-up. A randomized trial like the Women's Health Initiative can cost a billion dollars.

• 07

  Huberman's paper: being told a bigger dose produced a bigger brain response.

  In a not-yet-peer-reviewed preprint from Xiaosi Gu's lab at Mount Sinai (first author Ofer Perl), experienced smokers who had abstained vaped the same low-nicotine pen, confirmed by blood levels, while told it was low, medium or high strength. Their perceived effect matched what they were told. Thalamic activation in regions rich in nicotinic receptors scaled with belief, significantly between the low and high groups, during a market-prediction task in the fMRI scanner. Signaling from the thalamus to the ventromedial prefrontal cortex rose significantly at each step. Reward-pathway activity didn't change. Huberman calls it the first study of dose-dependent belief effects, while noting it lacked a zero-nicotine arm and had movement artifact from coughing smokers.

• 08

  Belief effects aren't the same as placebo.

  Placebo is binary: you think you got the drug or you didn't. Belief effects scale with the information people get. Huberman cites Alia Crum's work. People taught that stress enhances performance did about as much better as those taught it impairs performance did worse, very roughly 10–30% each way. A shake labeled high-calorie suppressed the hunger hormone ghrelin more than the identical shake labeled light. Hotel workers told their work was exercise lost more weight without measurably moving more. Telling patients that side effects mean a drug is working lessened how bad the side effects felt.