Remixing music to treat Alzheimer's

It turns out if you flash a light at a mouse 40 times a second, its neurons die less.

This came out of Li-Huei Tsai and Ed Boyden's labs at MIT, and a company called Cognito Therapeutics has taken it all the way to a 670 patient pivotal trial in Alzheimer's, results from which will be out later this year. In their earlier, smaller trial, an hour a day of flickering light and clicking sound slowed brain shrinkage by about 69% vs. sham. There are many caveats (they missed their primary efficacy endpoint!), but I think it may become one of the most exciting therapies for neurodegenerative conditions.

Cognito's version is a set of goggles and headphones, which flicker and play a "click" sound at 40hz. TLDR: I think there's a version of this that's just music, which could reach a lot more people, and perhaps even become a widely deployed prophylactic. And I think we could find out whether it helps for ~$5m.

In 2016, Tsai's lab saw that mice engineered to get Alzheimer's had weak gamma activity (the ~40hz brain rhythm associated with attention and memory). So they tried to put the rhythm back.

They did this with optogenetics, the technology Karl Deisseroth just shared this year's Nobel for. Specific neurons are genetically engineered to make a light-sensitive protein (borrowed from algae), allowing you to turn those neurons on and off by shining light on them. That means getting a gene into the brain and running an optical fiber in through the skull, so it's very much a mouse thing. They used it to make a set of neurons in the hippocampus fire at 40hz for an hour, and amyloid levels dropped. Other frequencies didn't do much.

An interesting therapeutic in mice, but good luck getting optogenetics into a person! Maybe Merge will someday manage an ultrasound version of this in people, but until then, mapping this 1:1 to humans seems pretty implausible.

Luckily, it turns out you might not need to. Your visual system tracks a flickering light, so a light flashing 40 times a second can drive rhythmic activity in the visual cortex without surgery, and reduce amyloid there in mice. Later work found that combining light and sound reduced amyloid across more brain regions, while sound alone could evoke hippocampal responses in both mice and humans undergoing epilepsy monitoring. Separately, daily 40hz light stimulation for several weeks reduced neuron and synapse loss in mouse models of neurodegeneration when treatment began early.

Cognito ran a trial of 76 people with mild to moderate Alzheimer's for 6 months, randomized 2:1 to either an hour a day of 40hz light and sound (via goggles and headphones), or a sham that looked similar but didn't produce the same rhythm.

It missed its primary efficacy endpoint, but hit some very interesting secondary ones:

The whole-brain volume result is pretty exciting! An Alzheimer's brain shrinks ~1–3% a year, and you can probably measure that with a scanner way more precisely than you can measure memory with a pen and paper test. So you need far fewer people to see an effect. The antibody drugs needed ~1,800 patients to show a cognitive benefit, while an atrophy trial can get away with a few hundred.

Intuitively, your brain shrinking seems obviously bad:

• Comparing MRI with autopsy, a smaller hippocampus lines up with fewer neurons.
• People whose brains shrink faster decline faster. Nondemented people in the fastest-shrinking third progressed to dementia at about 3.6x the rate of the slowest-shrinking third.
• It's faster than normal aging: roughly 0.5% a year in healthy older adults vs. ~2% in Alzheimer's.

A counterpoint, however, is the antibody drugs, which actually make brains shrink faster while slightly slowing decline. Nobody is sure why – it could be fluid shifts, real neurodegeneration, or inflammation moving in either direction! So a slowdown in brain volume shrinkage during treatment isn't automatically less neurodegeneration – but I think it is probably a reasonable proxy when combined with a quality of life measurement.

Blarcamesine is an unsuccessful drug for Alzheimer's which also slowed reduction in whole brain volume (37.6% less) – but this change was accompanied by no change in ADCS-ADL score compared to controls: caregivers did not think participants had better daily functioning.

ADCS-ADL questions try to evaluate the degree of assistance needed for daily functioning – stuff like has the person dressed themselves? Groomed themselves? Made telephone calls? Prepared food? Gone shopping? Used appliances?

In MMSE, participants are made to identify dates and location; learn three words and recall them later; perform repeated subtraction; name objects; follow instructions; copy a figure. ADAS-Cog14 is another cognitive test where they are supposed to recall a longer word list; recognize previously presented words among distractors; recall words after a delay; follow commands; copy shapes; navigate a maze; find target numbers on a page

In Cognito's case, I think it's quite possible that improvement wasn't obvious to clinicians with cognitive tests over 6 months, but clear to caregivers and the MRI scanner. They launched a Phase 3 with new endpoints aiming to measure an improvement in daily functioning, and we can expect to see results later this year.

What if you could build a Spotify plugin to drive a measurable 40hz brain response?

You can get a 40hz rhythm into music by modulating it -- eg by pulsing the volume 40 times a second. The problem is the stronger the pulse, the more the brain follows it, and the worse the music sounds. I don't think anyone's listening to a buzzsaw for an hour a day for a year. A plugin that keeps listening to music enjoyable may be one of the most valuable public good products. How do we build and evaluate it?

I think the first step is a search in a Kaggle-esque competition. We recruit a group of volunteers, and teams would submit modulators that sit on top of a shared, diverse music library. Volunteers listen with an EEG cap on, and rate how much they liked it and whether they'd listen again tomorrow. Then we look for the modulators with the best tradeoff between how strongly they drive 40hz activity and how listenable they are.

Whether a track has a strong 40hz rhythm is verifiable in code. And listenability is something AI music generators like Suno are already astonishingly good at. It's probably possible to train a model to remix music with the rhythm baked in, using the audio check as a cheap reward and the EEG and listener ratings from the competition as the expensive one to post-train your generator with.

We can use the same setup to find a rigorous sham: something that sounds almost the same as the real thing (people can't reliably tell them apart, and we keep checking that over months of listening), but that the brain responds to like normal music.

Then we take the best modulators (best ~= on the pareto frontier of listenability and gamma response) and run a randomized, double-blind trial with one primary endpoint: how fast the brain shrinks over a year. Roughly:

• Older adults with mild Alzheimer's (perhaps confirmed with a p-tau217 blood test?)
• An hour of modulated music a day, through an app that logs listening
• ~3 music conditions, each with its own matched sham, ~500 people per condition
• An MRI at the start and one a year later — along with ADCS-ADL interviews and scores
• Run somewhere with cheap brain MRIs (India is ~$200 / scan)

One study with scans a year apart estimated ~81 people per arm to see a 25% slowdown in total brain shrinkage in Alzheimer's. The real number depends on many more factors, but a few hundred per arm is probably the right estimate. Two whole brain MRI scans a year apart is ~$400. Say about $1k per person for other logistics (recruiting, screening, interviewing, the remixing app, follow ups), so call it about $1.5k per person. Three music conditions at ~500 people each (half on the matched sham) is ~1,500 people, or about $2.25m. Add ~$2m for the modulator competition and you're still under $5m.

I think this may have uniquely large public health benefits because of the scale this could be deployed at. While this is definitely stretching it, imagine a world in which you just swap some of your music consumption to a version that remixes the original song to trigger a gamma response as a potent prophylaxis against neurodegenerative conditions!

If you think this is cool (and want to help fund it?) or if you think this is dumb and want to tell me why, please reach out: samarth [at] jajoo.sh. I'll try to publish most communications on this page, in public!