1. Say everything from lights-out to lights-on is sleep

2. Test only on healthy sleepers

This is a pretty good algorithm because healthy sleepers fall asleep quickly and stay asleep the whole night. You’ll probably get >90% sleep/wake accuracy from this alone. You will miss 100% of the awakenings that happen for them during the night, but these are healthy sleepers, so there aren’t many of those. If they’re awake for five minutes in an 8-hour night, you can miss all five of those minutes and still report 98.958% accuracy.

Of course, you’ll do terribly if you apply this algorithm to a person lying awake in bed for hours, staring at the ceiling and ruminating about every mistake they’ve ever made. But you’re testing only on perfectly healthy sleepers, so that person isn’t there.

The above is why people in the sleep field have moved away from sleep vs. wake accuracy as a metric to brag about: it’s extremely low information. A great algorithm and a terrible one can look the same, depending on who’s in your dataset. You can get very high accuracy with the dumbest algorithm you can think of.

In place of accuracy, people often ask: What fraction of the true wakes did you correctly label as wake? And what fraction of true sleeps did you correctly label as sleep? These numbers are specificity and sensitivity by convention in the sleep world, and our “everything is sleep” algorithm would have a very good sensitivity (gets all true sleep right) and ass specificity (gets all the true wake wrong). An algorithm like “if you are moving, you are awake; otherwise you are asleep” gets around 90% sensitivity and 40% specificity. In general, if you tell me that you’ve made an algorithm for sleep using wristworn signals like acceleration and heart rate that has 95% sensitivity and a specificity over 65%, I’m impressed.

But this, too, is hackable. All you have to do to get an amazing specificity is include people with a lot of extremely easy-to-classify (high motion) wake in your evaluation population. You can do this by starting the window of time you’re going to use to calculate metrics a bit early (like when participants are walking around getting ready for bed instead of at lights-out) or including people who move a lot when they’re awake during the night (e.g., extremely restless insomniacs, or frequent bathroom users) in your participant pool.2

What this easy-to-classify wake does is juice the numbers with a lot of easy dunks. If I only have three minutes of true wake in my night, and I misclassify one of them as sleep, I’m down to 67% specificity, all from that one wrong minute. If I have 30 minutes of true wake and 27 of them are super easy-to-classify—unambiguous, I’m walking around, no classifier on the planet would say this is sleep—then my specificity is going to be at least 90%, even if I get all three hard-to-classify minutes wrong.

So accuracy is no good and the usefulness of sensitivity and specificity hinges on the participant pool you use for testing. It gets worse. Most validation studies are, what, like thirty people? Assume you’ve got a big population, with a lot of spread in the amount of easy-to-classify wake over the course of the night; something like this:

Assume also that you’ve got an algorithm that gets all the easy-to-classify wakes right, and all the hard-to-classify (motionless) wakes wrong. Now pick thirty people at random from your big population. How good does your classifier look? Here’s nine runs of this:

In one of your universes, the model looks amazing: 71% specificity! In another, it’s a terrible 22%. In all cases, it’s the same model. All that’s changing is who you tested it on. And this spread will only get bigger if you systematically select for healthy people in one study and people with a sleep disorder in another.

Of course you can imagine a generalization of this where instead of “sleep vs. wake” you’re doing sleep staging—light (bundled N1/N2), deep (N3), REM sleep and wake.

I’m writing all of this because of the sleep algorithm lawsuit that dropped this week, in which Oura is accused of overstating its ability to match lab sleep tracking, or polysomnography (PSG)3. When I saw the headlines, my first thought was: “Wow, that’s weird. Oura’s really good.” (It is; I like them a lot).

As far as I can tell, the claims are:

• Wristworn sleep tracking is not as good as EEG (✅, yes obviously; sleep stages were defined by EEG, the wrist is an imperfect proxy for the brain)

• Wristworn sleep staging by Oura is basically a coin flip in terms of accuracy (❌, see below)

• Oura did a bad thing by saying they had “95% Sleep Staging Accuracy” in marketing materials (🟡, yeah, they shouldn’t have done this— the defense is that they meant “two-class sleep staging” as in “wake vs. sleep,” and not “light/deep/REM/wake sleep staging.” But, c’mon. I’m never going to assume two-class is what you mean when you say “Sleep Staging Accuracy.” And yet this also seems to me to be a clear case of Marketing Team did a Bad Ad, not some grand corporate conspiracy.)

• Oura did a bad thing by reporting 79% agreement with PSG from a peer-reviewed study (❌, see below)

On the coin flip claim: What?? No! The coin flip claim comes from this paper where Oura was found to have 53.18% accuracy at four-class (light/deep/REM/wake) sleep staging in a clinical population (people with suspected sleep disorders, not healthy sleepers). But that’s not a coin flip when you’re trying to predict four classes. Consider if there existed 100 stages of sleep, and you picked them randomly over the course of the night. You would only expect to get 1% of the night correct! If you got 50% of the night right, it would be a sign that your algorithm is quite good at picking up on real signals.

Yes, okay, you could frame it as “for any given chunk of time, it was a coin flip as to whether or not it was the right or wrong stage.” But I don’t think people have good intuition for what 50% accuracy on a night of sleep looks like. It can look like this:

It’s not great, but it’s not terrible. You’re clearly picking up something in the signals.

Speaking of real signals: Have you ever seen a person score a night of sleep? I remember the first time I did. “Yeah, that looks mostly like REM,” the technician said, as we watched brain waves swim across the screen. “Eh, their eyes aren’t moving, but the sawtooth waves are there…” He shrugged and made the best judgment call he could for that 30 second chunk of time.

This is the gold standard! Of course you’re not going to be able to nail it!

For a quantification of just how squishy this all can be, see this paper, in which six human scorers tackling five-class sleep classification (N1/N2/REM/deep/wake) agree on only 46% of the night. It’s not that they’re bad at their job! It’s that we’re trying to fit a complex, high-dimensional process that’s nonuniform across the brain into a bucket of labels that you can count on one hand. If someone’s partially in REM, partially in deep sleep, they have no way to express that. It ends up being a true coin flip. And a sleep tracker’s accuracy can be punished for coming down on the other side.

Finally, on the 53.18% vs. 79% accuracy: Yes, 53% is low. But that was in a population with sleep disorders, not healthy people. And even if it wasn't a different patient population, it’s important to remember the lesson from eight paragraphs up, where you could get 22% specificity through bad luck when the typical value was more than twice that. Just as two random groups of participants can give you a markedly different picture of sleep vs. wake accuracy, so too could two random groups of participants give you a markedly different picture of sleep staging accuracy4. It doesn’t mean you did something wrong. You cast your fishing line into the pond four times, and in three studies, you got a catfish, and in the other, you got a minnow.

Where the analogy breaks down is that fishing is cheap, so you could cast your net out over and over again in an effort to arrive at a conclusive quantitative treatment of “the probability of getting a catfish from this pond.” Sleep studies are expensive, so you’ve basically got just a handful of lines to throw out there. As a non-lawyer, I can’t really say how much Oura should have added “*in healthy people” caveats to their advertising. But I don’t see an issue with them using 79% in their ads. The catfish still happened. You still caught those fish.

P.S. I’ve got my own company, Arcascope, but I have no relationship with Oura and nobody asked me to write this.

I’m out for breakfast with a friend of mine.

“My sleep is horrible,” he says. “I feel terrible during the day, then I get in bed at night and don’t feel tired.”

“Wow,” I say. “That’s so rough.”

“Yeah,” he says, pulling a bottle of melatonin pills from his pocket and twisting it open. “And I’ve tried everything.”

“Everything?” I repeat.

He pops a melatonin pill in his mouth and washes it down with a sip of orange juice. “Everything,” he replies.

It’s 9:42 in the morning. The sun is blazing in our eyes. He’s just taken a pill with five hundred times more night-hormone than the human body naturally produces on its own.

If the above was a real thing that actually happened, my next words would have been “Okay, well, maybe don’t do that, the thing you just did.” Taking melatonin in the morning is a bad idea for many reasons: it makes you groggy during the day, it makes it harder for you to fall asleep that night, it screws with your glucose tolerance. It’s also something basically everyone knows not to do: melatonin is a sleep pill, and you take it around the time you want to sleep1. Obviously.

Except people do this all the time! It’s just that instead of it being melatonin in the morning, it’s light at night.

Melatonin is negative light; light is negative melatonin.

First, accept that light is a drug. It is. When you are exposed to light, a chemical effect is triggered in your body that’s comparable to many things you take in pill form. In terms of “ability to change the time your body thinks it is,” a bright pulse of light is better than melatonin as a pill. You’re hooked up to an IV through your eyes anytime the lights are on, and sometimes it’s making you healthier and sometimes it’s making you sick.

Once you’ve accepted light as a drug, then you can start to see all the ways in which it is anti-melatonin, and melatonin is anti-light.

• Light suppresses melatonin production; melatonin makes you less sensitive to light.

• Light in the biological evening phase-delays you (or “late-ifies” you, makes you more of a night-owl); melatonin around then phase-advances (“early-ifies”) you.

• Light in the biological morning phase-advances you; melatonin at that time phase-delays you.

• Light in the middle of the biological night suppresses circadian amplitude. Melatonin at that time: well, you’ve got more melatonin! That specific rhythm, at least, is now higher amplitude.

They’re yin and yang; push and pull. When you get light at night, you’re doing the upside-down version of popping melatonin at brunch.

“Yeah, I’m not really getting that much light at night.”

Look me in the eyes and be honest. Do you turn the lights off at the same time every night.

You look away. OLIVIA reaches out and grabs your chin, turning your face back towards her.

OLIVIA

Do you… really?

A long beat. THEN:

Well, I sure don’t! And I’m a maniac about this stuff. I’ve experienced phenomenal benefits from having a very, very consistent bedtime, and I still pull “just one more email” shenanigans all the time.

Here’s the reason this matters: Light at night and light irregularity are the same thing.

If someone tells me they get bright light at 3:00 am, my first response is “probably not good.” But it’s probably instead of surely because if you get bright light at 3:00 am as part of a schedule where you, for whatever reason, stay in the light from, say, midnight to 4:00 pm every day and avoid it outside that time, I’d change my answer to “oh, never mind, you’re fine.”

Because if you do that every day, 3:00 am is no longer your biological night. That’s been shifted to start around when you turn the lights off, at 4:00 pm. You’re getting light during your body’s day, an extremely normal thing to do.

To really grind the pencil point down to a stump: Light at night is bad when it’s light during your biological night. There’s nothing intrinsic to the hour number on the wall that makes it friend or foe. Biological night is set by what you do regularly. Irregularity means, definitionally, that you are getting light at a different time than you did yesterday. Unless you live in a cave, this means you’re getting enough light that some of it is sloshing over into your body’s biological night.

A.k.a. the bad time.

So all we have to do is turn the lights on and off at the same time every day, and all problems in life will be solved

Yes.

Alright, actually no. Because people by now have generally heard the good news about sleep stability and circadian health, and they’re not turning off the lights at the same time every day regardless. I, knowing everything I know, still do not put the computer down when I have one more email to send.

When I say “light regularity is the low-hanging fruit that nobody wants,” I’m talking about this revealed preference. We gave humans the power of the light switch, and they immediately started chugging photons all hours of the day. Yes, I chug fewer photons at night than I would if circadian health weren’t my literal job, but if even my track record isn’t perfect, why should I expect anyone else’s to be?

This is the part where I reveal myself to be a Gundam villain: We need to take the choice away from people. Humans cannot be trusted with the light after dark. My mobile suit is LUME-004, and I use it to knock out the electric grid. Soon, all of us will once again experience the sweet embrace of true dark at night.

Here’s what we actually need:

1. Real-world-compatible light guidance: A reason people reject “turn the lights off at the same time every night” is that sometimes you can’t. Sometimes you just have to get the thing done. When that happens, your rhythms get jostled. The right lighting can help you efficiently jostle them back into place.
   
   When I get light late at night, I compensate the next morning with extra morning light. Clairoe, my company’s circadian health app, was designed to make sleep and light plans for people who can’t always keep a fixed schedule, but want to get back on track as fast as possible when they veer off course. Real life is going to cause dynamic lighting disruptions; there are better tools than “static bedtime” at our disposal.
   

2. Closed-loop lighting systems that adjust to your biology without you having to do anything. The lights in your house dim and change spectrum automatically based on what you, specifically, need for your sleep/circadian health. This helps you feel sleepy on a natural schedule, instead of tricking your brain into believing the sun is still up. Yes, of course, you can override it.
   
   I’ve got a version of Clairoe that I’ve hooked up to smart lights in my house for this exact purpose. My dream is to ultimately build this out further and hook it up to motorized blackout curtains, too. You don’t have to give up lighting control all the way to experience the benefits of delegating. Your home could adjust your environment to take care of you.
   

3. To make light’s effects as intuitive as taking a pill: I often say that circadian rhythms are like an audio recording of a sentence that’s been slowed down to 0.1x speed: you hear something, but because it’s so slow, you can’t derive any meaning from it. It’s just sound. With circadian rhythms, we know there’s something there, but we don’t map “light I got this morning” to “falling asleep faster tonight” the same way you don’t map sounds slowed beyond recognition to subject-object-verb.
   
   A pill is grokkable in a way that hours and hours of cumulative lighting history is not. In our apps, we’ve built features that highlight how things you’ve done in the last 24 hours could affect your sleep tonight, and we’re always trying new things aimed at the same goal. If someone plays the audio recording at normal speed, you get it instantly. We can unlock that for circadian effects on health as well.
   

4. More innovation: A pill that boosts your circadian amplitude and makes you more robust to light at night! A pill that dulls your light sensitivity while maintaining visual acuity at night! Special glasses! Special screens!
   

5. A little more faith: “Nobody wants to turn their lights out at the same time every night, even though it will make them healthier” is a clear overstatement. For as much as I was playing up the bit, I’m really very good about it. Lots of people are: I truly believe that when you start down the road of light and sleep regularity, you get addicted to it.
   
   Spend two weeks getting bright days and dark nights on the same schedule every day, then try to go back to leaving on overhead bulbs after dark without hissing at them like a raccoon. Once you get a taste of the low-hanging fruit of light regularity, you start to crave it.

But until you get that taste, closed-loop systems will surely make it easier. LUME-004 as backup.

Wait, sorry, I’m getting reports that that’s not what happened. Rewind.

It’s 2015, and someone at Peloton Therapeutics is making the call for how to dose their new molecule, PT2399, in its first in-mouse experiments. They choose to dose once at the start of the day and once at the end of the day. The molecule works: it suppresses tumorigenesis in 56% of patient-derived xenografts and beats sunitinib in efficacy. In humans, its descendant PT2977 beats everolimus on progression-free survival and response rate in a Phase 3 trial. Merck acquires Peloton for $1.05 billion dollars, with another billion in milestones pending, and the FDA approves PT2977, now belzutifan/Welireg, for the first of several indications in August 2021. As of writing, sales have climbed each year to a cumulative $1.6 billion+ worldwide. (That’s drug development for you.)

Blink and you’ll miss it, but the difference between these two worlds is that in one case the original drug was dosed once a day, late in the day, while in the second, it was dosed both morning and evening. This matters because PT2399 doesn’t work in mice if you only give it late in the day. Look, here’s the paper if you don’t believe me:

Wow, you might be saying. That’s lucky. I guess there was one time in history that a drug could have failed if they dosed at the wrong time.

It’s 2007, and Amgen is trying to decide how to dose motesanib, their short half-life VEGF inhibitor, in the MONET1 Phase 3 trial. The drug has had promising Phase 2 trial results, and poor outcomes with twice-daily dosing have pushed them to once-daily dosing. They decide that the protocol will instruct patients to take the drug once daily, first thing in the morning. The drug shows signs of activity but fails to meet its primary endpoint of improving overall survival—a narrow miss, one where a 5% change in hazard ratio would likely have been enough for a significant effect. After one more failed Phase 3, the drug is shelved.

This is the world we live in. But maybe it didn’t have to be that way. After all, motesanib is much more effective and less toxic when it’s dosed daily at night. Look, here’s the data so you can see what I mean:

They didn’t know this when motesanib was being trialed, the same way Peloton (presumably) didn’t know about the time dependence of PT2399. The difference between the two groups is that one was unlucky in dose timing; the other was lucky. That, and two billion dollars.

Why I’m a zealot about timing in drug development.

The PT2399 result comes from the lab of Dr. Katja Lamia at Scripps Research Institute; the motesanib data comes from my company, Arcascope, where we’ve identified whole classes of drugs that we think failed because timing wasn’t considered. Briefly, the logic goes like this:

1. The body is not constant in time. It’s rhythmic all the way down.

2. Therapeutic windows aren’t constant either.

We act like they’re this:

When really, they’re this:

3. Many approved drugs could probably be time-optimized to be more effective or less toxic. See: belzutifan.

4. Many failed drugs could also probably be time-optimized to be more effective or less toxic. Like motesanib. Like the others we’ve found.

5. This is the opportunity. We can rescue drugs by delivering them differently in the fourth dimension. Time is a whole additional dimension to explore and exploit. Let’s explore and exploit it.

Why Hasn’t This Been Done Already?

1. “People can’t measure circadian phase in the real world.”

Ah, but we can. We’ve been collecting data for years to build best-in-class methods for tracking circadian rhythms in real world conditions from devices you already own. Not generic, not static in time: we can figure out the best time for this specific drug, on this specific day, for you.

2. “People don’t know what drugs have timing effects, how many are out there, and what the best time is.”

We can figure it out with mechanistic models. Arcascope’s drug discovery platform, chronoRCT™, takes circadian expression datasets and builds mathematical models of the underlying systems, allowing us to probe optimal timing in silico and confirm it at the bench, like we’ve done with motesanib (and more!).

How many drugs like this are out there? We think many. After all, timing effects are extremely easy to miss.

For starters, we systematically dose nocturnal rodents during their rest period (the light) while dosing humans during the active period (also the light). There are many, many reasons why drugs can work in mice and fail in humans, but one of them surely is this consistent bias wherein dosing in mice vs. humans is literally night and day.

After all, do you know anybody testing delivering a drug every four hours for 24 hours that isn’t a circadian biologist? It’s an incredibly annoying experiment to run because it means everything is six times more expensive and someone is losing sleep. If I tried to sell you “Make Everything Six Times More Expensive (And Someone Is Losing Sleep)” as a product, would you sign off on it? Would your boss?

And let’s say you get promising enough results to go into human trials. The natural spread of human circadian rhythms is wide across the population and moves within an individual from day-to-day. Without circadian-tracking technology like ours, you’re shooting in the dark. You could get lucky in a Phase 2 trial through a happy collision of protocol-specified timing and biological timing of your participants making it so the overlap between drug and target is good enough for an effect. The odds of this luck continuing in a Phase 3 dwindle as N scales.

The biggest proof that people are missing this is that they missed it. Belzutifan’s been out for five years now; Dr. Lamia’s lab only published their findings a year ago. All once-daily short half-life VEGF inhibitors tried in the US failed. In the post-mortems dissecting what went wrong, nobody considered that the timing was just wrong.

What a time-based drug rescue looks like

Here are three potential shapes:

1. A drug works, but only when it’s temporally aligned with its target. The only way to make it work is to time it right.

2. A drug works but has significant toxicities from off-target effects. The only way to make it work is to time it right. (This is the last one wearing a different hat.)

3. A drug interacts with a rhythmic process. If it flattens the rhythm out, there are negative downstream consequences that outweigh the benefits of the drug. The only way to make it work—to keep the rhythm it interacts with healthy—is to pulse it.

I’ve got examples of all three of these that I’d like to try to rescue. And for that, I need some help.

Call me if this is you

I’m a computational biologist by training. I need someone with clinical development experience to join my team.

I’m looking for other people, too: industry advisors who can guide me on the path of the first in-humans timing rescue, drug developers who hear echoes of an asset they worked on in what I’m saying, and skeptics to argue with me so I don’t go down bad roads.

But a clinical development person is who we need most. We want you if you’re frustrated by the 90% failure rate of drugs in clinical trials, if you’ve been on the inside of bringing drugs to market and think there’s a better way, and if what I’m saying about “drugs can resonate with their targets” resonates with you.

What you’ll do:

• Help us pick which drug we should move forward first. We need to balance the likelihood of success with market size and path to patients. We’ve got ideas, but we want yours.

• Help us design our trial. Build the protocol that makes the best possible case for our drug, up against the closest competitors in standard of care. We want our drugs to be clear winners, not also-rans.

• Chart the regulatory path. Timing rescue is a new idea. It’ll be new to the FDA, too. We’ll need you to help us all navigate it together.

• Bring in assets. Reach the companies and academics sitting on drugs that didn’t succeed and make the case that theirs could, with different timing.

• Be our guide. We know a lot, but there’s plenty we don’t know yet. You’ve got the on-the-ground expertise with drug development; help us learn from your example.

Time is a whole extra dimension. Like going from still frames to movies, or static maps to GPS, when we let medicine move in time, there’s incredible value to be had. Somebody is going to hit play and start chronomedicine rolling. I’d like it to be us. Sound interesting? Reach out.

With thanks to Katja Lamia for her feedback and Franco Tavella for coming up with the therapeutic window plots.

The boat race is taking place in a circular current. You and your competitors are looping around and around, pushed along by the movement of the water.

The goal of the race is not to do the most loops. Instead, one boat has been labeled the target boat.

The goal of the race is to catch that boat.

You have exactly one tool at your disposal: your sail. The wind in the part of the world where this curious race is taking place always blows from east to west. You can put your sail up to catch the wind, or put it down to opt-out of the wind’s effects.

Q: How do you catch the target as quickly as possible? I’ll put some Jeopardy! music on while you think.

A: It depends on where the target boat is relative to you!

If it’s a little bit ahead of you, then you want to put your sail up when the wind is at your back, and put the sail down when the wind is blowing against you. You want to get wind when it speeds you up, and avoid it when it slows you down:

If it’s a little bit behind you, then you want to put your sail up when the wind is blowing against you, and put the sail down when the wind is pushing you ahead. You want to get wind when it slows you down and avoid it when it speeds you up:

Here’s the mapping to jet lag:

• The boat’s location in the loop is your biological time—the “time zone” you’re starting in.

• The target boat is where you want to be in your destination time zone.

• The wind is light exposure.

• Putting your sail up is availing yourself of it: getting light.

• Putting your sail down is avoiding light.

You get over jet lag by getting light at the times when it helps you catch the target and avoiding light at times when it pushes you away from the target.

Obviously it is not that simple or you would not have made this a multipart blog post

Yeah, but consider how much heavy lifting is done by this analogy:

1. People’s bodies change over 24 hours as we complete a circadian day. Jet lag is the process of catching a looping target while also looping ourselves.

2. Just like the wind does different things to your boat (speeds it up, slows it down) depending on how far through a cycle your boat is, so too does light do different things to your body (speeds it up, slows it down) depending on how far through the circadian day you are.

3. There isn’t a “one-size-fits-all” rule—you have to do very different things depending on whether the target boat is ahead or behind you.

4. The right choice for you in a given moment (wind or no wind, light or no light) doesn’t depend on what the target boat is doing in that moment. In other words: adjusting as quickly as possible can mean dimming the lights when it’s bright in your destination, or getting light when it’s dark in your destination.

But shouldn’t I just follow the sun?

Sometimes, but not always!

Here’s a case where it’s great to follow the sun: You normally get up at 7:00 am. You’re going from New York to London. You land in the U.K. at 9:00 am London time. This is 4:00 am in your home time zone. You are probably in the biological zone where light speeds your rhythms up. Speeding yourself up is what you want to do for this trip. Getting a ton of light at this time will help you catch the target. In boat race terms, it’s like this:

Here’s a case where it’s bad to follow the sun: You normally get up at 7:00 am. You’re going from New York to Istanbul. You land in Turkey at 9:00 am local time. This is 2:00 am in your home time zone. You are probably in the biological zone where light slows your rhythms down. Slowing down your clock is the opposite of what you want to do for this trip. Getting a ton of light at this time will not help you catch the target. In other words:

Following the sun is a very natural thing to do. But jet lag is inherently unnatural—abruptly changing light/dark schedules is not a thing we evolved for— so natural solutions don’t apply.

So how do I know what the “right times” to get and avoid light are?

In my boat animations, it’s easy to tell when having your sail up speeds you up or slows you down: It speeds you up when you’re going the same direction as the wind, and it slows you down when you’re going against the wind.

The equivalent information for your body’s clock is encoded in a phase response curve (PRC). To figure out how to get over jet lag for a given trip, you can sit down with a copy of the human phase response curve and apply the following algorithm:

1. Start with the PRC shifted so it represents your current biological time.

   1. If you’re going eastward, get light during the phase advance part of the PRC, and darkness elsewhere.

   2. If you’re going westward, get light during the phase delay part of the PRC, and darkness elsewhere.

2. Slide the PRC to the left or right to reflect how much you shifted today

3. Repeat until you’re in sync with the new time zone.

This will give you a schedule to get over jet lag. Congratulations; you did it. The “Part 1” in the title of this blog was a fake-out. We’re done here.

Just kidding

That schedule probably sucks! It sucks for any number of reasons, including:

1. If you don’t follow it exactly—your flight gets delayed, you deviate from the schedule because you want to go out—you will have no idea what to do from then on.

2. You won’t want to follow it exactly, because it tells you to get a bunch of darkness during the day.

3. You won’t want to follow it exactly, because if you try to shift before you leave, it’ll have you getting a ton of light at night.

4. The darkness times and the times when you want to be asleep will be out-of-sync.

5. The darkness times and the times you can fall and stay asleep will be out-of-sync, and, of course,

6. It’s not actually the best schedule because it reduces the circadian clock to a single dimension (phase).

Not to mention, you could only make a schedule to begin with if “Start with the PRC shifted so it represents your current biological time” meant something to you, and I purposefully threw you no lifelines there.

So that’s a way of getting a schedule, but probably not a great one. Want to do better? Stay tuned for Part 2.

To my enormous chagrin, circadian rhythms are maybe like the fifth most famous biological rhythm. I’m putting them behind breathing, walking, and menstrual cycles, here—don’t break my heart by telling me they should be even further down.

But I’m grateful to the #1 most famous rhythm (heart rate) because it makes the point I want to make today very neatly for me: When your heart rate goes flat, that’s about as bad as it gets, health-wise. When your circadian amplitude goes flat: bad too.

No, you don’t instantly die. Flattened circadian rhythms are all around us these days, sagging under the weight of aging and disease or squished underfoot by our jobs and behaviors. That people can work nights, eat around the clock, and live with dementia for decades shows that this isn’t an instantly fatal prognosis.

It does eventually get you, though. And what’s unique about the way in which it gets you is that it almost never takes the blame. Death, asphyxiation, paralysis, menopause: when other biological rhythms go flat, we give it a name. You could tell me, for any one of those terms in the last sentence, exactly which rhythm has been lost.

Circadian amplitude suppression, on the other hand… Well, insomnia, ADHD, idiopathic infertility, chronic inflammation, and non-dipping blood pressure come to mind.

You seem to have skipped a lot of steps there, Olivia

This is fair. After all, we don’t really have a gold standard way of quantifying circadian amplitude in the first place. It’s awfully bold to suggest that the root cause of the conditions above and many more is “thing we can’t easily measure or define; just trust me.”

But being bold is what you get to do in blog posts. And in that spirit, I will now lay out my case for 1) what circadian amplitude is, 2) why it’s the best framework for capturing the idea of “clock bad,” 3) how we, in modern life, chronically suppress circadian amplitude through our choices and actions, and 4) what we need to prove all this is right.

What circadian amplitude is.

Well, you know (waves hands). It’s when a rhythm that used to look like this:

starts looking like this:

When I talk about circadian rhythms having “terrible UI” (that is, having a bad user interface as a concept for people learning about them for the first time) I usually mention the fact that there are so many circadian rhythms, which is overwhelming, which causes people to short-circuit and fall back on the much more mentally tractable “circadian rhythms = sleep.” If you, exposed to the prior sentence, are starting to affix a framed version of it to the walls of your mind, please imagine me spiking it out of your hands.

Yet amplitude also suffers from this problem. There are so many ways I could define it that it’s hard to have a coherent conversation about it, which means we just mostly don’t talk about it at all. But let’s give it a go while we’re here.

I could define circadian amplitude with any subset of the roughly 50% of the genome that’s rhythmic. Example: PER1 rhythms in this cell are blunted—they go up and down less than they normally do in healthy people—therefore you have low amplitude.

I could define it at the tissue/organ/whole body level in the same way, with behaviors dampened by gene or protein rhythms being low amplitude. But I could also define it at those scales as arising from high amplitude constituent parts that are out-of-sync with each other.

Think: teens jumping up and down together in an elevator trying to break it. If they’re in sync, their combined force causes the elevator to lurch and stall out and everybody screams and they get a good TikTok out of it. If they’re jumping out-of-sync, nothing happens. Now swap in “neurons in the suprachiasmatic nucleus” for “teens” and “breaking the elevator” for “sending an unambiguous day/night signal to the body.” That’s amplitude.

Finally, I could define a functional “utilization” amplitude, along the lines of “you need Process X at these times, but you have it at these other times.” Like this:

Your utilization is flat. This, too, is low amplitude.

Fundamentally, amplitude is confidence. Amplitude is the clarity of the day-night signal; the body’s certainty that its estimate of time is correct. High amplitude means your body can take more risks, get more gain. Low amplitude is a state of chronic uncertainty—a wobbling bike, a sputtering motor. A refusal to take the plunge and say with full-throated conviction “yes, THIS is the time for peak insulin sensitivity!”

Measuring anything resembling amplitude is a nightmare.

Say I want to try to estimate whole-body circadian amplitude using temperature.

It needs to be core body temperature to be a true circadian marker, which means you’re swallowing a thermometer pill or inserting a rectal thermometer; plus, we have to make sure you don’t sleep or move too much to avoid losing the circadian signal under the masking effects of sleep and motion. Plus we need to get the full 24 hours to make sure we’re seeing the full up-and-down of the rhythm over the course of the day.

Enjoy lying awake in the dark for 24 hours!

Even a full day of data collection might not be enough. Say I want to estimate your amplitude by taking the area under the curve of your melatonin rhythm. Once again, I forbid you from moving too much or sleeping. Once again, I take samples (spit, probably) from you over hours and hours to get a plot that looks like this:

Except some people are just really high melatonin producers, for reasons that probably don’t have much to do with their circadian clock. Or, in other words:

Which means we can’t really use the entire day’s worth of melatonin we sampled to say something generalizable about their circadian amplitude. We can say things about changes within the individual, like “your melatonin amplitude is 20% higher than it was 7 days ago, the last time I made you sit in the dark for a full day as I collected your spit.” But what we want is a simple, one-off test that can say “hey, your amplitude is low,” and we don’t have that yet.

What we often default to is using “amount of daily activity measured from a wearable” as our definition of “circadian rest-activity rhythm amplitude.” Which, okay, I’m not going to say there’s nothing there, but there are certainly many more things driving what we see (movement) than what’s going on in the circadian part of your brain. Or:

A shift worker could run a marathon one day and be categorized as high amplitude. A perfectly healthy person could be stuck in meetings all day and read as low amplitude. I could kidnap you and put you in a Saw-style trap where your arm is attached to a moving windshield wiper all day, until you agree with me that what we’re measuring is not a good representation of your body’s confidence of day-versus-night. During the time your arm is moving? Under a “did you move a lot” definition of amplitude, you’d look incredible.

So amplitude has no one definition, means different things at different scales, and is practically impossible to accurately measure with current tools.

Why circadian amplitude suppression is still the right way to talk about diseases of the circadian clock.

Plenty of terms out there are aimed at capturing the idea that you’ve screwed up your circadian system. Circadian misalignment, chronic jet lag, circadian disruption—these all imply some form of circadian injury. What drives me crazy, though, is that none of them go the extra step of spelling out why that injury is bad. There’s a gap where mechanism should be.

Amplitude neatly solves this problem. If rhythms are flat, then there is either not enough of something when you want it or too much of something when you don’t, both of which are bad.

Why is chronic jet lag bad? Because it squishes your rhythms.

Why is circadian misalignment bad? Because if you are forced to wake up or get light or eat at a time your body isn’t anticipating, it squishes your rhythms.

Why is circadian disruption bad? It’s not because you’re knocking your clocks into a healthier, more robust state. It’s because somewhere there’s a rhythm, and it’s flatter now than it was before you messed with it.

My claim: Behind every circadian problem there is, somewhere, a low amplitude.

Contrapositive: If you don’t have reduced amplitude somewhere, you don’t have a circadian problem.

Consider, for a moment, the person who is a night owl but a very happy, consistent night owl—who goes to bed and gets up at the same time every day, gets plenty of light during their waking hours through sun exposure and full spectrum artificial lighting, who has extremely dark blackout curtains during their rest periods, and who never has to get up early for work or school. This person might be flagged for circadian misalignment simply by virtue of being a night owl, but do they actually have a circadian problem? No! But if you made them start waking up at 5:00 am, the light and activity hitting their system at that time would blunt their rhythms.

Consider next the person who has the most middle-of-the-bell-curve chronotype you can imagine; who keeps their wake-up and bed times precise down to the minute, but whose days are very dim (works in a basement) and whose nights are very bright (light pollution from outside). Does this person have a misalignment problem? Nope, nor do they have chronic jet lag or circadian disruption. But their amplitude is surely lower by virtue of the fact that they have massively reduced the contrast of the day-night signal to their brain.

And of course we can construct the example of a rhythm that is full amplitude, just horizontally shifted relative to where it’s needed, and all of the problems this person has in their life stem from the fact that this rhythm—appearing full amplitude to any passerby!—is not where it needs to be. But this is just my low “utilization” amplitude from above! Misalignment is low amplitude.

Uh, excuse me, what about gain-of-rhythmicity?

(Rusty geezer voice) Well, well, well. I thought you might bring up gain-of-rhythmicity.

A fun fact about aging and circadian rhythms is that you don’t just see existing rhythms flatten out as you get older. You also see previously flattened rhythms start oscillating.

Surely the appearance of so many new rhythms during old age—a period of life characterized by worsening health outcomes—implies circadian amplitude suppression is a woefully inadequate way of describing what’s going on! Checkmate, etc., etc.

Ah, but remember: the claim is that the presence of a circadian problem implies the presence of a low amplitude somewhere, not “every rhythm in every circadian problem is squished.”

Here, I’m going to draw on my own life. Consider the rhythm of my toddler breaking stuff daily in my home. Generally, this is a low amplitude rhythm: he breaks nothing at night while he’s asleep and maybe 0.5 things per week during the daytime hours. Averaged over many days, it would look like a very flat rhythm with some elevation during the times he’s awake and a flat zero (0 broken things per hour) from 7 pm to 7 am.

But this is only because his ability to break things is constrained by my high amplitude inhibition rhythm, where I am actively trying to prevent him from breaking things during the hours I’m awake. If that rhythm gets knocked out—I get sick or distracted by writing a 3,000-word blog post about circadian amplitude—my inhibition decreases, and the amplitude of his destructo-rhythm increases.

Except instead of a toddler breaking things, it’s cortisol, and instead of me, it’s SCN inputs to the HPA system growing weaker. You see where I’m going with this.

How we, in modern life, continually suppress circadian amplitude through our choices and behaviors.

Amplitude suppression might occur acutely. If you typically turn off the lights at 10:00 pm, and I break into your house and turn them all on max brightness at 11:00 pm, your melatonin levels, which have probably been rising for the last three hours, will go down. This is very straightforward, as melatonin is inhibited by light. It’s the night hormone. Light means not night.

Amplitude reduction might also be a state you’ve been driven to, either by disease (e.g., Alzheimer’s), diet (e.g., high-fat), or a long history of confusing inputs to your body (shift work). A person in a more persistent state of low amplitude doesn’t need to be in the light for their melatonin amplitude to be suppressed. Remember earlier, when I said the amplitude of melatonin is best understood relatively, since some people are naturally high producers and that makes it so we can’t use the same scale for everybody? That’s still true. But it’s undeniable that shift workers as a class tend to have lower, blunted rhythms than non-shift workers. Take this person, for example:

This is real data from a shift working person who was in the dark the whole time as data was being collected. This is a person who should have had melatonin start to clearly rise and fall at some point in the 24 hours they were assessed. Instead, it’s a 24-hour slump to zero. The endpoints on the far left and right, which meet up in rhythmic data, are nowhere near each other. They’re arrhythmic! They’re low amplitude.

But I’m not a shift worker!

But you are a person living in a place with internet in 2026. Which means you’re probably not getting light and dark as regularly as a person who lives by the sun and firelight alone. And you’re almost certainly not getting as bright a bright or as dark a dark as them.

You know what it’s like to be in a groove on a swing on a playground. You’ve got clear momentum. It feels good. If you started bucking your legs wildly back and forth at random times, this would not help you swing higher. Even if you’re not totally random, anything other than resonance with the tempo of the swing will dampen your swing. You’ll be in less of a groove; you’ll have lower amplitude.

The leg bucking here is the stand-in for the clock inputs (light, activity, wakefulness, sleep) that you want to do rhythmically, in sync with the twenty-four-hour day. Most of us aren’t all that rhythmic. Modern life has conditioned us that sleep duration is what we should care about (8 hours a night) without much attention paid to when those eight hours happen. Yet it turns out that sleep regularity is a better predictor of mortality than sleep duration. Why? Surprising no one, I say amplitude suppression caused by irregularity of inputs to the body’s clock (see Figure 2 in here for a visual). You’re lower amplitude, so you get less of the things you want—say, tumor suppression—when you want them, and more of the things you don’t want—say, inflammation—when you don’t.

And it’s not just the inputs being irregular that keep us out of a circadian groove. We get way less light during the day and way more at night, the equivalent of wimpily flopping our legs forward and back on the swing in place of good, strong kicks. And we make lifestyle choices, like high-fat diets, that impair our ability to groove at all, the equivalent of trying to swing in big honking clown shoes that scrape the dirt on each back-and-forth.

Hear, as I’m talking to you, the beep-beep-beep of a heart monitor in your medical drama of choice. When the regularity of those beeps starts to break down, it’s a sign that the big flat is coming; the lowest amplitude of them all. This is what we’re doing to ourselves when we keep the lights on late and eat at 11:00 pm. We are all low amplitude.

You’re healthy because you’re flat?

Of course, no biological story is ever just that simple. When you knock out the circadian clock, mice die significantly sooner than they otherwise would. There are some knockout mouse lines where they’re more vulnerable to certain cancers. Yet there are some knockout mouse lines where they’re less vulnerable to certain cancers. What gives?

When you lose amplitude, you lock in at some value—the MESOR of your dampened oscillations. Low amplitude, I claim, is bad because you have either too little of something at some time, too much of something at some other time, or some combination of both. But there’s no reason why you couldn’t also, as part of being low amplitude, lock something in at a level that’s both higher than it should be and actively beneficial in some contexts. Imagine you flatten your rhythms in a way that happens to make your baseline DNA repair higher, for instance. You’ll still die earlier from lack of clock, but you won’t die of cancer.

I’m sticking this here at the end because biology is a great humbler. I could come out and say “High amplitude for all! Brighten the ICUs!” and then it could turn out that it’s actually good to be low amplitude when you’re very ill in order to keep things from swinging too low at any point. I could tell cancer patients to make their rhythms higher amplitude, and it could end up boosting the vitality of both tumor cells and healthy cells alike. With anything health and biology related, blog posts are just about the only place you can be carefree and silly. Actually translating to patient care means taking careful steps and being obsessively paranoid of the potential pitfalls at every turn.

But at the very least we should be taking those steps in the right direction. And the right direction is towards reframing much of chronic disease as a form of circadian death.

So what do we need?

We need a good, easy way to measure amplitude. We need something that, with a single sample or lightweight tracking, can give you a quantitative assessment for whether someone’s flat or not. There are better ways of using activity to get an estimate of amplitude than just “how much motion did you do,” and blood and hair tests of gene expression are ripe with potential. More work on these is critical. We can’t communicate about amplitude if we don’t get very precise about what we mean, and shared measurements are the first step towards getting there.

We should make sure people know the recipe for high amplitude by heart: bright days, dark nights, same time every day. Same principle for food. Don’t eat at times when your body doesn’t expect it. Avoid things that can trigger low amplitude by clogging up the clockwork of your body’s machinery, like high-fat diets.

And we should probably stop saying “circadian misalignment” if we want people to really, actually care about circadian rhythms. Misalignment sounds like something an orthodontist could fix. Cosmetic, temporary; “the frame was misaligned on the wall so I aligned it.”

People out there are starving their system of DNA repair at night and deluged by inflammation during the day. They’re glutting their microglia with cholesterol beyond what the daily flush can clear, and frying their SCN with light at its most vulnerable hours. There should be an urgency to rescuing dying circadian rhythms, not a shrug. Circadian amplitude is whole-body health. Let us not go gentle into that long flat.

Thanks to Marc Ruben and John Hogenesch for comments and feedback on this post!

“What a nice young man,” I say as I shut the door behind him five minutes later. “I’ll remember him fondly.”

(Jump one universe to the left).

My new neighbor knocks on my door at 12:00 am. “Hi,” he says. “Just wanted to introduce myself. Hope I didn’t interrupt your evening.”

“What a freak,” I say as I shut the door behind him thirty seconds later. “Who does that? Who introduces themselves for the first time at [expletive] midnight?”

The same event—meeting the same neighbor—hits five hours apart in two identical universes. In one, our relationship starts congenially. In the other, he’s forever memorialized in my brain as Weird Guy (Do Not Interact).

The question posed today isn’t whether I’m right to respond this way (of course I am) or if my estimate of 9:00 pm as the time when appropriate-to-stop-by becomes don’t-do-it is correct (yes).

The question is whether or not this happens with immunotherapy for cancer.

An extremely fast primer to get everyone up to speed

Taking drugs at different times can make them better or worse at their jobs. We know this for sleep drugs. Nobody takes Ambien at 7:00 am to help them sleep better that night. Take melatonin in the evening and it will speed up your body clock, shifting your rhythms and making you feel sleepy earlier. Take it in the morning, and it will have the opposite effect.

But sleep is not the only circadian rhythm, just the most visible one. Plenty of other things in your body have rhythms you’re at best dimly aware of. I only notice rhythms in grip strength when I’m trying to open a jar at 6:00 am in the morning, and I only notice rhythms in immune response when my sinus infection symptoms get abruptly worse at night. Me noticing these things is not a precondition to them mattering, though, and it stands to reason that plenty of non-sleep drugs should have a “best time” of administration.

…if they have a short half-life.

Short half-life means they get in and out of your body quickly enough for timing to matter. Caffeine has a short half-life, which is why I can drink cold brew in the morning and have it cleared out of my system in time for bed sixteen hours later. Hypothetical 20-day half-life caffeine would be a terrible idea—it would mean that the amount of caffeine in my system would take 20 days to fall to half its initial level. If my first sip of long-acting coffee happened at 6:00 am, then 98.28% of my dose would still be floating around my body 12 hours later at 6:00 pm. Who cares that I drank it in the morning? I’m consigned to weeks and weeks of bad sleep either way.

It’s simply very, very clear: if something’s going to have a time-of-day effect, it better not have a 20-day half-life.

Hmm.

Except immunotherapy for cancer kind of seems like it might?? Or rather: Over and over again, retrospective studies have found time-of-administration effects, with morning dosing being better than afternoon dosing. This is weird, and from just about the first moment retrospective results dropped, people have been calling it out as weird.

After all, the immunotherapy drugs in question have VERY long half lives (22-25 days). Everything in my long half-life caffeine example applies to them. The amount of drug in your system 12 hours after an initial dose is essentially identical to that initial dose. How could timing the dose possibly matter when the difference it makes is on the order of 1%?

A lot of people have spent a lot of time thinking about this and have arrived at the same conclusion: it has to be a first exposure thing. If it’s real, it has to be something like what’s going on with my neighbor. The half-life of my neighbor, after all, is maybe something like 40 years1. The same amount of my neighbor is on my street every day of the week for as long as we both live there. But a five hour difference in time-of-first-introduction is all it took for me to permanently remember him as “fine” or “freak.”

But is something like that actually what’s going on?

The case for No

1. Not every retrospective study finds a morning benefit to immunotherapy for cancer.

Many do, see: [Karaboué et al., 2022], [Catozzi et al., 2024], [Qian et al., 2021], [Gonçalves et al., 2023], [Patel et al., 2024], [Nomura et al., 2023], among others (17 covered in [Karaboué et al., 2024] alone).

But not all of them do, see: [van Rensburg et al., 2022], [Janopaul-Naylor et al., 2024].

So if this is a phenomenon that only appears sometimes, it is natural to trust it less than if every study reported a positive result. It’s not “being struck by lightning is bad for you” levels of clear-cut causality. And speaking of casualty,

2. There are non-circadian things that could drive a time-of-day effect.

This one is extremely compelling! Whenever I’ve been pregnant, I’ve done everything in my power to schedule appointments as early in the day as possible because I’ve got other stuff to do later in the day, and I don’t want that stuff to be interrupted. I’m an early bird, and earlier is just easier for me.

Early birds tend to be healthier than night owls, and it seems completely plausible that early bird patients could be both healthier and choosing to come in earlier.

Or better yet: I’m less likely to schedule an infusion for 8:00 am if I live far away from the hospital. If I live far away, I’m more likely to live in a rural area, which correlates with worse health and fewer resources. So people who are more likely to be doing worse are systematically scheduled to be later in the day. That’s all you’d need for an effect like this to arise!

This is the argument of a paper from this week where the authors include “distance to the treating institution” in their statistical model and find no relationship between timing and survival (while also finding that the farther you live, the less likely you are to have an early appointment).

The natural way to address these confounders is with a randomized controlled trial (RCT), where the study team, not the patient, picks when they get their infusion. But, unfortunately,

3. The one prospective randomized controlled trial just got retracted.

Adam Rochussen of the Salk Institute has a really great write up on this, so go read that first. Briefly:

• Mid-trial updates to the clinical trial record

• Protocol submitted to the journal with the wrong date

• Fewer dropouts than expected (interestingly, also flagged in another questioned chronomedicine study)

• Kaplan-Meier curves looked less stair-steppy than expected given the typical scan schedule

And perhaps most eyebrow-raising of all, the effect size was enormous. Which is surprising because

4. A very large morning-to-afternoon swing would be weird.

Big swings in circadian rhythms can happen, of course. Melatonin goes from nothing to ON over the course of a few hours.

But a big morning-to-afternoon swing implies that a hypothetical efficacy curve looks either like this:

Or this:

Given that nobody’s reported that extreme night owls getting infusions in the morning (i.e., their biological night) go into complete remission right away, I’m doubtful that that the first one is what’s going on, and suspicious of the second one as a hypothetical and sharply skewed circadian rhythm without any clear mechanism behind it.

But I’m four pages into this blog post, and I’ve mostly just remade arguments people made before the immunotherapy RCT was retracted. So in the interest of adding something new, let me now make

The case for “well, hang on”

1. Not every retrospective study finds a morning benefit

Let’s assume there IS a real circadian time-of-day effect to immunotherapy for cancer. What could we expect—knowing nothing about what’s played out in the last six years—to happen?

Well, for starters, wall clock time (AM vs PM) is not biological. Within individuals living in the same time zone, you can get a ~10-hour spread of biological times at the same wall clock time2. Biological time is set by the history of synchronizing signals a person gets (food, light, activity) which can be wildly different between people living in the same home, let alone hospital service area.

This means you’ll have some people coming in for an afternoon appointment during their body’s biological morning or a morning appointment during their body’s biological afternoon. The spread in biological time will blunt the observed effect size. Or, in graph form:

Participants can also have flatter or more pronounced rhythms (circadian amplitude), which could similarly blunt or enhance a real effect.

Plus the authors in these retrospective studies are all using different definitions of “morning” and “afternoon” and different ways of deciding which infusions to use in their analysis (all of them, first one only, first few, etc.), making direct comparison a challenge.

Finally, if there’s one lesson I’ve learned from reading the cancer literature, it’s that all tumors act the same way and that if something is true for one of them, it’s true for all of them. (I pull a baseball bat out from behind the desk.) No! Of course not!

So to “not every retrospective study finds a morning benefit,” I’d say “yeah, as expected”.

2. There are non-circadian things that could drive a time-of-day effect.

Sure, but there’s preclinical data showing that the timing of the first dose (of two) can dominate tumor response ([Wang et al., 2024], Figure 5E):

In the above, the red lines show what happens to tumor growth in mice when the first of two doses is given at the start of the rest period, while the blue lines show what happens when the first dose happens at the start of the active period. Regardless of when the second dose happens, the first dose drives the effect. A very valid response to “what if it’s all people who drive long distances to the hospital being sicker?” could be to say “trust us, that’s not happening in mice.”

But I want to engage more with “distance from hospital site” as a possible explanation for observed time-of-day effects, both because mice aren’t people, and it’s a very reasonable explanation to propose. The thing about it that gets me is this: People come to the hospital for a lot more than just immunotherapy. If it really is just rural, far-away people being more likely to be worse-off, shouldn’t we see a morning benefit for lots of other things as well…? Chemotherapy, appointments in general3— why not morning superiority in all those cases? Why only immunotherapy?

You might say “well, maybe we DO see that morning is best for all of those,” but I’ve looked at infusion times for chemotherapy in fairly large datasets and seen no time-of-day effect4. With so much historical data out there, I’d expect that other people have done the same. Why aren’t there a slew of papers showing equally huge benefits to morning for all the many things people drive to the hospital for that aren’t immunotherapy?

Once again, we need a prospective RCT. And, of course,

3. The one prospective randomized controlled trial just got retracted.

Yep, but “this study is not reliable” does not mean “this phenomenon is disproven.”

Lotta people out here mistaking If P Then Q for If Not P Then Not Q. That’s not how it works.

The contrapositive of “if there’s a randomized controlled trial supporting the timing hypothesis, then it’s likely a real effect” is “if it’s not likely a real effect, then there isn’t a randomized controlled trial supporting the timing hypothesis.” Not “no trial, no effect.”

You could say that if there had been a real effect in the retracted study, then the authors wouldn’t have needed to change the trial registration/protocol, etc. But about all of this, there’s a lot we don’t know. I’m more inclined to assume unfortunate sloppiness than malfeasance until I’ve got a blindingly obvious reason to believe otherwise.

And, finally:

4. A very large morning-to-afternoon swing would be weird.

But hey, weirder things have happened.

Final thoughts

Is immunotherapy in humans like my neighbor example, where a few hours difference in time-of-first-introduction for something with a long half-life has a huge impact on the response to that thing forever after?

My guess in this moment is: ehh.5

Not because of any certainty I have that such a thing is biologically impossible (the mouse data in [Wang et al., 2024] is this, after all), but because it’s just not a slam dunk in the human data. Consider [Fletcher et al., 2025], where late immunotherapy shows up as bad for you when you use all the infusions in your calculations, but not when you use only the first four. This is the opposite of [Nomura et al., 2023], where they find a much clearer signal from the first infusions, but it’s enough to make me look askance at the “first infusion is the only thing that matters” hypothesis.

I’ve sometimes wondered why more people haven’t put out a compounding argument, and I’ve usually landed on “because it’s so simple it’s embarrassing, and people will say `that’s not how cancer works, idiot’.” But I, fearless, will make the argument anyway: Let’s say you get one dose in the morning, and it reduces your tumor to 98% of the size it would have been with afternoon dosing, a near negligible difference. Then, weeks later, you do it again and again, always in the morning, always getting that 2% edge. Eighteen sessions later, 0.98¹⁸ = 0.695—a thirty percent improvement over the always-afternoon dosing regimen! Let’s say it’s a 1% edge instead: 0.99¹⁸ = 0.835. Still pretty good!

Yes, most people won’t be all-morning or all-afternoon, and plenty of people don’t get to 18 sessions, so you’d expect the effect size to be smaller in practice. And, yes, for sure, this is a massively simple model. But at least it’s a model I understand, and—best of all—one that can pair nicely with other hypotheses, including a weaker form of the first infusion effect, where first exposure matters but isn’t everything.

Maybe the thing that unifies all the data to-date is some combination of compounding small benefits, a reasonable first infusion effect, and the inherent noise of biological time making clear morning/afternoon cut points unstable. Or maybe it’s all confounders! At the end of the day, it’s still retrospective data. We need an RCT that sticks around.

There are ongoing randomized controlled trials at other sites looking for timing effects right now. They’re surely going to be very careful about protocol updates and providing accurate documents to the journal. They’ll be highly motivated to make sure the characteristics of the different timing groups are as balanced as possible. They might find absolutely nothing when they do this, which will be a glorious vindication for those who questioned the results from the beginning and a textbook example of why RCTs matter.

Or they might, even after controlling for everything they can think of, still find a benefit to morning dosing. If that happens, there will be some people who nonetheless doubt the result. Some because timing medicine sounds like astrology to them, some because debunking is objectively fun, and some because it would be a massive pain in the ass to have to change clinic times. I’m sympathetic to all of these, minus the “timing medicine is astrology” take. Chronobiology grad students did not stay up all night in the lab sampling every two hours to be disrespected like this. They stayed up to be disrespected for being grad students.

References

Catozzi, S. et al. (2024) ‘Early morning immune checkpoint blockade and overall survival of patients with metastatic cancer: An In-depth chronotherapeutic study’, European journal of cancer (Oxford, England : 1990), 199. Available at: https://doi.org/10.1016/j.ejca.2024.113571.

Fletcher, K. et al. (2025) ‘Immune checkpoint inhibitor infusion times and clinical outcomes in patients with melanoma’, The oncologist, 30(1). Available at: https://doi.org/10.1093/oncolo/oyae197.

Gonçalves, L. et al. (2023) ‘Immunotherapy around the Clock: Impact of Infusion Timing on Stage IV Melanoma Outcomes’, Cells, 12(16). Available at: https://doi.org/10.3390/cells12162068.

Huang, Z. et al. (2026) ‘RETRACTED ARTICLE: Time-of-day immunochemotherapy in non-small cell lung cancer: a randomized phase 3 trial’, Nature Medicine, 32(4), pp. 1233–1240.

Janopaul-Naylor, J.R. et al. (2024) ‘Effect of time-of-day nivolumab and stereotactic body radiotherapy in metastatic head and neck squamous cell carcinoma: A secondary analysis of a prospective randomized trial’, Head & neck, 46(9), pp. 2292–2300.

Karaboué, A. et al. (2022) ‘Time-Dependent Efficacy of Checkpoint Inhibitor Nivolumab: Results from a Pilot Study in Patients with Metastatic Non-Small-Cell Lung Cancer’, Cancers, 14(4). Available at: https://doi.org/10.3390/cancers14040896.

Karaboué, A. et al. (2024) ‘Why does circadian timing of administration matter for immune checkpoint inhibitors’ efficacy?’, British journal of cancer, 131(5), pp. 783–796.

Nomura, M. et al. (2023) ‘Timing of the infusion of nivolumab for patients with recurrent or metastatic squamous cell carcinoma of the esophagus influences its efficacy’, Esophagus : official journal of the Japan Esophageal Society, 20(4), pp. 722–731.

Othus, M. et al. (2026) ‘Addressing Biases in Analysis of Time of Infusion: NCI/SWOG Trial S1404 Among Participants With High-Risk Resectable Melanoma Who Received Adjuvant Anti-PD-1 Therapy’, JCO oncology practice, p. OP2501413.

Patel, J.S. et al. (2024) ‘Impact of immunotherapy time-of-day infusion on survival and immunologic correlates in patients with metastatic renal cell carcinoma: a multicenter cohort analysis’, Journal for immunotherapy of cancer, 12(3). Available at: https://doi.org/10.1136/jitc-2023-008011.

Qian, D.C. et al. (2021) ‘Effect of immunotherapy time-of-day infusion on overall survival among patients with advanced melanoma in the USA (MEMOIR): a propensity score-matched analysis of a single-centre, longitudinal study’, The Lancet. Oncology, 22(12), pp. 1777–1786.

van Rensburg, H.J.J. et al. (2022) ‘The effect of circadian rhythm on clinical outcome in patients receiving pembrolizumab in the INSPIRE pan-cancer trial’, Journal of Clinical Oncology [Preprint]. Available at: https://doi.org/10.1200/JCO.2022.40.16_suppl.2589.

Wang, C. et al. (2024) ‘Circadian tumor infiltration and function of CD8 T cells dictate immunotherapy efficacy’, Cell, 187(11), pp. 2690–2702.e17.

With thanks to Adam Rochussen, Eric Canton, Doug Marttila and Franco Tavella for thoughts on this blog!

I’m selling caffeine pills to college kids so they can pull all-nighters. Here ya go, buddy, I say, slipping them a blister pack. Pop one of these at breakfast, and you’ll have no trouble staying awake tonight.

The pills have a five-hour half-life. The alerting effect from breakfast is gone by early evening. My customers pass out in their study carrels at 2:00 am, and my caffeine pill business gets a slew of one-star reviews. Bad end.

Now I’m selling melatonin to insomniacs. I’ve made it so easy for you, I assure them. This is a super slow release formulation. Just take it once a month and watch the good sleep roll in.

The pills have a three-week half-life. My customers walk around in a stupor as the night hormone seeps into their system at all hours of the day. “It was better when I wasn’t sleeping at all,” reads one unhappy comment on our product page. Another game over; another bad end.

Where did I go wrong? I forgot time. I told people to put caffeine into their body in the morning when they actually needed it at night. I told people to put melatonin into their body at all times, when they only actually needed it at night.

The first is an example of a treatment that’s mistimed. The second is an example of a treatment that’s mis-designed for the time-sensitive process—sleep—it’s supposed to target.

These examples are silly, of course. For starters, the number of college kids pulling all-nighters to get work done is probably vanishingly small in 2026, unless they’re staying up to keep tabs on thirty-eight Claude agents overnight. And people generally get that sleep drugs, like melatonin, work best if they’re around when you’re sleeping and gone when you’re not.

But people get this because sleep is such an obvious circadian rhythm. You close your eyes, you make your room dark, you put on your Sleepytime Tea Bear costume—it’s impossible to miss that this happens about once a day. Yet there are many, many other circadian rhythms that exist right outside of the periphery of our awareness and play a huge role in how we act, heal, and feel.

You know this already. You know this because you’ve tried to open a jar of jelly first thing in the morning and found it incredibly difficult (musculoskeletal circadian rhythms). You know this because you’ve felt your mucus thicken ominously as night approaches during allergy season (inflammation circadian rhythms). You know this because you’ve thought about your life at 3:00 am and decided you, specifically, were uniquely horrible out of all the people to have ever lived (mood circadian rhythms).

But most non-sleep drugs don’t have much to say on the label about timing apart from “with meals” or “in the morning” with an unspoken “[so you don’t forget, dummy]” tacked on at the end. So why is timing so obvious for sleep and such a non-consideration for everything else?

Yeah, yeah, you say. It’s the title of the blog. Medicine forgot time.

And now, the rug-pull: Medicine didn’t actually forget time. It forgot to come back to it after the challenges of chronomedicine finally became tractable.

A brief history of chronomedicine

Look in the literature for papers on timing drugs, and you’ll notice a certain pattern.

PAPER from GROUP 1: Many have speculated that circadian rhythms in X could affect outcomes in Y. In this work, we sought to understand if treatment with drugs in class Z is modulated by rhythms in X. We compare dosing with Z at 8:00 am to dosing with Z at 6:00 pm and find that outcomes in Y are improved in the 6:00 pm group. We therefore conclude that, for improving outcomes in Y, dosing with Z in the evening is optimal.

PAPER from GROUP 2 (18 months later): no it’s not.

It isn’t actually this bad, but it’s close. Sometimes the data from Group 1 is suspect. Sometimes the critiques from Group 2 are overstated. I’ll be writing about a lot of them here, in future blogs. But if we put aside the universes where all Group 1s are nefarious data fakers or all Group 2s are determined naysayers and assume—as has been the case in my own experience—that people are generally doing their best and acting in good faith, what could explain this disconnect happening as often as it does?

It’s because we’ve been trying to do conceptual surgery while wearing giant, four-fingered Mickey Mouse gloves. We’re brute-forcing things, not finessing them. We’re taking a complex, dynamic strategy game—hitting a moving target in the human body—and button-mashing A and B. Dose in the morning or dose at night.

Sometimes button-mashing works; sometimes it doesn’t. The only way to reliably win is to learn the controls of the game.

So what are the controls?

At its core, the idea I’m talking about boils down to hitting a rhythm that goes up and down once a day with a drug in a way that maximizes or minimizes a desired outcome, like efficacy or toxicity.

Thinking from first principles, what should matter here? For starters, the rhythm itself. Strike when your enemy’s guard is up, and you’re going to be pretty ineffective. Wait until they’re wide open, and it’s a different story.

Two people in the same time zone can have radically different rhythms (shifted ten hours apart from each other, say) by virtue of having different genetics and living on different schedules. Give these two people the same drug at 8:00 am, and you’re hitting their biological rhythms at very different points. One might have a target that’s wide open, while the other is relatively shielded. Average them together and you blunt the effect.

Same goes for rhythm amplitude. People with a flat rhythm will have the same outcome no matter when you dose them. Include enough of those people in your patient population, and you kill the effect size.

Basically no chronomedicine studies control for participants’ circadian rhythms because it’s too hard. The best ones use sleep timing as a proxy, but even that’s a static, imperfect approximation that ignores amplitude.

What else matters? How about the pharmacokinetics and pharmacodynamics (PKPD) of the drug? A drug with a 6-hour half-life will have a different best time than a drug hitting the exact same target with a 12-hour half-life. Not to mention the time until peak, or the more complicated dynamics of how byproducts of the drug decay and are sequestered.

Drug interactions matter. Someone on a CYP inhibitor won’t have the same PKPD as someone not on one. Someone taking dexamethasone might have their circadian rhythms scrambled—or have negative consequences from the dexamethasone mask positive outcomes from optimal timing of a different drug. If one drug is best in the morning and the other is best at night, and you take both of them at the same time, you’ve (once again) blunted any effect you could hope to observe.

Drug metabolism matters—rhythms in the liver can gate what gets through to the rest of your body. Meal interactions matter. Sex matters. Dose matters. Different amounts of a drug can make for different best times. No one talks about this! It matters! No wonder different groups get different answers!

And there’s absolutely nothing at all in biology that says a best time has to be morning or night.

A treacherous landscape stretches out before you. The ground is composed of shifting blocks, meteors rain down from on high, pillars of lava surge up at random intervals. If you can just keep track of all the moving pieces and get to the other side, you win.

You gaze out upon it all, lit from below by the flames, face impassive.

Then: A A A A A A A B B B B B B B A A A A A A A A A A

Despair, despair

The reason we don’t just try all possible times in drug development is that it’s too expensive—nobody wants to multiply the costs of a clinical trial by 2, 12 or 24. Not to mention the fact that circadian rhythms are dynamic and mutable, making it so the best time for a drug technically depends on the circadian state of the person taking it in the moment they take it. Combine that with everything I mentioned above, and it’s easy to be a pessimist. Why bother thinking about time at all? Keeping track of everything will take, frankly, a lot of math. It’s never going to be worth the work.

If you’d said that in 1999, or 2010, or 2020, then okay, sure. Keeping track of everything would take a lot of math. I came of age in an era where much of computational biology’s cutting edge lived in folders of MATLAB scripts maintained by grad students. You had to squint a bit to see the latent explosive potential in files like runSodiumChannel4.m.

But we live in 2026

The nice thing about living in the future is that you sometimes get to do futuristic-sounding things. In my case, at my company, Arcascope, this is building digital twins— combinations of equations and data that attempt to, in real time, tell us more about what’s going on inside the body than we can figure out from data alone.

There are many ways to put digital twins to work; chronomedicine, to me, is the best and most obvious one. If the field has been held back by a lack of specificity—in how we talk about the notion of a best time, in how we attempt to study it—then digital twins, which replace words like “we need to consider the role of CYP3A4” with equations that do exactly that, are clearly part of the solution. Deep digital twins; ones that grow and evolve over time with new experimental data, new structural insights, and new forms of input.

A phrase I love, but that lands differently depending on the audience, is that chronomedicine is “Medicine in the Fourth Dimension.” As in: You’ve heard of x, y, and z as the spatial dimensions—now here’s TIME with the STEEL CHAIR. To me, adding a new dimension speaks of near unimaginable potential, blasting open the realm of the possible. An entirely new dimension to add to our definition of a therapeutic window. Something that can illuminate muddled or confusing data into a clear, mechanistic, time-dependent story. Remember the fake super-long-acting melatonin pill from the top? How many real therapies have been done in by off-target effects because they flooded the system with a drug when they should have pulsed it? How much more can we learn and build when we bring time into the picture?

But it can also come across as a bit… same-old. Time’s not a sexy new drug or delivery mechanism. Time’s been here this whole time. Haven’t you guys figured Time out already?

We haven’t. We forgot about it because it was hard and expensive and inconsistent, and it was easier to wave it away as irrelevant because the effect sizes couldn’t possibly justify the effort. But a caffeine pill at the wrong time isn’t “10% less effective”, it’s useless. And there are problems out there (real problems, persistent problems, time-relevant problems) with much higher stakes than “I want to stay up late.” Your brain cancer treatment might be twice as effective if you dose in the early morning. Your diagnosis might be time-dependent. And these findings are just from papers I’ve got a personal connection to—there are lots and lots of great people out there doing great science in this space.

We’re only just now scratching the surface. Time to get to work.