Here’s a neutrophil, the most common white blood cell, chasing bacteria in a dish.
Our best method for studying that cell is to kill it and measure 20,000 variables.
Once. Meanwhile, the cell keeps on sensing, deciding, acting, re-sensing, until it catches what it’s chasing. That’s a closed control loop: you check where you are against where you want to be, and keep correcting until you get there. It’s the same thing you do when you catch a ball. The cell does this in real time without a nervous system, a preplanned trajectory, or a world model. And we get one frame of it.
If we want to solve all diseases in the next 10 years1, we need instruments that match the cell, not more of the same data.
Fast, alive, broad — pick two
Every tool in molecular biology trades off across three dimensions: speed (temporal resolution), viability (non-destructiveness), and breadth (bandwidth). You can pick two, but nothing gives you fast, alive, and broad, which is exactly what a closed control loop requires.
Say you want to know what that neutrophil will do next. That depends on its state: which regions of the genome are accessible, which genes are turned on, which proteins are present and where, what shape it’s in. The loop you just watched is the cell updating its state against the signals from its environment. In practice, researchers equate cell state with the genes being read right now; the definition has shrunk to fit the instruments.
You have two main approaches for reading that state.
The first is to figure out a few proteins that run the sense-decide-act loop and make them glow (using fluorescent proteins). Then you can watch them under a microscope in real time. The cell lives, but you can only tag 3-4 such proteins before the colors start blending2, so you get a fast but narrow view.
The second approach is to read the genes (RNA sequencing), which tell you what the cell is equipped to do (what proteins it can make), not what it’s doing right now. This layer moves on hours rather than seconds3, and reading these genes gets you thousands of measurements instead of a few, but you have to break the cell open and wait several days for the answer4. You can approximate a time course by killing a new batch of cells at each time point, and people do, but you’re stitching together different cells, not following the same one.
I’ve used a variety of these tools for the past 10 years to build predictive models of human cell aging, and I’ve always been frustrated by the compromises. I didn’t have measurements from the same donor or cell across aging (time-resolved data), so I, like everyone else, had to repeat experiments hundreds of times and average the readouts across cells, donors, and ages.
You might argue that’s fine and that populations are sometimes the right unit. But we’re trying to control a cell, not predict the population average, which means we need to act on the cell at hand, not the average of a million dead cells. And, sometimes, the average of many trajectories is a trajectory no cell ever followed.
Is biology even controllable?
Some might say that none of these measurement problems matter because cells are just too complex to steer, and no instrument would change that.
But more often than not, we fail to steer complex systems because we use open-loop control: we execute a pre-planned set of actions without using real-time feedback to correct errors (since the world changes around our system). In contrast, closed-loop control uses continuous measurement of the system’s state against a target to dynamically adjust inputs and stay on track.
Take the V-2 rocket, for example. The engineers knew the mechanics, calculated the trajectory, aligned all the parts, and fired them off. There was no intervening after launch, and only a rough idea afterward of where it had gone. To get better, they fired again. Prototype tests in 1943 landed half the rockets within 3 miles of the target5. Today, Falcon 9 lands on a barge roughly 300 by 170 feet6. The physics is the same, but we can now measure and act on the system as it evolves. The rocket can make thousands of microadjustments using real-time data from its onboard sensors. What if we could do that for biology?
Well, there is one subfield where we did that.
In 1952, Hodgkin and Huxley figured out how nerve cells produce electrical signals, and that’s the only physics-level quantitative predictive model of a cell process that we currently have7. A nerve cell holds an electrical charge difference (voltage) across its membrane, similar to a battery, and an impulse is a fast swing in that voltage. Using an instrument called a voltage clamp, scientists can fix the cell voltage at a specified number: it measures the cell continuously, compares against the set level, and feeds in exactly enough current to hold it there. You can think of it as cruise control for a cell membrane.
The cool part is that it wasn’t only used to modify nerve electrical signals, but also to learn how they’re generated. The same way the gas pedal position tells you about the slope of a hill, the current that the clamp has to feed in tells you about the cell membrane gates (ion channels) that let charge flow (how many are open and when). You can read and write the same variable (electrical signal), in the same loop, faster than the membrane can change.
Those equations from 1952 still predict nerve impulses today. Hodgkin and Huxley provided a proof of concept at bandwidth one, modulating a single knob (membrane voltage) out of thousands that define cell state, and spent weeks solving four equations by hand to do it8. Today, computation is no longer scarce, but their core insight stays the same.
Biology is controllable, just within a narrow scope for now.
AI changed the bottleneck
For most of history, biology was constrained by ideas. You had to come up with a story about how your system worked, then design an experiment to test it. Hodgkin and Huxley needed a theory of ion channels before they could write down an equation. Now, you can point a neural network at your dataset, and it will find a pattern. You don’t need the theory first anymore, so the binding constraint is measurement.
Dario Amodei argues in Machines of Loving Grace9 that AI could compress 50-100 years of biological progress into only 5-10. However, he names experiments and trials as the rate limit, not intelligence. I’d go further and say that the problem isn’t only that experiments are slow, but that the key experiments can’t be run at all, at any budget, because no instrument exists that can measure and modulate a cell’s state without destroying it.
Consider what happens when you try. You train a transformer (the architecture behind most modern AI models) on tens of millions of cells that have been interfered with (“perturbed,” in fancy lingo), then ask it what a cell would look like after you switch a gene on or off. When the first careful benchmarks came out, the predictions looked almost exactly like the average of the training data10. Basically, the model concluded that the perturbation added no predictive signal beyond the population mean.
However, that result is now contested (and by the time you read this, the consensus might change again). One group says the failures were due to badly calibrated metrics, and under better ones most models actually beat the average baselines11, but that was contested only a few months later by another study12. Arc’s Virtual Cell Challenge, with >1,000 teams competing, concluded that models don’t consistently beat naive baselines, and the winners used both deep learning and hand-built statistical features (not scale)13. Meanwhile, the data itself turns out to be thinner than presumed, with 65% of the 7,170 perturbations across 29 datasets being unreliable14.
Now say the optimists are right. Say the models are better than naive baselines. The problem is still that they’re predicting the wrong thing: the endpoint. None of them are asked for a trajectory, and even if one were, there’s no data to train it on.
If a model trained on endpoints alone can predict trajectories in cells it has never seen, I’ll concede that the information is recoverable from data we already have and that scaling is all we need. That’s a higher bar than the field has set for itself, though. Most of the effort is moving towards “the virtual cell”. It’s a model you can perturb on a computer rather than in a dish, and it has institutes, atlases, and billions of dollars behind it.
If it works, great. But the mechanics were all figured out by 1943 and the V-2 still missed London by miles because the rocket wasn’t capturing data about its environment. To steer a cell we need to know its state in real time because true control requires continuous measurement, not just a static prediction. Either way, the next step is building the instruments. In the worst case, they make training models cheaper; in the best case, they allow the models we already have to do things they currently can’t.
The next two sections present the technical case for why AI has solved biology’s chemistry problems, and why scaling alone can’t unlock dynamic cell state. If you want to jump straight to the proposed solution, skip to “Reading and writing biology”.
Chemistry problems first
The obvious counterargument is that AI has already revolutionized biology: it essentially solved protein folding (AlphaFold15), drug-protein interaction (DiffDock16), and custom binder (mini-antibody) design (BindCraft17).
Yes, but those are all chemistry (not biology) problems. These technologies do a good job of showing us the shape of individual Lego blocks and how two pieces click together, but they tell us nothing about how those blocks assemble and interact inside a living, changing system. When you put biological Legos together, you don’t get a Lego statue; you get a machine that continuously rearranges itself while you’re trying to look at it, like this cell morphing into a complex neuronal shape:
A folded protein is an equilibrium property, which you can measure with an arbitrarily slow instrument. Cell state, though, is a dynamic property where measurement speed matters. AlphaFold was trained on roughly 170,000 experimentally determined structures, built on 50 years of crystallography and cryo-EM (techniques that lock a protein in a crystal or ice so its structure can be read out by X-rays or electron beams, respectively). We were lucky enough to have five decades of physical data before the models arrived.
Everybody’s favorite AI-in-Bio (AIxBio) success story is actually a story about instrumentation.
We have a decent catalog of stable states (cell and protein atlases), but we have very little clue how living cells transition between them. Yamanaka published the recipe for making stem cells two decades ago, yet we still don’t know which cells convert, when, or by what route18. We know the destination, but not the path.
Why scaling alone won’t work in Bio
The traditional AI playbook says that scale is all you need. Since scaling already solved language, it should also solve Bio. But text had a few cheat codes that biology doesn’t:
Recorded intermediates. You have the first word, the last word, and everything in between, in the right sequence.
Training task = output task. Next-token prediction (guessing the next word) is both the practice run and the performance.
Free, self-supervised data. Frontier labs have access to ~10¹³ tokens at near-zero marginal collection cost19, and the text labels itself.
Lossless recording. Every character I type is the character you read.
Digital verification loops. Models can test millions of code or math solutions digitally, for pennies.
We decided to voluntarily record our thinking at full temporal resolution, in sequence, for decades and make it open-access. Similar to protein folding, LLMs came after a 30-year, trillion-dollar instrumentation project called the Internet.
Bio has no Internet of cells. Nobody has been recording cell state at full temporal resolution for 30 years for free. Our primary workhorse, Perturb-seq20 (barcoding cells to measure active gene expression after switching specific genes on or off), fails on every axis:
No time resolution. We only see the beginning and the end (first and last word of an entire paragraph).
Disconnected task. Models predict static endpoints without seeing the path there.
Expensive and destructive. The largest atlases have ~10⁷–10⁸ (dead) cells21 compared to 1013 tokens for an LLM.
Lossy data. You capture ~10% of cellular RNA, and about a third of the sequencing data never map to the reference genome.
Physical verification. Biological experiments take weeks, cost thousands of dollars, and are deeply noisy.
Brute-forcing model parameters (like Xaira’s 4.9B-parameter model trained on 25M cells22) yields only incremental gains because scaling can’t hallucinate unrecorded intermediate steps.
Instead, if you want to overcome this by generating “more data”, you hit three fundamental walls:
Combinatorial Explosion: Screening 20,000 human genes yields 108 pairwise and 1012 triplet combinations. Right now, a good platform can maybe do ~106 conditions a week23, and you can’t shortcut it by learning the singles and adding them up, because gene effects don’t compose: pairs routinely produce effects that aren’t the sum of their parts24.
Unmeasurable Intrinsic Dimensionality: Intrinsic dimensionality (d) is the number of independent “knobs” driving a system. Even though language has tens of thousands of word-fragments, it sits around d ≈ 42, which means that you need roughly 1,450 times more data to halve your prediction error25. Estimators for single cell state dimensionality don’t converge26, and the numbers keep moving as you add more samples. So no one knows if you need 10x or 10,000,000x more data.
Bandwidth ≠ Trajectory: Multiomics (measuring RNA, proteins, and epigenetics from the same cell) increases bandwidth per endpoint, but yields marginal predictive gains, while still killing the cell and ignoring time.
Even when we do have non-destructive repeat sampling, like the blood protein clocks27 that allow you to learn in days (not months) whether an intervention worked, reading a trajectory is still not enough without real-time steering. Measurement is just half the loop.
Scaling LLMs worked because we spent decades instrumenting human thought in real time. AIxBio is trying to speedrun LLMs without the basis to train them.
Reading and writing biology
So how do we make progress in biology?
I’d argue we already are; it’s just slower and less flashy than another foundation model trained on 842,657,227 cells, capable of predicting about 2% better than the mean.
Here are some results on both sides of the read/write (measure/modulate) loop.
On the read side, Biohub28 recently showed that we can infer a lot about a cell’s state after a perturbation from just shining light through it. Turns out we can track a rocket’s trajectory from the ground using its contrail. This method matched the resolution of RNA sequencing, and with enough cells exceeded it. The first time anything has managed all three: non-destructive, fast, and broad. Or as the authors put it, a foundation for profiling the trajectories of living cells.
Scaling this to whole organisms, a study that just came out demonstrated real-time, organ-wide functional imaging across live fish29. In this video below, you can see neural and cellular signals firing across different body parts of a live fish:
Now, optical imaging is just a proxy, not a full readout. You aren’t reading 20,000 genes from a contrail. But the good news is that control doesn’t require measuring every internal variable or knowing the complete mechanism. It just requires a signal that tracks the state fast enough to correct against. Remember how Hodgkin and Huxley controlled the nerve cell with just one number? They didn’t need to know how a specific ion channel worked, just the current required to hold a constant voltage.
On the write side, JURA Bio30 developed a method for designing and testing ~1016 antibodies in one experiment. The chemistry itself does the sampling, so the synthesis reaction is the actual generative model. It’s the equivalent of building ten quadrillion rockets with different designs, launching them all at once, and only keeping whichever one lands closest to the pad. Real write-bandwidth scaling, but screening 1016 designs once is still open-loop, with no control.
And now we even have a foundation model with time in its data. A model trained on (only) 1.67 million cancer patients (DNA, RNA, pathology, and clinical episodes over time) was given two patients, one who lived longer and one who didn’t, and it picked the right one 77% of the time. Hand-picked clinical and molecular markers (the standard approach) managed 56%, which is barely better than a coin flip. More importantly, the model ranked treatment benefit three times better across eleven different cohorts31.
So we have half a loop on the read side, half on the write side, built in isolation by separate groups, and nobody has put them together. Yet.
Closing the loop
So let’s go back to what’s missing.
Now, some might argue we don’t need a loop. Most medicines don’t work this way. You give a drug, it (hopefully) shifts a population, and enough patients get better.
That is factually true. Population-level medicine has worked really well, and it’s a big part of the reason we live as long as we do. But it works by shifting a distribution, and it gets approved on that basis. If you want most cells to arrive at the thing you’re aiming for, or most patients to respond to the drug you make, you need control over biology.
To steer a cell, you have to read it and write to it faster than it changes. In control theory, engineers sample at roughly 10 times the rate of the thing they’re steering32. Signaling inside cells moves in minutes: proteins like ERK and NF-κB oscillate on 10-100 minute cycles33, which means that for a 10x margin we need at least one non-destructive reading per minute from a live cell.
Here’s a target spec for full cellular control:
Read: ~10³ channels, non-destructively, at least once a minute, from the same living cell, for at least one cell cycle (enough signals to capture the cell’s complete internal state without getting lost in noise)34.
Write: ~10² independent, reversible channels, with under a minute of latency (enough control knobs to steer the key master switches that decide cell state)35.
Close the loop at 1 minute.
The obvious (and fair) objection is that you can’t measure or modulate that much inside something as small as a cell without changing it. But you could do it with label-free optics and light-activated switches. These two methods can reduce traditional chemical toxicity by using light to both measure (read with scattered light) and modulate (write with patterned light) cell state36. And scaling light projectors, sensors, and computer vision models for throughput is just an engineering problem.
Then, we need to change what we score. Right now the leaderboard rewards getting the destination right: how closely your predicted cell resembles the real one some time later. We should score trajectories instead. Can you predictably and robustly drive a population of stem cells toward a specific neuronal fate?
People are looking for biotech’s version of “cost per kg to orbit” (a common metric for measuring space tech progress), but the candidates proposed so far37, like “experiments per unit time” or “aging-clock accuracy”, don’t compound. Cost per kg fell from $54,500 to under $2,000 because rockets learned to land38: a dynamic control problem, solved with real-time sensors and actuators.
Here’s what control looks like when the right instruments exist. When you point a ground telescope at a star, atmospheric turbulence distorts incoming light roughly 100 times per second. Adaptive optics cancels this distortion in real time by correcting the image 1,000 times per second (10x margin over the disturbance)39. It took decades of hardware engineering to solve, but it gives ground telescopes space-grade clarity. We see this same architecture in an artificial pancreas: a glucose monitor reads blood sugar every 5 minutes, an algorithm compares it against a target, and a pump delivers insulin40. Sense, decide, act, re-sense. Yet, our current tools in biology sit at opposite ends of a broken spectrum: live-cell fluorescence is fast enough, but at low bandwidth; single-cell sequencing is broad enough, but at a rate of zero. We need something in the middle.
The good news, though, is that the models are not the problem. If we build the right instruments, we already have the architecture and we can use it right away. That neutrophil chasing bacteria has been closing the control loop for hundreds of millions of years.
AI for Bio is not bottlenecked on intelligence. It’s bottlenecked on instruments. Until we can read and write cell state at the rate cells actually run, scaling up models will only get us better guesses about the endpoint with no ability to steer.
Open-loop biology.
Thanks to Russell Gould, Mike Riggs, Lars Doucet, and Latham Turner for feedback on early drafts of this essay!
Standard live-cell fluorescence imaging can differentiate 3-4 distinct probes before overlap becomes limiting; more advanced setups can scale up to 8 probes, but that’s still fairly low bandwidth for biological information. [Kumar et al., Nature Photonics 2025].
Neutrophil transcriptional responses peak around 9 hours [Sanz-Santos et al., BMC Proceedings 2011], while chromatin accessibility changes (via ATAC-seq) are observed between 1 and 4 hours post-stimulation [Ram-Mohan et al., Life Science Alliance 2021]. The chemotaxis in the video is protein-level (actin polymerization) and runs in seconds [Weiner et al., Nature Cell Biology 1999].
Single-cell RNA-seq on 10x Genomics platforms detects a median of roughly 4,100 to 6,900 genes per cell. Technical dropouts still heavily impact the data depending on sequencing depth and baseline expression [Gezelius et al., NAR Genomics and Bioinformatics 2024].
V-2 prototype tests in 1943 achieved a 4.5 km (2.8 mi) circular error probable (the radius within which 50% of the shots impact), with all shots falling within 18 km (11.2 mi); operational accuracy against London was worse, with 518 of 1,225 rockets landing inside the Greater London defense zone [Encyclopedia Astronautica].
Falcon 9 boosters land within a few meters of target on autonomous droneships, against a deck tolerance of roughly 10 m. The landing deck of the first droneship, Just Read the Instructions, measured 52 × 91 m (170 × 300 ft) [Blackmore, National Academy of Engineering 2016, CBS News 2014].
Hodgkin and Huxley's quantitative model of the action potential [Hodgkin & Huxley, Journal of Physiology 1952].
Since the Cambridge computer was offline for six months, Huxley integrated the equations using a hand-operated Brunsviga calculator. It took three weeks of manual calculation to simulate just a few milliseconds of nerve activity [Alan Hodgkin, quoted via University of Pennsylvania].
Dario Amodei, Machines of Loving Grace: "my basic prediction is that AI-enabled biology and medicine will allow us to compress the progress that human biologists would have achieved over the next 50-100 years into 5-10 years." Amodei explicitly identifies the speed of experiments and clinical trials, rather than intelligence, as the binding constraint. [darioamodei.com]
Benchmarks evaluating five foundation models on gene perturbation found none outperformed predicting the simple mean of the training data or using a basic additive model [Ahlmann-Eltze et al., Nature Methods 2025]. An independent evaluation across four major datasets similarly found a basic random forest or train-mean correlation consistently outperformed models like scGPT and scFoundation [Csendes et al., BMC Genomics 2025].
Under a calibration framework using positive and negative controls, most deep learning models outperform simple baselines on unseen-gene prediction. The authors argue earlier negative results "largely reflect shortcomings of evaluation rather than of modeling" [Miller et al., bioRxiv 2025].
An independent evaluation reaches the opposite conclusion: across all datasets and criteria, deep learning models do not outperform simple baselines, especially under distribution-aware metrics [Heidari et al., bioRxiv 2026].
The Arc Institute's Virtual Cell Challenge, drawing over 1,200 teams, concluded that prediction models do not yet consistently beat naive baselines. Winning entries relied on a combination of deep learning and classical statistical features rather than pure scale [Arc Institute].
Across 29 datasets, 65% of 7,170 perturbations were classified as unreliable. Training on the reliable subset alone matched or outperformed models trained on the full dataset [Wang et al., bioRxiv 2026].
AlphaFold2 was trained on roughly 170,000 real protein structures from the Protein Data Bank along with an additional ~350,000 self-distillation predictions (AI-generated structures used for training the model on its own output) [Jumper et al., Nature 2021].
DiffDock uses a diffusion generative model (the AI framework behind image generators, adapted here to model 3D molecular physics) to predict small-molecule binding poses and drug-protein interactions without requiring prior knowledge of the target binding pocket [Corso et al., arXiv 2023].
BindCraft can generate mini-antibodies (binders) entirely from scratch (one-shot de novo design) with 10–100% accuracy, bypassing the need to test millions of candidates by trial and error (high-throughput screening) [Pacesa et al., Nature 2025].
Cellular reprogramming is thought to be stochastic (random) since while a good number of donor cells can be converted to stem cells, the exact path and timing of any specific cell is unpredictable [Takahashi & Yamanaka, Cell 2006; Hanna et al., Nature 2009].
Llama 3 was pretrained on over 15 trillion tokens from publicly available sources, and GPT-4's training set (never disclosed) is commonly estimated at around 13 trillion [Educating Silicon].
Perturb-Seq pairs droplet-based single-cell RNA sequencing with barcoded CRISPR perturbations to recover perturbation identity and gene expression profiles simultaneously from pooled screens (multiple different perturbations thrown applied to a single group of cells all at once) [Dixit et al., Cell 2016].
The largest single-cell perturbation atlas to date is Tahoe-100M, comprising 100 million transcriptomic profiles across 1,100 small-molecule perturbations in 50 cancer cell lines, spanning 56,829 unique experimental conditions [Zhang et al., bioRxiv 2025].
Xaira's X-Cell, a diffusion language model scaling to 4.9 billion parameters, trained on X-Atlas/Pisces (25.6M perturbed single-cell transcriptomes across 16 biologically diverse contexts) reports up to 5x improvement on Pearson Δ over prior state of the art, plus zero-shot prediction in unseen cellular contexts [Wang et al., bioRxiv 2026].
Recursion’s top-tier automated screening facility runs up to 2.2 million experiments per week, driven largely by lower-dimensional morphological imaging (Cell Painting/brightfield microscopy) rather than full single-cell transcriptomics [Recursion].
Combinatorial Perturb-seq across gene pairs yields a principled classification of genetic interactions, which identifies synergistic pairs whose joint phenotypic effect can’t be predicted by summing their individual parts [Norman et al., Science 2019].
Neural scaling laws link empirical dataset exponents to data manifold dimension (α ≈ 4/d), deriving an intrinsic dimension of d ≈ 42 for natural language [Kaplan et al., arXiv 2020; Sharma & Kaplan, arXiv 2020]. Note that d is the number of degrees of freedom in the data and sets the learning cost. It’s not a list of 42 independent variables (or knobs).
Intrinsic dimension estimates for single-cell transcriptomics are far below the total ~20,000 gene space (typically in the tens to low hundreds) but fail to converge due to undersampling, causing the estimators to shift with sample size [Biondo et al., NAR 2025; Staiano et al., bioRxiv 2025].
Machine-learning models estimate the biological age of 11 major organs from plasma protein levels, validated across 5,676 donors. Since the assay relies on peripheral blood draws, individuals can be repeat sampled longitudinally [Oh et al., Nature 2023].
Biohub profiled ~57 million single-cell profiles across 1,000 CRISPR knockouts using fluorescence (39 live, 13 fixed markers), label-free phase imaging, and scRNA-seq. At scale, optical morphology matched or exceeded scRNA-seq in perturbation classification accuracy and state resolution, establishing non-destructive optical proxies for trajectory tracking [Liu et al., bioRxiv 2026].
The WHOLISTIC platform images cellular and physiological activity simultaneously across all major organs in a live vertebrate. By continuously tracking multi-tissue dynamics in real time rather than taking static snapshot assays, the authors mapped previously unobservable systemic phenomena: high-speed brainstem-driven vascular regulation, synchronized renal calcium waves, and rapid multi-organ metabolic shifts during acute stress [Ruetten et al., Nature 2026].
Instead of using a computer to generate sequences and then trying to synthesize them one by one, tuned chemical assembly directly translates a generative model’s probability rules into physical DNA synthesis. The chemical reaction itself acts as the computational sampler, producing and screening ~1016 designed antibody candidates in a single batch [Weinstein et al., Nature Biotechnology 2026]. (Full disclosure: My PI is a co-founder of JURA Bio.)
Multimodal oncology foundation model (oFM) trained on longitudinal data from 1.67 million cancer patients integrating clinical trajectories, DNA, RNA, and pathology achieved 77.4% survival prediction accuracy (vs. 56.3% for standard markers) and a 3-fold higher pooled treatment-benefit across 11 comparative cohorts [Vorontsov et al., arXiv 2026].
The minimal measurement frequency for having enough information to reconstruct the signal is 2x the system’s frequency (the Nyquist limit). For active feedback control you want 10x-20x faster measurement, to prevent phase delay and reaction lag, which would lead to overcorrection and destabilization [Åström & Wittenmark, Computer-Controlled Systems: Theory and Design].
Master signaling regulators move in and out of the cell nucleus on rapid cycles: ERK cycles every ~15 minutes (under EGF stimulation), while NF-κB cycles back every ~100 minutes [Shankaran et al., Molecular Systems Biology 2009; Tay et al., Nature 2010].
RNA sequencing datasets show that while cells have roughly 20,000 genes, their behavior can be accurately summarized using only 10-100 key dimensions [Heimberg et al., Cell Systems 2016; Moon et al., Nature Biotechnology 2019]. In control engineering, observing a system with an intrinsic dimensionality of d requires measurement of ~10 x d (1,000 features in this case) to eliminate ambiguity caused by noise and indirect optical proxies.
Major cell decisions (growing, dividing, or dying) are regulated by a core network of roughly 100 master proteins [Lambert et al., Cell 2018]. Using multi-color light patterns you can flip these switches on and off in real time without waiting for chemical drugs to wash in and out [Toettcher et al., Cell 2013; Müller et al., Nucleic Acids Research 2014; Bugaj et al., Journal of Cell Biology 2017].
Quantitative phase imaging can measure cells below toxicity thresholds over multiple cycles, by using low-power, long-wavelength light [Mir et al., PLOS ONE 2014]. Optogenetic modulation of cell states relies on short, low-duty-cycle pulses to actuate pathways in real time [Toettcher et al., Nature Methods 2011].
Jason Carman explores different cost-per-kilogram to orbit metrics for biotech progress, including “healthy years / $”, “experiments / unit of time”, and “aging clock accuracy” [Carman, Story Company 2025].
The price of getting a kg to orbit fell from $54,500 using the Space Shuttle (1981-2011) to between $1,500 and $2,720 using the Falcon 9 (a >95% reduction mainly due to booster reusability) [Orbital Intel, Launch Cost History 2026].
Adaptive optics corrects atmospheric distortion in real time: a wavefront sensor measures incoming distortion and a deformable mirror compensates for it. The loop rate is set by the atmospheric coherence time (how long the distortion pattern stays the same), which is about 1 ms at optical wavelengths. So adaptive optics loops should run at around 1 kHz (1,000 times per second), with systems spanning ~400 Hz to 5 kHz. Deformable mirrors with ~100 actuators (control points) can be driven to 1 kHz; systems built for extremely large telescopes use far more [Jenkins et al., Monthly Notices of the Royal Astronomical Society 2018; Paschotta, RP Photonics Encyclopedia].
Hybrid closed-loop insulin systems read blood glucose every 5 minutes from a continuous monitor, compare it against a target, and adjust insulin pump delivery in real time. Several of these “artificial pancreases” are FDA-approved and in routine use, like Tandem’s Control-IQ, which was validated in a 6-month randomized trial [Brown et al., New England Journal of Medicine 2019].
The most insightful essay i've read on AIxBio. I want to solve the domestic instrumentation layer--longitudinal population-scale data on all that comes and goes--high-throughput single cell and molecule sensing in cheap compact devices in the home, in air vents, toilets, kitchen sinks, etc.. (This is several years away and very very very very hard.) Your essay made me realize that nutrition, drugs, exercise, reported mood, posture, body composition, and a thousand other personal macroscopic things will enable closed-loop personal biological control systems. Like what athletes and bodybuilders have always done, but with deep and broad molecular information every time one uses the bathroom... every meal and every exercise session and every medication becomes an experiment.
Great discussion Alex. We are actually developing a sensor aimed at addressing the closed loop (alive, fast and broad) offering non-destructive readout of in-vitro systems state (particuladiagnostics.com). By "system state" I mean an integration of multiple molecular inputs; just like the cell state representation created by an LLM from omics data but already compressed in the form of a sensor's signal. Do you think a cell/system state tracking sensor could be impacting AI efficency in predicting biologic trajectories? maybe in combination with some selected omics snapshots?