How Molecular Recorders Record Biological History in DNA
Most of our biological experiments are designed to give us a picture of what a cell is doing at the time of collection. If we do RNA sequencing, we can see what genes are active. If we measure proteins, we can infer which pathways are active. If we do single-cell sequencing, we can even isolate thousands or millions of cells and compare them.
But what we can’t typically see is what happened before the measurement.
Maybe that cell from the tumor we collected is expressing a stress-response gene now, but we can’t really tell from sequencing data if it was in a low-oxygen environment three days ago, if it got an inflammatory signal yesterday, or if it came from a specific ancestor that behaved differently from other cells around it. Or maybe it’s development. By the time you sequence that neuron or liver cell, you can tell what it’s become, but a lot of the sequence of events that led up to that identity is already gone.
This is why molecular recording caught my attention. Instead of trying to continuously monitor every cell, researchers are starting to engineer cells to record important events in their own DNA. Then we can go back, sequence that DNA, and piece together parts of the cell’s history.
In other words, the genome can be more than just a place to get instructions for biology. It can be a notebook.
Why a Snapshot Isn’t Enough
We’ve become really good at measuring the state of cells in modern biology. With RNA sequencing, you can measure thousands of transcripts. With ATAC-seq, you can look at accessible chromatin. With proteomics, you can measure proteins. With microscopy, you can look at where molecules are. And with single-cell methods, we’re taking it further, describing individual cells instead of averaging them together.
But biological systems are dynamic.
Imagine two cells that have very similar RNA profiles when you collect them. One might have gotten to that state because it got a strong inflammatory signal. Another might have gotten to that state through a completely different pathway. But if you just take a snapshot, they might look the same even though their histories are different.
With development, it’s even more obvious. All the cells in your adult body ultimately come from that one fertilized egg. But there’s no obvious family tree connecting one liver cell to one neuron. The cell divisions happened and then they’re gone in the past. Sometimes you can infer relationships based on natural mutations or similarities in cell state. But inferring isn’t the same as directly recording the lineage as it’s happening.
To tackle this, scientists are developing what are often called **molecular recorders**. These are systems that can take transient biological events and turn them into stable changes in DNA. Because DNA sticks around when cells divide, that record can be copied to the descendant cells and then recovered.
It sounds simple enough in concept, but it’s a tricky engineering problem. A good recorder needs to be able to record information without too much damage to the cell. It needs to be able to distinguish between events. It needs to preserve the order of events. It needs to store enough information to be useful. It needs to be readable after many cell divisions.
DNA Is an Attractive Storage Medium
We already have an incredibly rich amount of information stored in DNA, in a very small physical space. Cells copy it when they divide. And we have the ability to read millions of DNA molecules in parallel with sequencing technologies.
So DNA is a logical place to put a synthetic biological record.
But the challenge is that normal DNA doesn’t automatically record what a cell is experiencing. Maybe a signaling pathway is on for twenty minutes and then it’s gone. But it doesn’t leave an obvious permanent mark in the genome.
So a molecular recorder needs two things.
First, you need a sensor that can detect the event you’re interested in. And second, you need a writer that can take the activity of the sensor and turn it into a stable change in DNA.
Early recording systems used recombinases, CRISPR-Cas enzymes, or base editors to make permanent changes in DNA. They showed that you could record biological events, but they often had limited storage capacity or had trouble preserving the order of events.
If you have a cell that gets signal A, then signal B, then signal C, you don’t just want to know that you got all three signals. The order can contain biological information.
This is what led to DNA Typewriter.
How DNA Typewriter Uses Prime Editing as a Recording System
In 2022, Junhong Choi and colleagues described **DNA Typewriter**, a molecular recorder using prime editing (Choi et al., 2022).
Prime editing is a type of genome editing that uses a modified Cas9 protein fused to a reverse transcriptase (an enzyme that can copy RNA information into DNA). It’s guided by a prime editing guide RNA (pegRNA) that both targets the editor to a site and specifies what sequence to write.
DNA Typewriter uses this system in a really clever way.
They’ve engineered a genomic region they call **DNA Tape**. This tape is a series of repeated target sites. At first, only the first target site can be edited. When a pegRNA puts a short DNA barcode into that target site, the same edit makes the next target site accessible. Then you can move forward and write another barcode.
So each insertion does two things at once. It records a symbol and advances the writing position.
Think of it like the typewriter head moving across a line. If the first edit records A, the second edit records C, the third edit records B, then when you sequence the finished tape, you can recover A-C-B in that order.
This is fundamentally different from just marking three independent sites. Maybe you have separate marks that say you got A, B, and C, but you don’t necessarily know which one came first.
In their original *Nature* paper, they showed how DNA Typewriter can encode thousands of different possible symbols and recover ordered event histories in mammalian cells. They even showed how you can store short text-like information as a proof of concept for the capacity of the system (Choi et al., 2022).
The goal wasn’t to necessarily make cells store English sentences. The goal was to show that you could use a living genome as a sequential recording medium.
Study 1: ENGRAM Lets Cell Signaling Write Its Own Barcode
So DNA Typewriter showed us a recording architecture. But how do you link a real biological signal to that writing process?
In 2024, Wei Chen, Junhong Choi, and colleagues published **ENGRAM** (enhancer-driven genomic recording of activity in multiplex) (Chen et al., 2024).
The idea is to link the activity of regulatory elements to the production of a pegRNA.
A **cis-regulatory element** (CRE) is a stretch of DNA that helps control gene expression. Enhancers are a type of CRE. They become active when specific transcription factors or signaling pathways are active in a cell.
With an ENGRAM circuit, you can have a regulatory element control the production of a guide RNA with a barcode that corresponds to that biological signal. So when the regulatory element becomes active, you get production of that guide RNA, you get prime editing, and you get a specific DNA insertion into the recording site.
So you turn a transient signal into a permanent barcode.
They tested this out with a few different regulatory programs. Instead of just measuring pathway activity while a signal is present, you can go back later and recover evidence of past activity by sequencing the genomic record.
ENGRAM is also multiplexable. You can have different signals lead to different barcode sequences. This lets you record multiple biological events into the same genomic site.
Even with a 4-base DNA insertion, you have 256 possible sequence combinations. Not every sequence might work equally well in practice, but you can see how the different possible symbols can scale differently than just turning a genomic target on or off. In the later ENGRAM protocol, they explicitly mention 4-base insertions as potentially representing up to 256 different regulatory element activities.
They also linked ENGRAM with DNA Typewriter. So in principle, you could record not just which signals you got, but the order of those signals.
This is where things can get much more powerful.
Maybe eventually you could record a history like inflammatory signal, then growth signal, then differentiation signal, instead of just having a single measurement at the end.
Molecular Recording as a Bioinformatics Problem
Recording information is only half the system. Now you have to be able to read that information.
After you do your experiment, you can sequence the DNA from thousands or millions of cells. But each read might have a different combination of insertions. Cells didn’t all experience the same signals. Or maybe the efficiency of the editing was different.
So you don’t have a simple diary written in full sentences as your raw output. You have a big pile of partial records.
You need computational analysis to figure out what barcodes are present, in what order, and at what frequencies. For lineage recording, you can look at the shared editing patterns across different cells. If you have cells that inherited the same early edits, they probably share a recent ancestor. If you have edits that are only present in a smaller branch, maybe those happened later.
So now you have a phylogenetics problem.
Normally, phylogenetics is used to reconstruct evolutionary relationships between species or viruses based on genetic differences. With molecular lineage tracing, you can use similar logic within a single organism. Instead of asking which species share an ancestor, you’re asking which cells share an ancestor.
And you can represent that as a tree.
You have the root corresponding to an early ancestor. You have branch points that represent cell divisions or lineage separations. And you have leaves that represent cells collected at the end of the experiment.
If you also have single-cell RNA sequencing data from the same cells, you can assign cell identities to those leaves. Now you have a tree with history and present-day cell state.
And that can be incredibly useful for developmental biology.
Study 2: peCHYRON Tries to Make the Recording Open-Ended
With DNA Typewriter, you have a predefined array of recording sites. So there’s a physical limit on how many sequential events you can store.
But with a different system called **peCHYRON**, the goal is to make the recording process open-ended.
Theresa Loveless, Courtney Carlson, and colleagues published the system online in November 2024, with the final issue publication in *Nature Chemical Biology* in April 2025 (Loveless et al., 2025).
Like DNA Typewriter, peCHYRON uses prime editing. But the architecture alternates between target states. With each recording step, you insert a short variable sequence that serves as a symbol, as well as a constant propagator sequence that creates the target you need for the next edit.
So instead of moving across a fixed row of sites, you’re growing the recording locus as you insert more information.
They showed how you can get diverse insertion patterns with collections of pegRNAs. You can use that to reconstruct lineage relationships. They also linked specific guide RNAs to inducible signals, so you can record exposure to different stimuli.
One advantage of this architecture is that it can theoretically be open-ended. As long as you keep editing and the locus remains functional, you can keep appending more information.
But it also shows why biological recorders are trickier than digital hard drives.
The recording efficiency can decrease. Some components of the recorder can interact. Some cells can silence the engineered sequences. There are fitness costs to record over long periods. Sometimes recording loci can be lost. In the authors’ experiments, they directly observed issues like silencing the recorder, losing recording loci in long experiments, etc.
These aren’t just minor details about engineering.
If a recorder only records when it feels like it, then what does it mean if there’s no mark? Is it because the cell didn’t experience the signal? Or because the recorder didn’t record?
This is a key distinction in the field.
Recording Has Its Own Measurement Error
The process of measuring things always alters what you’re measuring to some extent.
In the case of molecular recording, the interesting thing about this problem is that the measurement system is physically inside the cell.
You have to express the prime editor. You have to make the guide RNAs. You have to modify the DNA. All of these things can potentially impact cell fitness, gene expression, genome stability.
Also, a recorder has some finite rate at which it can record. If a signaling event is very brief, the recorder might miss it. If two events happen too close together in time, it can be hard to figure out the order. If the editing is slow, then the DNA record can be delayed relative to the biology.
So, recording efficiency is a bit like the frame rate of a camera. If the camera is slow, you can’t really capture a fast event very well.
There’s also a tradeoff between recording activity and perturbation. You might be able to improve the amount of information you can record by increasing the expression of the editor, but you might also be causing more stress to the cell.
So, you need to characterize a good recorder not just in terms of what it can store, but also in terms of how faithfully it represents the process.
Study 3: Better Parts to Make Larger Recording Circuits
Synthetic biology often involves parts that seem trivial, but when you start putting a lot of them together, it becomes more complicated.
In 2025, Troy McDiarmid, Megan Taylor, and colleagues published a study in *Nature Biotechnology* on promoters and guide-RNA scaffolds for mammalian genome engineering and molecular recording (McDiarmid et al., 2025).
A promoter controls when and how much an RNA is transcribed. For molecular recording systems, you might need a lot of different guide RNAs, each with its own promoter.
But if you keep using almost the exact same promoter sequences, then you can end up with large genetic constructs that are unstable and hard to assemble.
So, they looked for thousands of diverse RNA polymerase III promoters in mammalian evolutionary sequences and synthetic sequence variants. They experimentally tested 3,566 ancestral, extant, or mutagenized promoter sequences in one big screen. Then, they made collections of components that were different enough in sequence to make them easier to assemble, but still useful in terms of activity.
They also made a ten-key DNA Typewriter construct with ten different barcode writing units. They showed that they could detect all the expected insertion barcodes for ten sequential recording positions.
This study isn’t as visually exciting as reconstructing a developing animal. But it addresses an important scaling problem.
A molecular recorder that can track one signal is an interesting experiment.
If you want a recorder that can track many pathways at once, you need a much larger library of reliable, noninterfering biological components.
This is often how synthetic biology progresses. You make the exciting circuit possible by having a collection of parts that are predictable enough to make it possible.
2026 Experiment: Recording the Development of a Mouse
The latest study is the one that really made this field feel much more concrete to me.
On October 8, 2026, Qi Yu, Haedong Kim, Sophie Seidel, and colleagues published a peer-reviewed *Science* paper on the application of DNA Typewriter to mouse development (Yu et al., 2026).
They wanted to reconstruct cell lineage from the zygote, the single fertilized egg, up to almost two weeks of development.
They engineered the recording system so that as cells divided, edits would accumulate. Since those edits were inherited, you’d have molecular evidence of the ancestry of descendant cells.
At the end of development, they did single-cell transcriptional profiling. And then they read the DNA recording information from the cells.
They were able to reconstruct a time-calibrated, zygote-rooted lineage tree with **1.28 million transcriptionally profiled cells** from one embryo.
That’s an impressive number, but the important thing is more about the structure of the experiment.
Each cell had two types of information.
The RNA profile told you what the cell was at the end of the experiment.
The DNA record gave you evidence of where the cell came from.
So, they could ask questions that you can’t really ask with a standard single-cell atlas.
When did different cell types start to diverge? What mature cell types had recent developmental ancestors in common? Did some early clones disproportionately contribute to later tissues?
They observed clonal dominance starting to emerge during gastrulation. This is the stage in development where the major body layers are established. Cells that were close together in the reconstructed lineage tree were much more likely to end up with the same cell type than random cells. But you could identify points of terminal differentiation where branches had different cell types.
They could also look at what types of cells tended to repeatedly show up together in clades. And from that, you could infer a temporal hierarchy of developmental relationships.
Now, this doesn’t mean they recorded every cell division perfectly.
This is still a computational reconstruction of the tree, based on an imperfect recording system.
But it’s a big shift in what you can measure in developmental biology.
Instead of just inferring lineage from the identities of the cells at the end, the cells themselves have a synthetic record of their own history.
Why This Is Different From a Cell Atlas
Single-cell atlases have been game-changing for biology. They’ve allowed us to catalog the types of cells and cell states in tissues.
But there’s a difference between an atlas and a lineage tree.
An atlas asks what is there.
A lineage tree asks where it came from.
Imagine you go into a school and you measure the age, classes, interests, test scores of every student. You can make a detailed map of the current population. But you don’t really know which students are siblings from those measurements.
Lineage recording is about adding the family tree.
This is important in developmental biology because you might have cells that look different at the end, but they’ve shared most of their developmental history. Or you might have cells that look similar, but they got to that state in different ways.
In the future, you could potentially combine lineage, transcriptomics, chromatin accessibility, spatial location, signaling history, etc. to make something much more rich than a static cell atlas.
It would be more like a movie of development, reconstructed from information left behind in the cells themselves.
Could Cells Record Disease Before We Notice It?
The obvious next applications are beyond embryonic development.
Cancer is an evolutionary process. You have tumor cells dividing, mutating, facing immune pressure, hypoxia, responding to treatment, etc.
In theory, you could record some of those events with a molecular recorder. And also keep the lineage relationships.
So, you might ask: is a drug-resistant clone derived from cells that experienced some specific stress pathway before treatment? Or do metastatic cells have a specific signaling history?
What about inflammation? Chronic inflammatory diseases involve repeated signaling events. It can be hard to reconstruct those from a biopsy months later. But in principle, you could have a biological recorder that preserves evidence of past pathway activity.
Another thing is environmental sensing. There’s been some exploration of using engineered microbes as living sensors. Bacteria can survive in some environments that are hard to continuously monitor. Like parts of the gastrointestinal tract.
But a lot of this is still conceptual or experimental.
It’s one thing to engineer recorders into cells in the lab or model organisms. It’s a whole other thing to use them routinely in humans.
You have to deliver the recorder. You have to make sure it’s stable. You have to make sure it doesn’t cause harm. You have to make sure you can reliably interpret the data. These are big hurdles.
There’s Also an Information Problem
The more signals you can track with a recorder, the more complicated the data is going to get.
Imagine you have a recorder that can track 100 possible cellular events in millions of cells over several weeks.
You don’t just get a nice clean chronological story when you sequence.
Some records might be incomplete. Some barcodes might be written more efficiently. Some cells might lose parts of the recorder. You can also have errors in sequencing.
So, you need computational models to try and infer the most plausible history from noisy observations.
This is where molecular recording starts to overlap with a bunch of areas in bioinformatics.
You have sequence alignment because you need to identify barcodes. You have phylogenetics because you need to reconstruct lineage trees. You have statistics because you need to estimate recording efficiency, uncertainty, etc. You have single-cell analysis because you need to connect the molecular histories with the cell state. You have experimental design because you need to figure out what signals are worth having separate recording channels before you even start the experiment.
So, even though it’s biological, the usefulness of this depends a lot on computation.
What Would Make Molecular Recording More Useful?
One thing we still need is better temporal resolution.
If the recorder writes too slowly, you can lose short events. If you have faster writing systems, then you can tell the difference between biological events that happen close together in time.
Capacity is another thing. If you have a fixed tape, it can eventually fill up. If you have open-ended systems, then you can have other stability issues.
Another thing is portability. If you optimize a recorder in HEK293T cells, it might behave differently in neurons, immune cells, stem cells, or a developing embryo. Promoters can have different activity in different cell types. DNA repair can be different. You can have epigenetic silencing of engineered constructs.
Most importantly, you need the recorder to be less intrusive.
Ideally, you’d want something that can observe the cell without noticeably altering its behavior.
We’re not quite there yet in biology.
What I Actually Think
The most interesting thing about molecular recording to me is that it changes how we think about the relationship between an experiment and time.
With most experiments, you have to decide when you’re going to look. You take the sample. You freeze the biology at a certain moment. And then you measure what’s left.
With molecular recording, you’re offloading some of that responsibility to the cell.
You still have to design the experiment in advance. But once you have the recorder in place, biological events can leave information behind as they happen.
So, in 2024, with the ENGRAM study, we saw that you could convert pathway activity into symbolic DNA records. With peCHYRON, we saw that you could append sequential information to a growing recording locus. Then, in 2025, we saw this work on parts for larger recording circuits. And finally, in 2026, with the mouse experiment, we saw how you could use DNA Typewriter to reconstruct a lineage tree with over a million cells from a single developing embryo.
I don’t think this means we’re close to having cells record everything that happens to them.
The systems we have now are engineered, selective, noisy, limited in capacity. You still decide what signals the recorder can detect.
But this limitation also highlights what’s new.
A microscope is looking at the cell while you’re looking at it. RNA sequencing tells you what the cell was doing when you collected it. A molecular recorder is trying to preserve something that’s hard to get at with either of those methods. It’s trying to preserve the history that happened before you got there.
This is a different kind of measurement.
If we keep improving these systems, then maybe some of the hardest questions in biology will be easier. Not because we figure out how to continuously watch every cell. But because the cells themselves figure out how to leave us a record.
References and Sources
Choi, J., Chen, W., Suiter, C. C., et al. (2022). “A time-resolved, multi-symbol molecular recorder via sequential genome editing.” Nature, 608, 98-107.
DOI: 10.1038/s41586-022-04922-8. Open DOI
Liao, H., Choi, J., & Shendure, J. (2024). “Molecular recording using DNA Typewriter.” Nature Protocols, 19, 2833-2862. Published June 6, 2024.
DOI: 10.1038/s41596-024-01003-0. Open DOI
Chen, W., Choi, J., Li, X., et al. (2024). “Symbolic recording of signalling and cis-regulatory element activity to DNA.” Nature, 632, 1073-1081. Published July 17, 2024.
DOI: 10.1038/s41586-024-07706-4. Open DOI
Loveless, T. B., Carlson, C. K., Dentzel Helmy, C. A., et al. (2025). “Open-ended molecular recording of sequential cellular events into DNA.” Nature Chemical Biology, 21, 512-521. Published online November 14, 2024, issue publication April 2025.
DOI: 10.1038/s41589-024-01764-5. Open DOI
McDiarmid, T. A., Taylor, M. L., Chen, W., et al. (2025). “A parts list of promoters and gRNA scaffolds for mammalian genome engineering and molecular recording.” Nature Biotechnology. Published November 11, 2025.
DOI: 10.1038/s41587-025-02896-2. Open DOI
Nathans, J. F., McDiarmid, T. A., Chen, W., et al. (2026). “Multichannel genomic recording of biological information with ENGRAM.” Nature Protocols, 21, 3918-3943. Published February 11, 2026.
DOI: 10.1038/s41596-025-01322-w. Open DOI
Yu, Q., Kim, H., Seidel, S., et al. (2026). “A DNA Typewriter records the cell lineage history of a mouse, from zygote to late organogenesis.” Science, eael0508. Published October 8, 2026. This is the peer-reviewed paper that superseded the earlier bioRxiv preprint.
DOI: 10.1126/science.ael0508. Open DOI
BioBytes archive and representative posts reviewed: What PCA Actually Does in the Real World, One Gene at a Time: How Perturb-seq Maps Cause and Effect, BioVerse Navigator and the Data Problem in Biology, How Artificial Intelligence Is Changing the Way We Find Medicines, The Ethics of Gene Editing, and the Designing a Protein Binder series.