Cells are constantly making decisions in response to the signals they receive. In an embryo, signaling cues guide where cells go, what they become, and how they interact with their neighbors. In disease, many of the same pathways can push cells toward proliferation, invasion, or resistance. To understand why a cell ends up in a particular state, we need to trace that outcome back to the signals that guided it there. While we are very good at measuring the state of cells at a single time point, it remains difficult to determine what those cells experienced before.
This challenge is what motivated this project. Many important biological signals are transient. They rise and fall, sometimes over hours, sometimes over days, and often in places that are difficult to image continuously. Yet cells can base long-term decisions on those fleeting experiences. We wanted to know whether it might be possible to let cells keep their own record of signaling history, and then read that record later from a simple endpoint measurement.
This idea sits within the broader vision of DNA recording: programming cells to write aspects of their history into their genomes so that a later snapshot can reveal not only their present state, but also part of their past. In our case, we wanted to record signaling activity quantitatively in single cells. That meant building a system in which genome editing events accumulate at a rate proportional to pathway activity, so that the final fraction of edited sites reports how much signaling a cell experienced over time.
At first glance, this may sound similar to lineage recording, where CRISPR editors generate mutations in genomic barcodes. But signal recording turned out to demand something fundamentally different. In lineage tracing, diversity is the goal: you want as many distinguishable barcode outcomes as possible. For signaling history, diversity by itself is not useful. What matters is whether the final edits can be interpreted as a faithful quantitative record of pathway activity.
That realization shaped the design of INSCRIBE from the beginning. Rather than asking each barcode to tell a unique story, we asked many repeated barcode sites to report on the same underlying process. In effect, each target site becomes one probabilistic measurement, and the collection of sites provides a more reliable estimate of signaling history. This statistical logic also simplified the readout. Because we did not need to decode every individual barcode in sequence, we could instead read out the fraction of edited sites in each array using only two competing probes. That ratiometric strategy was a key step toward making imaging-based molecular recording much more practical.
Design of the INSCRIBE writer, however, led us into a long period of trial and error. Our early designs failed in a very instructive way. We hoped to build an analog recorder, where each editing event would act like a probabilistic “hit” and the total number of edits would scale smoothly with signaling activity. Instead, our first versions often behaved in a “winner-takes-all” fashion. Once editing started on an array, many neighboring sites were edited together, so arrays were frequently almost entirely edited or almost entirely unedited. That was interesting, and it still carried some information at the population level, but it was not the kind of analog single-cell recorder we were trying to build.
That problem forced us to stop thinking of the writer as just a genetic construct and start thinking of it as a problem in kinetic engineering. The crucial question became which component should be coupled to signaling and how.
In our initial designs, gRNAs were expressed from a U6 promoter, which is strong and constitutive. That made it difficult to modulate editing in a graded way. The breakthrough came when we instead placed gRNA expression under a Pol II promoter responsive to signaling. Linking signaling activity to gRNA expression, rather than to base editor abundance, changed the behavior of the system in exactly the way we needed. Once the base editor was present above a threshold, it was the amount of gRNA that set the magnitude of recording. That gave us a much broader operating window and produced the analog behavior required for quantitative single-cell measurements. Importantly, Pol II-driven gRNA expression also provides a scalable framework for multiplexed recording, since gRNAs are programmable and can in principle be assigned to different signaling pathways.
Once the system worked, one of the most exciting aspects was what it allowed us to ask biologically. We engineered cells to record WNT or BMP signaling and asked whether past pathway activity could be reconstructed from endpoint measurements. It could. But the real surprise came when we compared responses to repeated stimulation. For WNT, the memory decayed quickly. For BMP, it did not. Cells that responded strongly the first time tended to remain more responsive to BMP again much later, and this correlation persisted for up to 18 days. That finding emerged directly from the ability to record signaling history over time rather than just measure a downstream snapshot.
To us, that points to the larger promise of this kind of approach. Many of the most important questions in development, regeneration, immunology, and cancer are questions about history: what signals did this cell experience, in what context, and how did those experiences shape what it later became? Sequencing-based recorders have shown the power of writing history into DNA, but they usually sacrifice spatial context. Spatial transcriptomic methods preserve tissue organization, but they largely provide static snapshots. INSCRIBE was designed to help connect those worlds by creating a genetic record that can be read out in situ and combined with other imaging-based measurements.
This is still an early step. The current system records cumulative activity, not the full temporal shape of a signal, and much remains to be improved. But the broader goal is clear: to move from snapshots of cells to reconstructions of cellular experience. If we can do that, then tissues may become readable not only as collections of cell types, but as maps of what those cells have lived through.